Keywords: Thermal sensors and systems, Wearable wireless sensors, motes and systems, Integrated sensor systems
Abstract: We fabricated a wearable sensor that can be attached to the skin surface and continuously measure core body temperature (CBT) wirelessly over a long period. CBT is calculated from skin-surface temperature and heat flux passing through the sensor. Since heat flux is lost to the surroundings of the probe, the slightest change in convection in daily life will degrade the measurement accuracy of the sensor. Accordingly, we previously proposed a heat-flux-path control structure to reduce the absolute amount of heat-flux loss. To make wearable sensors for long-term human trials, we proposed an integrated design in which a sensor probe, a circuit board, and a battery are stacked. We optimized the proposed design by computer simulation and evaluated the fabricated sensor by a phantom experiment in which the convectional state was changed. The evaluation results demonstrate that the sensor has limits of agreement (LOA) of [−0.13; 0.03]°C under 1-m/s-wind convection. Moreover, a preliminary human trial conducted under daily-life conditions (including convectional changes) demonstrated that the sensor has LOA of [−0.18; 0.22]°C. These results demonstrate that the fabricated sensor is suitable for CBT measurement.

Keywords: Chemo/bio-sensing - Chemical sensors and systems, Chemo/bio-sensing - Biological sensors and systems, Health monitoring applications
Abstract: Access to low-cost, rapid, individualized diagnostics at point-of-care and point-of-need is vital to minimize the impact of highly infectious viruses, such as influenza. Herein, a biosensor for detecting hemagglutinin (HA), an abundant capsid protein in H1N1 viruses, is demonstrated. A gold working electrode was functionalized with a thiol-modified, HA-binding aptamer derivatized with a methylene blue modification for redox reporting. The aptamer was characterized by surface plasmon resonance to confirm its biorecognition activity for HA. The aptasensor was characterized by square wave voltammetry to quantify the sensor’s response to varying concentrations of HA. The sensor exhibited a lower limit of detection of 1.5 pM with linear detection of up to 1.2 nM in both Tris buffer and simulated human saliva, thus encompassing the clinically relevant HA range in saliva. Average sensitivity was measured at 21.083 nA·nM-1 in Tris and 14.5 nA·nM-1 in artificial saliva across clinically relevant HA titers. Sensor stability across time was also investigated, providing a preliminary understanding of the translational viability of the aptasensors for mobile and remote diagnostic applications.

Keywords: Bio-electric sensors - Sensor systems, Bio-electric sensors - Sensing methods
Abstract: This work presents a method to minimize the inadvertent cutting of tissues in surgeries involving bone drilling. We present electrical impedance measurements as an assistive technology to image-guided surgery to achieve online guidance. Proposed concept is to identify and localize the landmarks via impedance measurements and then use this information to superimpose the estimated drilling trajectory on the offline maps obtained by pre-operative imaging. To this end, we propose an asymmetric electrode geometry, split electrodes, capable of distinguishing impedance variations as a function of rotation angle. The feasibility of the proposed approach is verified with numerical analysis. A probe with stainless steel electrodes has been fabricated and tested with a technical phantom. Although the results are impacted by a non-ideality in the phantom, we could show that the variation of impedance as a function of rotation angle can be used to localize the regions with different impedivities.

Keywords: Physiological monitoring - Modeling and analysis, Wearable sensor systems - User centered design and applications, New sensing techniques
Abstract: The thermal noise due to the resistivity of insulation materials can become a significant noise source in non-contact capacitive sensing, especially when measuring micro-volt-level physiological signals. Since both the impedance and the resistivity of practical insulation materials may be strongly frequency dependent, their thermal noise is often frequency dependent. This paper studies the impedance and noise behavior of different interface materials as function of frequency, by means of modelling, simulations, and experimental measurements. The results show that the inherent resistive noise of some fabrics (e.g., cotton, polyester) could outweigh the typical noise level of circuits for physiological sensing; and as a result, the interface noise can limit the quality of low-amplitude signal detection.

Keywords: Chemo/bio-sensing - Chemical sensors and systems, Bio-electric sensors - Sensor systems, Chemo/bio-sensing - Biological sensors and systems
Abstract: Microneedles (MN) are short, sharp structures that have the ability to painlessly pierce the stratum corneum, the outermost layer of the skin, and interface with the dermal interstitial fluid that lies beneath. Because the interstitial fluid is rich in biomarkers, microneedle-based biosensors have the potential to be used in a wide range of diagnostic applications. To act as an electrochemical sensor, the tip or the body of the MN must be functionalized, while the substrate areas are generally passivated to block any unwanted background interference that may occur outside of the skin. This work presents four different passivation techniques, based on the application of SiO2, polymethyl methacrylate (PMMA), an adhesive film, and varnish to the substrate areas. Optical, SEM and electrochemical measurements were performed to quantitatively assess the performance of each film. The data shows that whilst manual application of varnish provided the highest level of electrical isolation, the spin-coating of a 5 m thick layer of PMMA is likely to provide the best combination of performance and manufacturability.

Keywords: Chemo/bio-sensing - Micrototal analysis and lab-on-chip systems, Chemo/bio-sensing - Biological sensors and systems, Bio-electric sensors - Sensor systems
Abstract: Moore's law has enabled massive scaling of complex computing and sensing systems in modern-day chip-scale architectures allowing extremely high yield and system complexity at very low-cost. Exploiting such Moore's law, we explore silicon-based integrated circuits and chip-scale systems to interface with biological fluids to manipulate, sense, and detect cells in real-time for an end-to-end low cost, miniaturized, and high sensitivity point-of-care diagnostics platform. Eliminating the need for complex, expensive, large and bulky syringe pumps and optical-based cytometers, the proposed system allows pneumatic-free AC electro-osmosis bulk fluid driving capabilities controlled by the CMOS chip, and integrated dielectrophoretic cell actuation with 2μm focusing accuracy, impedance spectroscopy sensing, and separation capabilities. The paper presents, for the first-time, a CMOS-driven cellular sensing platform for microfluidics that can be translated to a wide range of biomedical applications.

Keywords: Data mining and big data methods - Biosignal classification, Data mining and big data methods - Machine learning and deep learning methods, Neural networks and support vector machines in biosignal processing and classification
Abstract: Automatic electrocardiogram (ECG) analysis plays a critical role in early detection and diagnosis of cardiac abnormalities and diseases. Data augmentation and automation strategies have been proposed to enhance the robustness of the machine and deep learning model for the classification of cardiac abnormalities. Here we propose 15 data augmentation and 6 filters, and an automation method using an end-to-end deep residual neural network (ResNet) model for automatic cardiac abnormalities detection from 12-lead ECG recordings. We evaluate the effectiveness of data augmentation/filtering and automation techniques using the proposed ResNet-based model on the China Physiological Signal Challenge (CPSC) dataset consisting of 9 diagnostic classes. The average F1 scores across 9 classes on the CPSC dataset trained with three data augmentation (baseline wander addition, dropout, and scaling) and a filter (sigmoid compression) were significantly higher than that without using augmentation/filters (baseline). The highest average F1 score with sigmoid compression method was significantly higher (relative improvement of 2.04 %) than the baseline while horizontal and vertical flipping augmentations were detrimental to the classification performance. Additionally, the results show that the random combination of four selected data augmentation and filter using the modified RandAugment technique provided a significantly higher average F1 score (relative improvement of 2.54 %) compared to the baseline. The proposed data augmentation, filters, and automation technique provide an effective solution to improve the classification performance of the end-to-end deep learning model from ECG recordings without changing the model hyperparameters and structure.

Keywords: Data mining and big data methods - Machine learning and deep learning methods, Data mining and big data methods - Biosignal classification, Time-frequency and time-scale analysis - Wavelets
Abstract: Atrial fibrillation (AF) is a common supraventricular arrhythmia. Its automatic identification by standard 12-lead electrocardiography (ECG) is still challenging. Recently, deep learning provided new instruments able to mimic the diagnostic ability of clinicians but only in case of binary classification (AF vs. normal sinus rhythm-NSR). However, binary classification is far from the real scenarios, where AF has to be discriminated also from several other physiological and pathological conditions. The aim of this work is to present a new AF multiclass classifier based on a convolutional neural network (CNN), able to discriminate AF from NSR, premature atrial contraction (PAC) and premature ventricular contraction (PVC). Overall, 2796 12-lead ECG recordings were selected from the open-source “PhysioNet/Computing in Cardiology Challenge 2021” database, to construct a dataset constituted by four balanced classes, namely AF class, PAC class, PVC class, and NSR class. Each lead of each ECG recording was decomposed into spectrogram by continuous wavelet transform and saved as 2D grayscale images, used to feed a 6-layers CNN. Considering the same CNN architecture, a multiclass classifiers (all classes) and three binary classifiers (AF class, PAC class, and PVC class vs. NSR class) were created and validated by a stratified shuffle split cross-validation of 10 splits. Performance was quantified in terms of area under the curve (AUC) of the receiver operating characteristic. Multiclass classifier performance was high (AF class: 96.6%; PAC class: 95.3%; PVC class: 92.8%; NSR class: 97.4%) and preferable to binary classifiers. Thus, our CNN AF multiclass classifier proved to be an efficient tool for AF discrimination from physiological and pathological confounders.

Keywords: Neural networks and support vector machines in biosignal processing and classification
Abstract: The fetal heart rate (fHR) plays an important role in the determination of the good health of the fetus. Beside the traditional Doppler ultrasound technique, non-invasive fetal electrocardiography (fECG) has become an interesting alternative. However, extracting clean fECG from abdominal ECG (aECG) recordings is a challenging task due to the presence of the maternal ECG component and various noise sources. In this context, we propose a deep residual convolutional autoencoder network trained on synthetic aECG simulations followed by a transfer learning phase on real aECG recordings to extract the cleanest fECG. Afterwards, we propose to use a non-negative matrix factorization based approach on the obtained fECG to estimate the fHR. Our method is evaluated on three publicly available databases demonstrating that it can provide significant performance improvement against comparative methodologies.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Data mining and big data methods - Machine learning and deep learning methods
Abstract: The non-invasive fetal electrocardiogram (FECG) derived from abdominal surface electrodes has been widely used for fetal heart rate (FHR) monitoring to assess fetal well-being. However, the accuracy of FECG-based FHR estimation heavily depends on the quality of FECG signal itself, which can generally be affected by several interference sources such as maternal heart activities and fetal movements. Hence, FECG signal quality assessment (SQA) is an essential task to improve the accuracy of FHR estimation by removing or interpolating low-quality FECG signals. In recent research, various SQA methods based on supervised learning have been proposed. Although these methods could perform accurate SQA, they require large labeled datasets. Nevertheless, the labeled datasets for the FECG SQA are very limited. In this paper, to address this limitation, we propose an unsupervised learning-based SQA method for identifying high and low-quality FECG signal segments. Specifically, a fully convolutional network (FCN)-based autoencoder (AE) is trained for reconstructing a spectrogram derived from FECG. An AE-based feature related to reconstruction error is then calculated to identify high and low-quality FECG segments. In addition, entropy-based features, statistical features, and ECG signal quality indices (SQIs) are also extracted. The high and low-quality segments are identified by feeding the extracted features into self-organizing map (SOM). The experimental results showed that our proposal achieved an accuracy of 98% in high and low-quality signal classification.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Signal pattern classification
Abstract: Automatic classification of cardiac abnormalities is becoming increasingly popular with the prevalence of ECG recordings. Many signal processing and machine learning algorithms have shown the potential to identify cardiac abnormalities accurately. However, most of these methods heavily rely on a large amount of relatively homogeneous datasets. In real life, chances are that there is not enough data for a specific category, and regular deep learning may fail in this scenario. A straightforward intuition is to use the knowledge learned from previous data to solve the problem. This idea leads to learning-to-learn: extrapolating the knowledge accumulated from the old dataset and using it in a different but somewhat related dataset. In this way, we do not need to have considerable data to learn the new task because the underlying features of the old and new datasets resemble one another. In this paper, a novel machine learning method is introduced to solve the ECG arrhythmia detection problem with a limited amount of data. The proposed method combines the popular techniques of meta-learning and transfer learning. It is shown that our method achieves much higher accuracy in ECG arrhythmia classification with a few data and learns the new task faster than regular deep learning.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Time-frequency and time-scale analysis - Time-frequency analysis, Signal pattern classification
Abstract: Epilepsy is a life-threatening disease affecting millions of people all over the world. Artificial intelligence epileptic predictors offer excellent potential to improve epilepsy therapy. Particularly, deep learning models such as convolutional neural networks (CNN) can be used to accurately detect ictogenesis through deep structured learning representations. In this work, a tiny one-dimensional stacked convolutional neural network (1DSCNN) is proposed based on short-time Fourier transform (STFT) to predict epileptic seizure. The results demonstrate that the proposed method obtains better performance compared to recent state-of-the-art methods, achieving an average sensitivity of 94.44%, average false prediction rate (FPR) of 0.011/h and average area under the curve (AUC) of 0.979 on the test set of the American Epilepsy Society Seizure Prediction Challenge dataset, while featuring a model size of only 21.32kB. Furthermore, after adapting the model to 4-bit quantization, its size is significantly decreased by 7.08x with only 0.51% AUC score precision loss, which shows excellent potential for hardware-friendly wearable implementation.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Signal pattern classification, Time-frequency and time-scale analysis - Time-frequency analysis
Abstract: Mobile and wearable electronics is one of the rapidly developing areas of high technologies, which regularly appear new devices that offer new features for monitoring our health, level of physical exertion and everyday activity. From the point of view of medicine and sports, usually wearable devices are used to solve one of two tasks: round-the-clock continuous heart rate monitoring, and the control of the heart rate during workout. One of the drawbacks of such devices is the calculation errors that are caused by the motion artifacts. The most critical of them are those that arise in some cases when performing high-intensity training's, which are conjugate with intense hands movements, such as sprint running or boxing. In addition, limited resources of wearable devices do not always allow usage of very complex algorithms for processing certain events. In our work, an ultra-lightweight framework for a precise real-time heart rate monitoring during the high intensity physical exercises is developed. The model is a combination of a digital signal processing and deep convolutional and recurrent neural networks approaches. From the personalization point of view, the effect of anthropometric parameters has been studied. The model shows an average mean absolute error of 2.4 ± 2.8 bpm during 5 fold cross-validation on an internal dataset, 2.9 ± 3.4 bpm when evaluated on 12 IEEE SPC subjects, and 4.8 ± 5.3 bpm when evaluated on 24 subjects from Wonkwang University dataset. Such results exceed the current state of the art solutions both in terms of the achieved accuracy of heart rate estimation and consumed computational resources.

Keywords: Time-frequency and time-scale analysis - Wavelets
Abstract: Electrocardiogram (ECG) signal provides a graphical representation of cardiac activity and is the most commonly adopted clinical tool for cardiac abnormalities detection. Heartbeat detection, as the first step in analyzing ECG signals, is required for an accurate diagnosis. Stationary wavelet transform (SWT) as a commonly used algorithm for heartbeat detection has a disadvantage of phase shift regarding the original signal. This work addresses this issue by presenting a new method that incorporates an SWT-based zero-phase filter bank with a voting scheme. Our results indicated that a superior performance in heartbeat detection was achieved from the upper arm compared to conventional SWT with a more accurate localization. We achieved sensitivity (SE) and positive predictive value (PPV) of 0.98 ± 0.04 and 0.95 ± 0.09 with the most distance of 50 ms from the actual heartbeats. The SE and PPV changed to 0.75± 0.15 and 0.73± 0.16, respectively for the distance of 20 ms.

Keywords: Signal pattern classification, Nonlinear dynamic analysis - Biomedical signals, Time-frequency and time-scale analysis - Nonstationary analysis and modeling
Abstract: The role of fetal surveillance for the prediction and timely assessment of fetal distress is widely established. Fetal ECG (fECG) monitoring via wearable devices is a feasible solution for performing continuous monitoring of fetal wellbeing and it has seen a net increase in popularity in recent years. In this paper, we propose a novel adaptation of the Smart AdaptiVe Ecg Recognition (SAVER) algorithm for the detection of fECG in long-duration recordings acquired in clinical as well as unconventional settings. The methodology was trained and tested on 50 recordings of duration 1 hour (59.33±5.54 min) obtained using the Monica AN24 fetal monitor. We validated the performance against the automatic extraction performed by the Monica DK software. Our results show superior reliability of the proposed methodology in extracting fECG and associated estimates of fetal heart rate (fHR).

Keywords: Time-frequency and time-scale analysis - Time-frequency analysis, Physiological systems modeling - Signal processing in physiological systems, Data mining and big data methods - Pattern recognition
Abstract: Photoplethysmogram (PPG) signal is extensively used for deducing health parameters of patients in order to infer about physiological conditions of heart, blood pressure, respiratory patterns, and so on. Such analysis and estimations can be done accurately only on high quality PPG signals with very minimal artifacts. PPG signals collected from fitness grade and smart phone scenarios are prone to muscle artifacts and hence there is a need to assess the signal quality before using the signal. Although there are approaches available in the realm of machine learning and deep learning, they are computationally expensive and may not be suitable for a wearable or edge computing scenario. In this paper, we propose the design of a quality checker to check the quality of the signal that can be directly implemented on edge devices like smartwatch. The algorithm is tested on PPG data collected from wearable, ICU and medical grade devices. In the wearable scenario where the noise levels are very high, our algorithm has performed significantly better with a Fscore of over 0.92. Further we show that by applying the proposed quality checker, the accuracy of the computed heart rate from a smart phone PPG-application significantly improves.

Keywords: Data mining and big data methods - Machine learning and deep learning methods, Data mining and big data methods - Biosignal classification, Neural networks and support vector machines in biosignal processing and classification
Abstract: Preterm babies in the Neonatal Intensive Care Unit (NICU) have to undergo continuous monitoring of their cardiac health. Conventional monitoring approaches are contact-based, making the neonates prone to various nosocomial infections. Video-based monitoring approaches have opened up potential avenues for contactless measurement. This work presents a pipeline for remote estimation of cardiopulmonary signals from videos in NICU setup. We have proposed an end-to-end deep learning (DL) model that integrates a non-learning-based approach to generate surrogate ground truth (SGT) labels for supervision, thus refraining from direct dependency on true ground truth labels. We have performed an extended qualitative and quantitative analysis to examine the efficacy of our proposed DL-based pipeline and achieved an overall average mean absolute error of 4.6 beats per minute (bpm) and root mean square error of 6.2 bpm in the estimated heart rate.

Keywords: General and theoretical informatics - Algorithms, Health Informatics - Informatics for chronic disease management, General and theoretical informatics - Data intelligence
Abstract: Undertreatment or overtreatment of pain will cause severe consequences physiologically and psychologically. Thus, researchers have made great efforts to develop automatic pain assessment approaches based on physiological signals using machine learning techniques. However, state-of-art research mainly focuses on verifying the hypothesis that physiological signals can be used to assess pain. The critical assumption of these studies is that training data and testing data have the same distribution. However, this assumption may not hold in real-life scenarios, for instance, the adoption of machine learning model by a new patient. Such real-life scenarios in which user's data is unlabeled is largely neglected in literature. This study compensates for the rift by proposing an adaptive transfer learning based pain assessment system (ATLAS), a novel adaptive learning system based on the transfer learning algorithm Transfer Components Analysis (TCA) to minimize the distance between training data and unlabeled testing data. Experiments were conducted on BioVid database, and the results showed our approach outperforms three existing traditional machine learning-based approaches and achieves an accuracy just 2.0% below the accuracy with labeled data.

Keywords: General and theoretical informatics - Artificial Intelligence, Sensor Informatics - Physiological monitoring, Health Informatics - Mobile health
Abstract: Passive assessment of obstructive pulmonary disease has gained substantial interest over the past few years in the mobile and wearable computing communities. One of the promising approaches is speech-based pulmonary assessment wherein spontaneous or scripted speech is used to evaluate an individual's pulmonary condition. Recent approaches in this regard heavily rely on accurate speech activity segmentation and specific, hand-crafted features. In this paper, we present an end-to-end deep learning approach for detecting obstructive pulmonary disease. We leveraged transfer learning using a network pre-trained for a different audio-based task, and employed our own additional shallow network on top as a binary classifier to indicate if a given speech recording belongs to an asthma or COPD patient. The additional network was a fully connected neural net with 2 hidden layers, and this was evaluated on two real-world datasets. We demonstrated that the system can identify subjects with obtructive pulmonary disease using their speech with 88.3% precision, 88.8% recall and 88.3% F-1 score using 10-fold cross-validation. The system showed improved performance in identifying the most severely affected subgroup of patients in the dataset, with an average 93.6% accuracy.

Keywords: General and theoretical informatics - Artificial Intelligence, General and theoretical informatics - Machine learning, Health Informatics - eHealth
Abstract: Since the emergence of the COVID-19 pandemic, various methods to detect the illness from cough and speech audio data have been proposed. While many of them deliver promising results, they lack transparency in the form of explanations which is crucial for establishing trust in the classifiers. We propose CoughLIME which extends LIME to explanations for audio data, specifically tailored towards cough data. We show that CoughLIME is capable of generating faithful sonified explanations for COVID-19 detection. To quantify the performance of the explanations generated for the CIdeR model, we adopt pixel flipping to audio and introduce a novel metric to assess the performance of the XAI classifier. CoughLIME achieves a DeltaAUC of 19.48,% generating explanations for CIdeR's predictions.

Keywords: General and theoretical informatics - Artificial Intelligence, Health Informatics - Precision medicine, General and theoretical informatics - Knowledge modeling
Abstract: In this work, an attempt has been made to discriminate drug with blood brain barrier (BBB) permeability using clinical phenotypes and extreme gradient boosting (XGBoost) methods. For this, the drug name and their clinical phenotypes namely side effects and indications are obtained from public available database. Prominent clinical phenotypes are selected using genetic algorithm and binary particle swarm optimization. Four machine algorithms namely k-Nearest Neighbours, support vector machines, rotation forest and XGBoost are used for classification of BBB drugs. The result show that the proposed clinical phenotypes based features are able to distinguish drugs with BBB permeability. The maximum number of clinical phenotypes (69%) is reduced by BPSA and GA for classification. The XGBoost method is found to be most accurate (F1=89.7%) is discriminating drugs with BBB permeability. The proposed approach are found to be capable of handling multi-parametric characteristics of the drugs. Particularyl, the combination of XGBoost with combination of side effects and indications could be used for precision medicine applications

Keywords: General and theoretical informatics - Artificial Intelligence, Health Informatics - Patient tracking, Health Informatics - Information technologies for the management of patient safety and clinical outcomes
Abstract: Abstract—This study aims to use computers to detect and recognize ventilation objects (masks and tubes) and their positions on the patient’s face. We created a YOLO model and a transfer learning (TL) model to perform this computer vision task. The development processes and comparison of performance are described in this paper. The TL model had a better performance (98%) compared to the YOLO model (93%). Clinical Relevance— Healthcare providers and researchers interested in the field of computer vision applied in medicine, specifically automatic object detection using video streams or real-time video streaming may benefit from findings reported.

Keywords: General and theoretical informatics - Causality analysis and case-based reasoning, Health Informatics - Outcome research, Public Health Informatics - Epidemiological modeling
Abstract: Propensity score matching (PSM) is a technique used in retrospective investigation of cohort matching as an alternative approach to the prospective matching that is typically used by a randomized control trial (RCT). The process of selecting untreated cases that are the best match tothe treated cases is the focus of this research. We created aPSM package for the python environment, termed PsmPy, to carry out this task. The PsmPy package debuted and proposed here is based on a logistic regression logit score where a match is selected using k-nearest neighbors (k-NN). Additional plotting and arguments are available to the user and are also described. To benchmark our method, we compared it with the existing R package, MatchIt, and evaluated our covariates’ residual effect sizes with respect to the treatment condition before and after matching. Using a Mann-Whitney statistical test, we showed that our method significantly outperformed MatchIt in cohort matching (U=49, p<0.0001)when comparing residual effect sizes of the covariates. The PsmPy demonstrated a 10-fold average improvement in residual effect sizes amongst covariates when compared with the package MatchIt, suggesting that it is a viable alternative for use in propensity matching studies

Keywords: General and theoretical informatics - Data privacy, General and theoretical informatics - Machine learning, General and theoretical informatics - Security and authentication
Abstract: Machine learning is playing an increasingly critical role in health science with its capability of inferring valuable information from high-dimensional data. More training data provides greater statistical power to generate better models that can help decision-making in healthcare. However, this often requires combining research and patient data across institutions and hospitals, which is not always possible due to privacy considerations. In this paper, we outline a simple federated learning algorithm implementing differential privacy to ensure privacy when training a machine learning model on data spread across different institutions. We tested our model by predicting breast cancer status from gene expression data. Our model achieves a similar level of accuracy and precision as a single-site non-private neural network model when we enforce privacy. This result suggests that our algorithm is an effective method of implementing differential privacy with federated learning, and clinical data scientists can use our general framework to produce differentially private models on federated datasets. Our framework is available at https://github.com/gersteinlab/idash20FL.

Keywords: General and theoretical informatics - Data privacy, General and theoretical informatics - Machine learning, Health Informatics - Behavioral health informatics
Abstract: As the most common neurodegenerative disease among older adults, Alzheimer's disease (AD) would lead to loss of memory, impaired language and judgment, gait disorders, and other cognitive deficits severe enough to interfere with daily activities and significantly diminish quality of life. Recent research has shown promising results in automatic AD diagnosis via speech, leveraging the advances of deep learning in the audio domain. However, most existing studies rely on a centralized learning framework which requires subjects' voice data to be gathered to a central server, raising severe privacy concerns. To resolve this, in this paper, we propose the first federated-learning-based approach for achieving automatic AD diagnosis via spontaneous speech analysis while ensuring the subjects' data privacy. Extensive experiments under various federated learning settings on the ADReSS challenge dataset show that the proposed model can achieve high accuracy for AD detection while achieving privacy preservation. To ensure fairness of the model performance across clients in federated settings, we further deploy fair aggregation mechanisms, particularly q-FEDAvg and q-FEDSgd, which greatly reduces the algorithmic biases due to the data heterogeneity among the clients.

Keywords: General and theoretical informatics - Data privacy, General and theoretical informatics - Deep learning and big data to knowledge, Sensor Informatics - Physiological monitoring
Abstract: Electrocardiogram (ECG) signals provide rich information on individuals' potential cardiovascular conditions and disease, ranging from coronary artery disease to the risk of a heart attack. While health providers store and share these information for medical and research purposes, such data is highly vulnerable to privacy concerns, similar to many other types of healthcare data. Recent works have shown the feasibility of identifying and authenticating individuals by using ECG as a biometric due to the highly individualized nature of ECG signals. However, to the best of our knowledge, there does not exist a method in the literature attempting to de-identify ECG signals. In this paper, to address this privacy protection gap, we propose a Generative Adversarial Network (GAN)-based framework for de-identification of ECG signals. We leverage a combination of a standard GAN loss, an Ordinary Differential Equations (ODE)-based, and identity-based loss values to train a generator that de-identifies a ECG signal while preserving structure the ECG signal and information regarding the target cardio vascular condition. We evaluate our framework in terms of both qualitative and quantitative metrics considering different weightings over the above-mentioned losses. Our experiments demonstrate the efficiency of our framework in terms of privacy protection and ECG signal structural preservation.

Keywords: General and theoretical informatics - Data privacy, General and theoretical informatics - Machine learning, Health Informatics - Behavioral health informatics
Abstract: Mental health disorders, such as depression, affect a large and growing number of populations worldwide, and they may cause severe emotional, behavioral and physical health problems if left untreated. As depression affects a patient's speech characteristics, recent studies have proposed to leverage deep-learning-powered speech analysis models for depression diagnosis, which often require centralized learning on the collected voice data. However, this centralized training requiring data to be stored at a server raises the risks of severe voice data breaches, and people may not be willing to share their speech data with third parties due to privacy concerns. To address these issues, in this paper, we demonstrate for the first time that speech-based depression diagnosis models can be trained in a privacy-preserving way using federated learning, which enables collaborative model training while keeping the private speech data decentralized on clients' devices. To ensure the model's robustness under attacks, we also integrate different FL defenses into the system, such as norm bounding, differential privacy, and secure aggregation mechanisms. Extensive experiments under various FL settings on the DAIC-WOZ dataset show that our FL model can achieve high performance without sacrificing much utility compared with centralized-learning approaches while ensuring users' speech data privacy.

Keywords: General and theoretical informatics - Data storage, Health Informatics - Clinical information systems, Health Informatics - Health data acquisition, transmission, management and visualization
Abstract: In this work we present the creation of a large structured database of CardioTocoGraphic (CTG) recordings, starting from a raw dataset containing tracings collected between 2013 and 2021 by the medical team of the University Hospital Federico II of Naples. The aim of the work is to provide a big, structured database of real clinical cardiotocographic data, useful for subsequent processing and analysis through state-of-the-art methods, in particular Deep Learning Methods. This organized dataset could lead to an increase of the diagnostic accuracy of CTG analysis in the discrimination of healthy and unhealthy fetuses.

Keywords: General and theoretical informatics - Deep learning and big data to knowledge, General and theoretical informatics - Machine learning, General and theoretical informatics - Data privacy
Abstract: The generation of synthetic genomic sequences using neural networks has potential to ameliorate privacy and data sharing concerns and to mitigate potential bias within datasets due to under-representation of some population groups. However, there is not a consensus on which architectures, training procedures, and evaluation metrics should be used when simulating single nucleotide polymorphism (SNP) sequences with neural networks. In this paper, we explore the use of Generative Moment Matching Networks (GMMNs) for SNP simulation, we present some architectural and procedural changes to properly train the networks, and we introduce an evaluation scheme to qualitatively and quantitatively assess the quality of the simulated sequences.

Keywords: Cardiac mechanics, structure & function - Ventricular mechanics, Cardiac mechanics, structure & function - Cardiac structure from imaging, Cardiac mechanics, structure & function - Heart failure
Abstract: Ejection fraction (EF) is considered to provide clinically useful information. Despite its enormous popularity, with more than 75,000 citations in PubMed, only few studies have traced the origin(s) of its foundation. This fact is surprising, as there are perhaps more papers published that criticize EF, than the number of publications that actually provide a solid (mathematical) basis for its alleged applicability. EF depends on two volume determinations, namely end-systolic volume (ESV) and end-diastolic volume (EDV). EF is defined as 1-ESV/EDV, yielding a metric without physical units. Previously we formulated a robust analytical expression for the nonlinear connection between EF and ESV. Here we extend that approach by providing a formula to illustrate that EF is strongly associated with half the sum (HS) of ESV and EDV. HS is not new, but forms a major component in the recently introduced Global Function Index. For 420 heart failure (HF) patients we found for left ventricular angio data: R(ESV, EDV) = 0.92, R(EF, ESV) = -0.90, and R(EF, HS) = -0.65. For echo (33 HF patients stages A, B, C and D): R(EF, HS) = -0.82. For the right atrium (CMRI in 21 acute myocardial infarction patients): R(EF, HS)=-0.65. For the left atrium (N=86) R (EF, HS)=-0.46. ESV indicates the level to which the ventricle is able to squeeze blood out of the cavity via pressure build-up. In contrast, EF refers to relative volume changes, not to the mechanism of pumping action. We conclude that for each cardiac compartment EF borrows its acclaimed attractiveness from the fact that for a wide patient spectrum the ESV and EDV correlate in a fairly linear manner. Attractiveness of EF features a straightforward mathematical derivation, rather than reflecting underlying physiology.

Keywords: Cardiovascular and respiratory system modeling - Cardiovascular control models, Cardiac mechanics, structure & function - Heart failure, Cardiovascular, respiratory, and sleep devices - Therapeutics
Abstract: Acute heart failure is caused by various factors and requires multiple drug therapies to remedy underlying causes. Due to the complexity of pharmacologic effects of cardiovascular agents, few studies have theoretically addressed the multidrug optimization problem. This paper proposes a drug infusion system for acute heart failure that controls cardiovascular performance metrics (cardiac output, left atrial pressure, and mean arterial pressure) within desired ranges as dictated by the cardiovascular parameters (systemic vascular resistance, cardiac contractility, heart rate, and stressed blood volume). The key to our system design is modeling and controlling cardiovascular parameters to yield the desired cardiovascular metrics. A `tailored drug infusion' technique controls parameters by solving the optimization problem in order to conquer the complexity of multi-dependencies and the different dosage limits among multiple drugs. A `cardiovascular space mapping' technique identifies the desired parameters from the desired metrics by deriving the analytical solutions of the metrics as functions of the parameters. To facilitate clinical discussions, parameters were set to realistic values in 5,600 simulated patients. Our results showed not only that the optimized drug combinations and dosages controlled the cardiovascular metrics to within the desired ranges, but also that they mostly corresponded to the recommended clinical use guidelines. An additional value of our system is that it proactively predicts the limitations of the tailored drug therapy, which supports the clinical decision of pivoting to alternative treatment strategies such as mechanical circulatory support.

Keywords: Cardiovascular and respiratory signal processing - Pulse transit time, Cardiovascular and respiratory signal processing - Lung Sounds
Abstract: A large portion of the elderly population are affected by cardiovascular diseases. The early prognosis of cardiomyopathies is still a challenge. The aim of this study was to classify cardiomyopathy patients by their etiology in function of significant indexes extracted from the characterization of the recurrence plot of the systems involved. Thirty-nine cardiomyopathy patients (CMP) classified as ischemic (ICM – 24 patients) and dilated (DCM – 15 patients) were considered. In addition, thirty-nine control subjects (CON) were used as reference. The beat-to-beat (BBI) time series, from the electrocardiographic signal, the systolic (SBP), and diastolic (DBP) time series, from the blood pressure signal, and the respiratory time (FLW) from the respiratory flow signal, were extracted. The recurrence plot from each signal considered were calculated and characterized by a total of 12 indexes. The best classifiers were used to build support vector machine models. The optimal model to classify ICM versus DCM patients achieved 92.3% accuracy, 95.8% sensitivity, and 86.6% specificity. When comparing CMP patients and CON subjects, the best model achieved 85.8% accuracy, 92.3% sensitivity, and 80.1% specificity. Our results suggest a more deterministic behavior in DCM patients.

Clinical Relevance— This study explores the recurrence plot for the classification of ICM and DCM patients.

Keywords: Cardiac mechanics, structure & function - Ventricular mechanics, Cardiovascular and respiratory signal processing - Non-linear cardiovascular or cardiorespiratory relations, Cardiac mechanics, structure & function - Cardiac structure from imaging
Abstract: Ventricular pump function is often characterized by the (non)linear end-systolic pressure-volume relationship (ESPVR). For each working point on that curve the tangent along with the intercept (Vo) reflect contractile state. Vo on the abscissa is an extrapolated point without physiological meaning, and may be negative. To obtain positive values for the intercept, investigators often choose a non-zero pressure level. Although this preference is mathematically sound, we demonstrate that statistical evaluations may yield different results, depending on the pressure level selected. Published data on 17 cardiac patients representing three diagnostic groups were analyzed, showing dicrotic notch pressure based values -14

Keywords: Cardiovascular regulation - Heart rate variability, Cardiovascular regulation - Baroreflex, Pulmonary and critical care - Bioengineering applications in Intensive care
Abstract: Fluid administration is one of the most common therapies performed on intensive care patients. However, no clinical evidence is available to establish optimal strategies for fluid management as well as characterizing the effects on the cardiovascular system after therapy initiation. Moreover, fluid overload showed a correlation with worse clinical outcomes.

This study aims at characterizing the response to the fluid intervention of intensive care unit patients. We extracted a population of 57 subjects with available electrocardiogram and arterial blood pressure recordings from the MIMIC-III database and evaluated the induced changes in cardiovascular and autonomic indices.

We compare autonomic indices obtained from a statistical model of heartbeat dynamics before and after the intervention. Results show significant differences in RR interval, blood pressure, autonomic and Baroreflex activities up to 60 minutes after fluid administration. Specifically, we observed a median increase in RR interval, Baroreflex activity, and overall activity both in pressure and RR time series, as well as a reduction in systolic blood pressure. Specifically, a subgroup of survived patients shows an imbalance toward sympathetic activity, whereas non-survivors have a persistent vagal state after fluid administration.

Clinical relevance - The observed differences in autonomic response after fluid administration, together with the assessment of their correlation with patients' mortality, paves the way for the inclusion of heart rate variability indices as markers for assessing fluid responsiveness as associated with ICU patients' state.

Keywords: Pulmonary and critical care - Bioengineering applications in Intensive care, Vascular mechanics and hemodynamics - Arterial pressure in cardiovascular disease, Cardiovascular and respiratory signal processing - Cardiovascular signal processing
Abstract: We investigated whether a statistical model used previously to predict hypotension from mean arterial pressure (MAP) time series analysis could predict hypertension. We performed a retrospective analysis of minute-by-minute MAP records from two cohorts of intensive care unit (ICU) patients. The first cohort was comprised of surgical and medical ICUs while the second cohort was comprised of acute spinal cord injury (ASCI) patients in a neurological ICU. At each time point with physiological MAP, time series analysis was used to predict the median MAP for the subsequent 20 min. This method was used to predict hypertensive episodes, i.e., intervals of 20 or more minutes where at least half of the MAP measurements were > 105 mmHg.Advance prediction of hypertensive episodes was similar in the two cohorts (69.15% vs. 82.61%, respectively), as was positive predictive value of the hypertension predictions (67.42% vs. 71.57%). The results suggest that the methodology may be usable for predicting hypertension from time-series analysis of MAP.

Keywords: Artificial organs (including heart, kidney, liver, pancreas, retina), Diagnostic devices - Physiological monitoring
Abstract: The aim of this study was to measure intraperitoneal volume (IPV) and ultrafiltration volume (UFV) by monitoring the abdominal resistance using segmental bioimpedance analysis (SBIA, Hydra 4200). Twenty peritoneal dialysis (PD) patients were studied during a fill with 2 L of 2.5% glucose peritoneal dialysate solution. UFVDrain (g) was measured as weight difference between fill and drain dialysate volumes. Ultrafiltration volume (UFVSBIA; ml) and absorption volume were calculated from the IPV curve derived by SBIA. UFVSBIA correlated with UFVDrain (R2=0.98, p<0.0001). This study may provide actionable clinical insights and help clinicians to better understand the function of the peritoneal membrane in individuals. Clinical Relevance— This technique may support personalized medicine by aiding the prescription of PD therapy on a patient-level.

Keywords: Physiological monitoring & diagnistic devices - Pulmonary disease, Wearable or portable devices for vital signal monitoring, Ambulatory Diagnostic devices - Point of care technologies
Abstract: In this work we present the development of a multi-frequency electrical impedance pneumography (EIP) system based on a portable acquisition device and a mobile platform. This design is intended as an upgrade to our previous device for clinical use in the screening of patients with pulmonary diseases. The acquisition device uses the bioimpedance analog front end MAX30001, a mux/demux stage, Bluetooth 4.0 communication and an ESP32 microcontroller unit. It generates an excitation current of 8 uApp in a range of selectable frequencies from 1 kHz to 130 kHz. The mobile platform provides a real-time respiration signal at 64 S/s and allows the configuration of the device. Results show better accuracy in higher frequencies, which are the most common in these applications, ensuring the feasibility of the system for use in humans. Some hardware constraints related to EIP integrated circuits are discussed, an their effect in the signal acquisition.

Keywords: Clinical engineering - Device safety and efficacy evaluation (electrical safety, electromagnetic compatibility and immunity)
Abstract: It is difficult to electrocute (induce ventricular fibrillation) with capacitive discharge shocks. With small capacitance values, the high voltages required for the necessary charge are rarely seen in industrial situations (e.g. electric vehicle charging stations). On the other hand, with large capacitance values, the discharge time is so great that the shock couples inefficiently with the cardiac cells. The update to IEC 60479-2 sets the C1 “mostly-safe” charge limit of 3 mC for a short “impulse function” pulse. We calculated the equivalent capacitor stored charge for an arbitrary capacitance value using the simple single membrane time constant model for the cardiac response. The peak membrane response was set equal to that of the 3 mC impulse function response to calculate the safe values for stored charge, voltage, and energy. The total stored charge, per se, cannot be used simplistically to estimate the danger of a capacitive discharge shock. A capacitive-discharge shock cannot be accurately compared to a rectangular shock with a duration equal to the shock time constant. The greater the capacitance, the larger the fraction of wasted charge in coupling to the heart and thus the shorter equivalent duration compared to the shock time constant. For a capacitive discharge shock this translates to a stored charge of 3 mC increasing up to 9 mC for a 10 µF capacitor using the assumed 575 Ω load for an electric-vehicle (EV) charging station. In the area of interest for 1 - 10 µF, the safe voltage ranges from 1300 to 4700 V, which includes the 1500-VDC scope of EV charger standard IEC 61851-23. For C > 100 µF, the voltage asymptote is 700 V.

Keywords: Robotic-aided therapies - Targeted therapy systems, Neuromuscular systems - Stroke therapy devices / technologies, Ambulatory Therapeutic Devices - Personalized therapeutic devices and emergency response systems
Abstract: This paper discusses the design, construction, and characteristics of a six degree of freedom (6-DoF) robotic upper limb stroke rehabilitation device. The device is primarily designed to be used by stroke survivors suffering from hemiparesis, post stroke therapy. The proposed device is aimed to substitute upper limb exoskeletons or end-effector robotic arm based rehabilitation systems, which are generally bulky, expensive due to customized design, and are used only in clinical settings. The device is capable of aiding patients perform rehabilitation exercises involving abduction, adduction, flexion, and extension movements for the wrist, shoulder joints and flexion, extension, pronation, supination for the elbow joint. The device has a mobile base and uses differential drive principles for movement. The device provides an end-effector plate having force-sensitive resistors (FSRs) on which the patient can rest their hand. The mobile base of the rehabilitation system enables it to aid for a greater range of movements when compared to other end-effector-based upper limb rehabilitation devices. Additionally, a camera is mounted on top of the 6-DoF robotic system to enable finger tracking from a remote system using the MediaPipe framework, and measure the hand instability metric over time to assess patient's performance.

Clinical relevance — The 6-DoF rehabilitation system is capable of aiding different range of motions for upper limb. The low-cost generic system is applicable for different physical personalities, enabling quick adoption for large patient populations who are in need of rehabilitation systems. The 6-DoF system is aimed towards clinical, and domestic usage by patients.

Keywords: Ambulatory Diagnostic devices - Point of care technologies, Clinical engineering -Verification and validation of diagnostic & therapeutic systems / technologies, Physiological monitoring & diagnistic devices - Heart failure
Abstract: Abstract— With heart failure (HF) and renal malfunction becoming global public health issues, there is an urgent need to monitor diseases at home or in the community. Point-of-care testing (POC) would shorten the patients waiting time compared with laboratory molecular analysis. This work evaluates two types of gold nanomaterials, and two assay protocols, to develop a lateral flow (LF) system for Cystatin C (CysC) quantification. Of the protocols, the ‘delayed-release’ shows the alleviation of the hook effects with 1% BSA running buffer (RB), albeit at increased complexity with three steps of washing. The standard method with sample dilution (1: 150 sample dilution for GNPs, and 1:10 sample dilution for GNRs) can ensure the clinical range detection of CysC as 1 mg/L with partial LF assays. GNPs have stronger optical signal intensity compared with GNRs and developed full LF assays with GNPs require 1:1.5 sample dilution in recombinant CysC detection. The ideal sample dilution ratio is different for partial and full LF assays.

Clinical Relevance— This work is the basis of future work that will use LF devices for human serum/plasma monitoring to assess kidney function related to heart failure during medication. The specificity, sensitivity, and limit of detection will be validated via a clinical trial before potential clinical use.

Keywords: Computer modeling for treatment planning, Glucose remote monitoring devices and technologies, Diagnostic devices - Physiological monitoring
Abstract: Over the years and with the help of technology, the daily care of type 1 diabetes has been improved significantly. The increased adoption of continuous glucose monitoring, the continuous subcutaneous insulin injection and the accurate behavioral monitoring mHealth solutions have contributed to this phenomenon. In this study we present a mobile application for automated dietary assessment of Mediterranean food images as part of the GlucoseML system. Based on short-term predictive analysis of the glucose trajectory, GlucoseML is a type-1 diabetes self-management system. A computer vision approach is used as main part of the GlucoseML dietary assessment system calculating food carbohydrates, fats and proteins, relying on: (i) a deep learning subsystem for food image classification, and (ii) a 3D food image reconstruction subsystem for the volume estimation of food. The deep learning subsystem achieves 82.4% and 97.5% top-1 and top-5 accuracy, respectively, for food image classification while the subsystem for volume estimation of food achieves a mean absolute percentage error 10.7% for the four main categories of MedGRFood dataset.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Image feature extraction, Magnetic resonance imaging - Other organs
Abstract: Prostate cancer is one of the most common cancers in men, with symptoms that may be confused with those caused by benign prostatic hyperplasia. One of the key aspects of treating prostate cancer is its early detection, increasing life expectancy and improving the quality of life of those patients. However, the tests performed are often invasive, resulting in a biopsy. A non-invasive alternative is the magnetic resonance imaging (MRI)-based PI-RADS v2 classification. The aim of this work was to find objective biomarkers that allow the PI-RADS classification of prostate lesions using a radiomics approach on Multiparametric MRI. A total of 90 subjects were analyzed. From each segmented lesion, 609 different texture features were extracted using five different statistical methods. Two feature selection methods and eight multiclass predictive models were evaluated. This was a multiclass study in which the best AUC result was 0.7442 ± 0.0880, achieved with the Naïve Bayes model using a subset of 120 features. Valuable results were also obtained using the Random Forests model, obtaining an AUC of 0.7394 ± 0.0965 with a lower number of features (52).

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Image classification
Abstract: Breast cancer is one of the leading causes of death among women. Early prediction of breast cancer can significantly improve the survival rates. Breast density was proven as a reliable risk factor. Deep learning models can learn subtle cues in the mammogram images. CNN models were recently shown to improve the risk discrimination in full-field mammograms. This study aims to improve risk prediction models using bilateral analysis. Bilateral analysis is the process of comparing two breasts to verify presence of anomalies. We developed a Siamese neural network to leverage the bilateral information and asymmetries between the two mammograms of the same patient. We tested our model on 271 patients and compared the results of our Siamese model against the traditional unilateral CNN model. Our results showed AUCs of 0.75 and 0.70 respectively (p = 0.0056). The Siamese model also exhibits higher sensitivity, specificity, precision, and false positive rate with values of 0.68, 0.69, 0.71, 0.31 respectively. While the CNN values were 0.61, 0.66, 0.67, 0.34 respectively. We merged both models by two techniques using pre-trained weights and weighted voting ensemble. The merging technique boosted the AUC to 0.78. The results suggest that bilateral analysis can significantly improve the risk discrimination.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, X-ray imaging applications, Machine learning / Deep learning approaches
Abstract: It is generally believed that vision transformers (ViTs) require a huge amount of data to generalize well, which limits their adoption. The introduction of data-efficient algorithms such as data-efficient image transformers (DeiT) provided an opportunity to explore the application of ViTs in medical imaging, where data scarcity is a limiting factor. In this work, we investigated the possibility of using pure transformers for the task of chest x-ray abnormality detection on a small dataset. Our proposed framework is built on a DeiT structure benefiting from a teacher-student scheme for training, with a DenseNet with strong classification performance as the teacher and an adapted ViT as the student. The results show that the performance of transformers is on par with that of convolutional neural networks (CNNs). We achieved a test accuracy of 92.2% for the task of classifying chest x-ray images (normal/pneumonia/COVID-19) on a carefully selected dataset using pure transformers. The results show the capability of transformers to accompany or replace CNNs for achieving state-of-the-art in medical imaging applications. The code and models of this work are available at https://github.com/Quantimb-Lab/DeiT_Covid.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Deformable registration, Functional image analysis
Abstract: The overriding clinical and academic challenge that inspires this work is the lack of a universally accepted, objective, and feasible method of measuring facial deformity; and, by extension, the lack of a reliable means of assessing the benefits and shortcomings of craniofacial surgical interventions. We propose a machine learning-based method to create a scale of facial deformity by producing numerical scores that reflect the level of deformity. An object detector that is constructed using a cascade function of Haar features has been trained with a rich dataset of normal faces in addition to a collection of images that does not contain faces. After that, the confidence score of the face detector was used as a gauge of facial abnormality. The scores were compared with a benchmark that is based on human appraisals obtained using a survey of a range of facial deformities. Interestingly, the overall Pearson's correlation coefficient of the machine scores with respect to the average human score exceeded 0.96.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Image analysis and classification - Digital Pathology, Image feature extraction
Abstract: The classification of cells extracted from Pap-smears is in most cases done using neural network architectures. Nevertheless, the importance of features extracted with digital image processing is also discussed in many related articles. Decision support systems and automated analysis tools of Pap-smears often use these kinds of manually extracted, global features based on clinical expert opinion. In this paper, a solution is introduced where 29 different contextual features are combined with local features learned by a neural network so that it increases classification performance. The weight distribution between the features is also investigated leading to a conclusion that the numerical features are indeed forming an important part of the learning process. Furthermore, extensive testing of the presented methods is done using a dataset annotated by clinical experts. An increase of 3.2% in F1-Score value can be observed when using the combination of contextual and local features.

Keywords: Magnetic resonance imaging - Parallel MRI, Image reconstruction - Fast algorithms, Image reconstruction and enhancement - Machine learning / Deep learning approaches
Abstract: Channel suppression can reduce the redundant information in multiple channel receiver coils and accelerate reconstruction speed to meet real-time imaging requirements. The principal component analysis has been used for channel suppression, but it is difficult to be interpreted because all channels contribute to principal components. Furthermore, the importance of interpretability in machine learning has recently attracted increasing attention in radiology. To improve the interpretability of PCA-based channel suppression, a sparse PCA method is proposed to reduce the most coils’ loadings to be zero. Channel suppression is formulated as solving a nonlinear eigenvalue problem using the inverse power method instead of the direct matrix decomposition. Experimental results of in vivo data show that the sparse PCA-based channel suppression not only improves the interpretability with sparse channels, but also improves reconstruction quality compared to the standard PCA-based reconstruction with the similar reconstruction time.

Keywords: Magnetic resonance imaging - Parallel MRI, Magnetic resonance imaging - MRI RF coil technology
Abstract: This paper demonstrates a rapid B1 field benchtop measurement system that is independent of an MR scanner and network analyzer. This system can be used to obtain radiofrequency (B1 field) strength distribution plots of multiple 2D slices (with an extension to 3D) of a liquid cylindrical phantom for multi-element phased arrays used in MRI. The system can be used in three modes- element, phased array, and multiple fixed point pattern measurement. These modes are demonstrated for a 7T 1H eight-channel dipole array and a corn-syrup based phantom. The system can measure complex phase and amplitude measurements from up to 8 elements in the first mode one or 8 different phase settings in the second mode at a rate of approximately 37 positions per minute, allowing a full 2D B1 mapping for 1303 points in 33.05 minutes. The scan patterns obtained using this setup are compared to the ones obtained using an HP network analyzer and simulations. This work can be extended to measure the E field, SAR and upon increasing the speed of measurement, could be used for applications such as Transmit SENSE.

Keywords: Magnetic resonance imaging - Pulse sequence, Image reconstruction and enhancement - Compressed sensing / Sampling
Abstract: Variable density sampling of the k-space in MRI is an integral part of trajectory design. It has been observed that data-driven trajectory design methods provide a better image reconstruction as compared to trajectories obtained from a fixed or a parametric density function. In this paper, a data-driven strategy has been proposed to obtain non-Cartesian continuous k-space sampling trajectories for MRI under the compressed sensing framework (greedy non-Cartesian (GNC)). A stochastic version of the algorithm (stochastic greedy non-Cartesian (SGNC)) is also proposed that reduces the computation time. We compare the proposed trajectory with a traveling salesman problem (TSP)-based trajectory and an echo planar imaging-like trajectory obtained by a greedy method called stochastic greedy-Cartesian (SGC) algorithm. The training images are taken from knee images of the fastMRI dataset. It is observed that the proposed algorithms outperform the TSP-based and the SGC trajectories for similar read-out times.

Keywords: Magnetic resonance imaging - Pulse sequence
Abstract: With sound pressure levels reaching up to 130 dB, acoustic noise in Magnetic Resonance Imaging (MRI) is one of the main sources of patient discomfort in otherwise one of the safest medical imaging modalities. In this work, a noise prediction-based approach, termed predictive noise cancelling (PNC), is applied, for the first time, to suppress noise in MRI. In PNC the noise from the scanner gradient coils is predicted based on linear time-invariant models, which relate the individual gradient coil (X, Y and Z) input to the acoustic noise output. A model setup was constructed of a custom speaker box and MRI-compatible microphone to demonstrate live noise reduction. Additional tuning steps, including output channel equalization and clock mismatch correction, were performed to maximize noise reduction. A calibration sequence was designed to determine the model and tuning parameters. Analysis of actual scanner noise shows an upper limit of 21 dB noise reduction with the proposed linear model. For the components of a clinical example sequence, the setup demonstrated in-bore live noise reduction of up to 10 dB (7.01 ± 0.31 dB, 6.42 ± 2.04 dB and 9.28 ± 0.26 dB for X, Y and Z gradient coils respectively) in the presence of system imperfections. The results indicate promising noise attenuation without the need to modify scanner hardware or compromises in acquisition speed or quality. This has potential to substantially and cost effectively improve patient comfort in clinical MRI.

Keywords: Magnetic resonance imaging - Dynamic contrast-enhanced MRI, Image reconstruction and enhancement - Machine learning / Deep learning approaches, Magnetic resonance imaging - Parallel MRI
Abstract: Dynamic contrast enhanced (DCE) MRI acquires a series of images following the administration of a contrast agent, and plays an important clinical role in diagnosing various diseases. DCE MRI typically necessitates rapid imaging to provide sufficient spatio-temporal resolution and coverage. Conventional MRI acceleration techniques exhibit limited image quality at such high acceleration rates. Recently, deep learning (DL) methods have gained interest for improving highly-accelerated MRI. However, DCE MRI series show substantial variations in SNR and contrast across images. This hinders the quality and generalizability of DL methods, when applied across time frames. In this study, we propose signal intensity informed multi-coil MRI encoding operator for improved DL reconstruction of DCE MRI. The output of the corresponding inverse problem for this forward operator leads to more uniform contrast across time frames, since the proposed operator captures signal intensity variations across time frames while not altering the coil sensitivities. Our results in perfusion cardiac MRI show that high-quality images are reconstructed at very high acceleration rates, with substantial improvement over existing methods.

Keywords: Image reconstruction and enhancement - Machine learning / Deep learning approaches, Brain imaging and image analysis, Magnetic resonance imaging - MR neuroimaging
Abstract: Mental disorders such as schizophrenia have been challenging to characterize due in part to their heterogeneous presentation in individuals. Most studies have focused on identifying groups differences and have typically ignored the heterogeneous patterns within groups. Here we propose a novel approach based on a variational autoencoder (VAE) to interpolate static functional network connectivity (sFNC) across individuals, with group-specific patterns between schizophrenia patients and controls captured simultaneously. We then visualize the original sFNC in a 2D grid according to the samples in the VAE latent space. We observe a high correspondence between the generated and the original sFNC. The proposed framework facilitates data visualization and can potentially be applied to predict the stage that a subject falls within a disorder continuum as well as characterize individual heterogeneity within and between groups.

Keywords: Motor learning, neural control, and neuromuscular systems, Neuromuscular systems - Postural and balance, Neuromuscular systems - Learning and adaption
Abstract: During object manipulation, our sensorimotor system needs to represent the object’s dynamics in order to better control it. This is especially important in the case of grip force control where small forces can cause the object to slip from our fingers, and excessive forces can cause fatigue or even damage the object. While the tradeoff between these two constraints is clear for stable objects, such as lifting a soda can, it is less clear how the sensorimotor system adjusts the grip force for unstable objects. For this purpose, we measured the change in the grip force of individual human participants while they stabilize five different lengths of an inverted pendulum. These lengths set different dynamics of the pendulum, ranging in their degree of controllability. We observed two main states during such manipulation, a marginally stable state of the pendulum and a stabilization state in which participants acted to stabilize the system. While during the stabilization state participants increased their applied grip force, for the stable state we observed a mixed behaviour. For small and less controllable pendulums, grip force increased while for larger pendulums, participants could modulate the the grip force according to the anticipated load forces. Based on these results, we suggest that the pendulum dynamics change the control strategy between predictive control and impedance control.

Keywords: Sensory neuroprostheses - Somatosensory, Neural interfaces - Body interfaces, Human performance - Sensory-motor
Abstract: Neurotraumas and neurological diseases often result in compromised proprioceptive feedback, which plays a critical role in motor control by delivering real-time position information. Frequency-modulated electrotactile feedback is a promising solution, as it can deliver proprioceptive information such as a joint angle. Prior works demonstrated that frequency-modulated electrotactile feedback successfully delivered distance information between the end effector and the target object. In this study, we implemented the electronic skin (E-skin) monitoring the elbow joint angle and delivering it to the nervous system via tactile channel. We also demonstrated that frequency-modulated electrotactile feedback improved the elbow joint angle control. The gyroscope measuring the elbow joint angle and electrodes delivering electrotactile feedback were integrated together as a skin using thin silicon coating and polyurethane film. We call this novel E-skin, monitoring and delivering joint angle information, as an electro-prosthetic E-skin. Elbow joint angle matching test with two healthy human subjects showed that the frequency-modulated electrotactile feedback, via electro-prosthetic E-skin, enhanced 101.7% accuracy and 63.8% precision in elbow joint angle control.

Keywords: Neuromuscular systems - Postural and balance
Abstract: Balance refers to the dynamics of body posture to prevent falls. For years, researchers have tried to find out which tasks and measures provide optimal detection of balance disorders, so that they can be quantified. This paper proposes the use of an accelerometer sensor located in the lower back to measure the center of mass accelerations and to characterize the subject’s static balance. For characterizing the static balance objectively, we propose using normality circles, a centroid, and a dispersion circle during the modified Clinical Test of Sensory Interaction in Balance (mCTSIB) test. The proposed methodology was tested using two groups of subjects (10 healthy and 3 unhealthy). Our methodology for the static balance was compared to the Berg Balance Scale score. The results shown that a subject with lower BBS score obtain lower dispersion circle and is outside the normality circle. Also, our method outperforms a new option since it characterizes the static balance in an objective, portable, simple, and low-cost way.

Keywords: Human performance, Human performance - Activities of daily living, Human performance - Ergonomics and human factors
Abstract: Trimanual operation using a robotic supernumerary limb is a new and challenging mechanism for human operators that could enable a single user to perform tasks requiring more than two hands. Foot-controlled interfaces have previously proven able to be intuitively controlled, enabling simple tasks to be performed. However, the effect of going from unimanual to bimanual and then to trimanual tasks on subjects performance and coordination is not well understood. In this paper, unimanual, bimanual and trimanual teleoperation tasks were performed in a virtual reality scene to evaluate the impact of extending to trimanual actions. 15 participants were required to move their limbs together in a coordinated reaching activity. The results show that the addition of another hand resulted in an increase in operating time, where the time increased in going from unimanual to bimanual operation and then increased further when going from bimanual to trimanual. Moreover, the success rate for performing bimanual and trimanual tasks was strongly influenced by the subject's performance in ipsilateral hand-foot activities, where the ipsilateral combination had a lower success rate than contralateral limbs. The addition of a hand did not affect any two-hand coordination rate and even in some cases reduced coordination deviations.

Keywords: Neural interfaces - Implantable systems, Neural interfaces - Biomaterials, Neurorehabilitation
Abstract: The loss of the ability to blink the eyelid is considered the most severe effect of facial nerve paralysis. The delicate homeostasis of the eye is disrupted, and without frequent intervention, the cornea can become damaged, ultimately resulting in blindness. The psychosocial impact is also significant, with individuals withdrawing from society to hide what they perceive to be a disfigurement. Surgical and engineering interventions have been devised to reanimate blink, however, a solution has yet to be designed which addresses both functional and aesthetic concerns. Here we describe an implantable electromagnetic actuator to restore the capacity to blink. Triggered synchronously with the contralateral eye, and externally modifiable to tailor treatment post-operatively to the individual, this implant restores complete blinking and a natural appearance. Cadaver studies (N=12) have been used to validate the device design, including the form factor and force required to elicit a blink, while a passive in vivo study (N=1) has verified the surgical protocol and recovery.

Keywords: Human performance, Human performance - Gait, Human performance - Activities of daily living
Abstract: A process of predicting fall events in patients with Parkinson’s disease (PD) by using a simple motion sensor is described in this paper. Causes of falls in people with PD can be postural instability, freezing of gait, festinating gait, dyskinesias, visuospatial dysfunction, orthostatic hypotension, and posture problems. This study uses only one motion sensor in collecting data. Thus, only fall events caused by festinating gait factors which are moments when the patient suddenly moves faster with smaller steps can be performed and tested. In this preliminary study, fall event scenarios of simulated test cases are performed by five healthy young subjects who are 20 to 28 years old. The acceleration mode in the motion sensor provides information that can detect how fast the subjects move. Data that is collected by the sensor will be analyzed by simple analysis methods and machine learning techniques classification. The proposed study achieved an accuracy of 70.3% for the 10-class model while for binary classification it was 99%.

Keywords: Ultrasound imaging - Breast
Abstract: Accurate breast lesion segmentation in ultrasound images helps radiologists to make exact diagnoses and treatments, which is important to increase the survival rate of breast cancer patients. Recently, deep learning-based methods have demonstrated remarkable results in breast lesion segmentation. However, the blurry breast lesion boundaries and noise artifacts in ultrasound images still limit the performance of the deep learning-based methods. In this paper, we propose a novel segmentation network equipped with a focal self-attention block for improving the performance of breast lesion segmentation. The focal self-attention block can incorporate finegrained local and coarse-grained global information. The finegrained local information is useful to enhance features of breast lesion boundaries, while the coarse-grained global information effectively reduces noise interference. To verify the performance of our network, we implement breast lesion segmentation on our collected dataset of 9836 ultrasound images. The results demonstrate that the focal self-attention block enhances features of breast lesion boundaries and improves the accuracy of breast lesion segmentation.

Keywords: Ultrasound imaging - Elastography, Image reconstruction - Performance evaluation
Abstract: The knowledge of the biomechanical properties of tissues is useful for different applications such as disease diagnosis and treatment monitoring. Reverberant Shear Wave Elastography (RSWE) is an approach that has reduced the restrictions on wave generation to characterize the shear wave velocity over a range of frequencies. This approach is based on the generation of a reverberant field that is generated by the reflections of waves from inhomogeneities and tissue boundaries that exist in the tissue. The Kelvin-Voigt Fractional Derivative model is commonly used to characterize elasticity and viscosity of soft tissue when using shear wave ultrasound elatography. These viscoelastic characteristics can be then validated using mechanical measurements (MM) such as stress relaxation. During RSWE acquisition, the effect of interface pressure, induced by pushing the probe on the skin through the gel pad, on the viscous and elastic characteristics of tissue can be investigated. However, the effect of interface pressure on the validity of the extracted viscous and elastic characteristics was not investigated before. Therefore, the purpose of this study was to compare the estimation of the viscoelastic parameters at different thickness of gel pad against the viscoelastic characteristics obtained from MM. The experiments were conducted in a tissue-mimicking phantom. The results confirm that the relaxed elastic constant can be depreciated. In addition, a higher congruence was found in the viscous parameter estimated at 6 and 7 mm. On the other hand, a difference in the order of fractional derivative was found.

Keywords: Ultrasound imaging - Interventional, Image reconstruction and enhancement - Machine learning / Deep learning approaches, Machine learning / Deep learning approaches
Abstract: Ultrasound exam output of large organs like liver has traditionally been limited to still images or 2D cine loops of key structures, without 3D context. Having 3D context for follow up studies makes ultrasound scanning much easier and for interventional applications such as biopsy. 3D context will reduce wrong sample selection thereby increasing patient comfort. As of today, there is no existing solution which provides 3D anatomical context to users during scanning for large organs like liver. Even for routine measurements like liver volume, patients have to undergo CT or MR scan. In this paper, we propose a novel approach to build-patient specific 3D anatomical surface models from B-mode ultrasound images and tracking information from position sensors. The complexity of the problem stems from the fact that liver boundaries are often not very clear in ultrasound images, in addition to large variability in liver size and shape across patients. Our work uses state-of-the-art deep learning algorithms to detect surface landmarks of liver followed by registering a geometric model to surface point cloud to build patient specific 3D liver model. Further, the developed models will be used to guide users to right lesion locations during the interventional procedure. Our proposed semi-automated workflow ensures the accuracy of the developed models are within acceptable limits for the targeted problem.

Keywords: Ultrasound imaging - Other organs, Image feature extraction
Abstract: The penetrating ability of ultrasound (US) combined with its real-time operation make it the perfect tool for investigating muscle contraction mechanics during complex functional tasks, e.g., locomotion.

Changes in fascicle lengths and pennation angles of muscle fascicles strongly correlate with the capacity of skeletal muscles to produce forces, thereby represent fundamental parameters to be tracked. While the gold standard for extracting these features from US images is still based on manual annotation, the availability of recording devices capable of generating big data of muscle dynamics makes such manual approach unfeasible, setting the need for automated muscle images annotation tools.

Existing approaches, however, are seriously limited, also in view of the continuous developments and technology advancements for ultrafast US and plane-wave imaging. In fact, they rely on conventional (slow) B-mode imaging, make use of point tracking approaches (which often fail due to out-of-plane motion), or can only operate on very high quality images. To overcome all these limitations, we present AEPUS, an automated image labeling tool capable of extracting pennation angles from low quality images using a very small number of plane waves, therefore making it capable of exploiting all the benefits of ultrafast US.

Clinical relevance — Ultrasound is a standard research tool to investigate alterations of spastic muscles in children with Cerebral Palsy. We propose a reliable and time-efficient method to track muscle features in ultrasound images and support clinical biomechanists in their analyses

Keywords: Ultrasound imaging - Other organs, Image analysis and classification - Machine learning / Deep learning approaches
Abstract: The coronavirus disease 2019 (COVID-19) evolved into a global pandemic, responsible for a significant number of infections and deaths. In this scenario, point-of-care ultrasound (POCUS) has emerged as a viable and safe imaging alternative. Computer vision (CV) solutions have been proposed to aid clinicians in POCUS image interpretation, namely detection/segmentation of structures and image/patient classification but relevant challenges still remain. As such, the aim of this study is to develop CV algorithms, using Deep Learning techniques, to create tools that can aid doctors in the diagnosis of viral (COVID-19) and bacterial pneumonia (BP) through POCUS exams. To do so Convolutional Neural Networks (CNNs) were designed to perform in classification tasks. The architectures chosen to build these models were the VGG16, ResNet50, DenseNet169 e MobileNetV2. Patients images were divided in three classes: healthy (HE), BP and viral pneumonia (VP) which falls under COVID-19. Through a comparative study, which was based on several performance metrics, the model based on the DenseNet169 architecture was designated as the best performing model, achieving 78% average accuracy value of the five iterations of 5-Fold Cross-Validation (5FCV). Given that the currently available POCUS datasets for COVID-19 are still limited, the training of the models was negatively affected by such and the models were not tested in an independent dataset. Furthermore, it was also not possible to perform lesion detection tasks. Nonetheless, in order to provide explainability and understanding of the models, Gradient-weighted Class Activation Mapping (GradCAM) were used as a tool to highlight the most relevant classification regions.

Keywords: Ultrasound imaging - Other organs, Fetal and Pediatric Imaging, Image analysis and classification - Machine learning / Deep learning approaches
Abstract: The use of Lung Ultrasound (LUS) as a tool to diagnose and monitor lung diseases in neonates has increased in urban hospitals. Lung ultrasound’s main advantages compared to chest CT or X-rays is that it is less expensive, more accessible, and does not expose the patient to radiation. Performing LUS on neonates and diagnosing the LUS images require highly trained specialist medical professional and clinicians. While availability of such specialists in general is not an issue in urban areas, there is lack of such personnel in rural and remote communities. Hence, an automated computer-aided screening approach as a first level diagnosis assistance in such scenarios might be of significant value. Many of the image morphologies (or patterns) used by clinicians in diagnosing the LUS have strong recurrence characteristics. Building upon this knowledge, in this paper, we propose a feature extraction method designed to quantify such recurrent features for classification of LUS images into 6 common neonatal lung conditions. These conditions were normal lung, chronic lung disease (CLD), transient tachypnea of the newborn (TTN), pneumothorax (PTX), respiratory distress syndrome (RDS), and consolidation (CON) that could be pneumonia or atelectasis. The proposed method extracts virtual scanlines from the LUS images and converts them into signals. Then using recurrence quantification analysis (RQA), features were extracted and fed to pattern classifiers. Using a simple linear classifier the proposed features can achieve a classification accuracy of 69.3% without clinical features and 77.6% with clinical features.

Keywords: CT imaging, Image reconstruction and enhancement - Machine learning / Deep learning approaches, Machine learning / Deep learning approaches
Abstract: Automatic and efficient liver tumor detection in multi-phase CT images is essential in computer-aided diagnosis of liver tumors. Nowadays, deep learning has been widely used in medical applications. Normally, deep learning-based AI systems need a large quantity of training data, but in the medical field, acquiring sufficient training data with high-quality annotations is a significant challenge. To solve the lack of training data issue, domain adaptation-based methods have recently been developed as a technique to bridge the domain gap across datasets with different feature characteristics and data distributions. This paper presents a domain adaptation-based method for detecting liver tumors in multi-phase CT images. We adopt knowledge for model learning from PV phase images to ART and NC phase images. To minimize the domain gap, we employ an adversarial learning scheme with the maximum square loss for mid-level output feature maps using an anchorless detector. Experiments show that our proposed method performs much better for various CT-phase images than normal training.

Keywords: CT imaging, Image segmentation, Machine learning / Deep learning approaches
Abstract: Kidney cancer is one of the common cancers in the world. Automatic segmentation of the kidney and kidney tumor from CT images is of great significance for the therapy treatment of kidney cancer. Due to the diversity of the kidney tumor in terms of location, size, and shape, current methods have limited performance on the tumor segmentation, especially on the boundary. This paper proposes an effective deep neural network with multi-task learning paradigm to improve the boundary segmentation accuracy. The network is an improved U-Net model enhanced by a novel boundary attention mechanism, named boundary attention U-Net (BAU-Net). It consists of a main branch to segment the target regions and an auxiliary branch to generate boundary attention maps to boost the segmentation. Our method is an extension of our original competition method, in which we improve the classical coarse-to-fine framework by using kidney information to segment tumors. Our method is quantitatively evaluated on a public dataset from MICCAI 2021 Kidney and Kidney Tumor Segmentation Challenge (KiTS2021), with mean Dice similarity coefficient (DSC) as 98.04% and 84.09% for the kidney and kidney tumor respectively, outperforming all competitive methods of KiTS2021.

Keywords: CT imaging, Image segmentation, Image feature extraction
Abstract: Automatic localization and segmentation of organs-at-risk (OAR) in CT are essential pre-processing steps in medical image analysis tasks, such as radiation therapy planning. For instance, the segmentation of OAR surrounding tumors enables the maximization of radiation to the tumor area without compromising the healthy tissues. However, the current medical workflow requires manual delineation of OAR, which is prone to errors and is annotator-dependent. In this work, we aim to introduce a unified 3D pipeline for OAR localization-segmentation rather than novel localization or segmentation architectures. To the best of our knowledge, our proposed framework fully enables the exploitation of 3D context information inherent in medical imaging. In the first step, a 3D multi-variate regression network predicts organs’ centroids and bounding boxes. Secondly, 3D organ-specific segmentation networks are leveraged to generate a multi-organ segmentation map. Our method achieved an overall Dice score of 0.9260 ± 0.18% on the VISCERAL dataset containing CT scans with varying fields of view and multiple organs.

Keywords: CT imaging, Image enhancement - Denoising, Image reconstruction and enhancement - Machine learning / Deep learning approaches
Abstract: With the increasing concern regarding the radiation exposure of patients undergoing computed tomography (CT) scans, researchers have been using deep learning techniques to improve the quality of denoised low-dose CT (LDCT) images. In this paper, a cascaded dilated residual network (ResNet) with integrated attention modules, specifically spatial- and channel- attention modules, is proposed. Further, an investigation regarding the effectiveness of per-pixel loss, perceptual loss via VGG16-Net, and SSIM loss functions is covered through an ablation experiment. By knowing how these loss functions affect the output denoised images, a combination of the these loss function is then proposed which aims to prevent edge over-smoothing, enhance textural details and finally, preserve structural details on the denoised images. Finally, bench testing was also done by comparing the visual and quantitative results of the proposed model with the state-of-the-art models such as block matching 3D (BM3D), patch-GAN and dilated convolution with edge detection layer (DRL-E-MP) for accuracy.

Keywords: CT imaging, Image segmentation, Image analysis and classification - Machine learning / Deep learning approaches
Abstract: Multiphase computed tomography (CT) images are widely used for the diagnosis of liver disease. Since each phase has different contrast enhancement (i.e., different domain), the multiphase CT images should be annotated for all phases to perform liver or tumor segmentation, which is a time-consuming and labor-expensive task. In this paper, we propose a dual discriminator-based unsupervised domain adaptation (DD-UDA) for liver segmentation on multiphase CT images without annotations. Our framework consists of three modules: a task-specific generator and two discriminators. We have performed domain adaptation at two levels: one is at the feature level, and the other is at the output level, to improve accuracy by reducing the difference in distributions between the source and target domains. Experimental results using public data (PV phase only) as the source domain and private multiphase CT data as the target domain show the effectiveness of our proposed DD-UDA method.

Keywords: CT imaging, Image analysis and classification - Machine learning / Deep learning approaches
Abstract: Non-Small Cell Lung Cancer (NSCLC) represents up to 85% of all malignant lung nodules. Adenocarcinoma and squamous cell carcinoma account for 90% of all NSCLC histotypes. The standard diagnostic procedure for NSCLC histotype characterization implies cooperation of 3D Computed Tomography (CT), especially in the form of low-dose CT, and lung biopsy. Since lung biopsy is invasive and challenging (especially for deeply-located lung cancers and for those close to blood vessels or airways), there is the necessity to develop non-invasive procedures for NSCLC histology classification. Thus, this study aims to propose Cloud-YLung for NSCLC histology classification directly from 3D CT whole-lung scans. With this aim, data were selected from the openly-accessible NSCLC-Radiomics dataset and a modular pipeline was designed. Automatic feature extraction and classification were accomplished by means of a Convolutional Long Short-Term Memory (ConvLSTM)-based neural network trained from scratch on a scalable GPU cloud service to ensure a machine-independent reproducibility of the entire framework. Results show that Cloud-YLung performs well in discriminating both NSCLC histotypes, achieving a test accuracy of 75% and AUC of 84%. Cloud-YLung is not only lung nodule segmentation free but also the first that makes use of a ConvLSTM-based neural network to automatically extract high-throughput features from 3D CT whole-lung scans and classify them.

Keywords: Scaffolds in tissue engineering - Patterned 3D, Scaffolds in tissue engineering
Abstract: Treatment for critical size defects (CSDs) in bone often use bone grafts to act as a scaffold to help complete healing. Biological scaffolds require bone extraction from the individual or an outside donor while synthetic grafts mostly suffer from poor degradation kinetics and decreased bioactivity. In this study, we investigated a 3D printed scaffold derived from a novel composite bioink composed of alginate and collagen augmented with varying doses from 2 mg/mL to 20 mg/mL of 1% strontium-calcium polyphosphate (SCPP) to control biodegradability and fluid uptake. Scaffolds with increased SCPP concentrations showed higher particle density, lesser swelling ratio and greater biodegradability indicating that these critically important properties for bone healing are fine-tunable and highly dependent on SCPP dosing.

Clinical Relevance— The dosing of 1% SCPP into porous alginate/collagen scaffolds provides adjustable long-term degradation and material properties suitable for potential in vivo CSD applications.

Keywords: Electromagnetic field effects and cell membrane, Electric fields - Tissue regeneration, Non-viral gene delivery
Abstract: In this paper, a broadband microwave device for cell poration is presented, that enables the analysis of the relation between frequency, electrical field strengths and temperature for a successful cell poration. Electromagnetic-thermal coupled simulations in the frequency range from 1 GHz to 10 GHz show that the device reaches electrical field strengths of 100 V/cm and temperatures lower then 40°C. Electroporation experiments with adherent C2C12 mouse myoblast cells show successful uptake of a anti-histone g-H2A.X nanobody at a frequency of 10 GHz.

Keywords: Micro- and nano-sensors, Scaffolds in tissue engineering - Patterned 3D, Biomaterials - Chemical and electrochemical sensors
Abstract: Fabrication of conductive and bioactive microdevices has garnered tremendous attention in the emerging biomedical fields, particularly organic bioelectronics and biosensing. Direct laser 3D printing based on two-photon polymerization (TPP) has shown great promise in construction of well-defined and multi-functional microdevices. Herein, we present a novel photosensitive resin for fabrication of highly conductive and bioactive microstructures via TPP. This resin is based on poly(ethylene glycol) diacrylate that is doped with poly (3,4-ethylenedioxythiophene)-poly(styrenesulfonate) (organic semicoductor), and laminin (extracellular matrix protein) or glucose oxidase (biorecognition enzyme). We demonstrate the fabrication of hybrid microelectrodes, bioactive microstructures for cellular adhesion / spreading, and high-performance glucose biosensors.

Keywords: BioMEMS/NEMS - Tissue engineering and biomaterials, Electric fields - Tissue regeneration, Micro- and nano-sensors
Abstract: Abstract— Functional electrical stimulation (FES) modifies red blood cells (RBC) flux in blood capillaries of muscle. In this work, we aim to investigate changes in the RBC flux in small and large capillaries due to FES using zinc oxide nanowires (ZnO NWs) based electrode at different stimulation parameters. The RBC flux was significantly increased immediately after stimulation, which was evident from decreasing light intensity measured in the region of interest.

Keywords: Micro- and nano-sensors, Micro- and nano-technology
Abstract: The long-term stability of platinum electrodes is a key factor that determines the life-time of biomedical devices, such as implanted neural interfaces like brain stimulation or recording electrodes, cochlear implants, and biosensors. The downsizing of such devices relies on the usage of microfabricated thin-film electrodes. In order to determine and investigate the causal degradation processes for platinum electrodes, it is essential to use potential-controlled experiments, which allow selectable polarization of the electrode without exceeding the water stability window boundaries. Therefore, the surface processes and redox reactions occurring at the electrode are known at all times. In this study, we present the continuous in situ monitoring of platinum-based thin-film electrodes along their complete life cycle in neutral pH with and without the presence of proteins. The usage of chronoamperometry for electrode aging, monitoring of surface processes and the tracking of analyte redox processes, together with cyclic voltammetry to determine the complete amount of surface charge, allows a reliable quantification of fundamental degradation processes. We found that platinum dissolution is primarily driven by the formation and removal of Pt oxide. Despite the significantly lowered charge transfer, the presence of proteins did not prevent material loss or increase electrode lifetime. These results should be considered when interpreting results from current-controlled methods as typically used for neural interfaces.

Keywords: Micro- and nano-technology, Microfluidic applications
Abstract: This paper presents experimental results of effects of a fluoroscopy agent on the radio opacity and steering performance of the steerable micro guidewire. The guidewire is driven by internal pressure, and made of the silicone polymer mixed with the barium sulfate, BaSO4, masterbatch. Steerable distal tips with different BaSO4 densities up to 30 % are fabricated. The radio opacity is measured by comparing CT (computed tomography) numbers of the steerable distal tips. The steering performance is measured by the bending angle at particular internal pressures while being bent up to 180°. Experiment results show that the radio opacity improves while the bending stiffness decreases as the concentration of the barium sulfate increases.

Keywords: Modeling of cell, tissue, and regenerative medicine - 2d and 3d cell modeling, Modeling of cell, tissue, and regenerative medicine - Tissue profiling , Modeling of cell, tissue, and regenerative medicine - Cells
Abstract: The physiological origin of the aperiodic signal present in the electrophysiological recordings, called 1/f neural noise, is unknown; nevertheless, it has been associated with health and disease. The power spectrum slope, - in 1/f, has been postulated to be related to the dynamic balance between excitation (E) and inhibition (I). Our study found that human cerebral organoids grown from induced pluripotent stem cells (iPSCs) from Schizophrenia patients (SCZ) showed structural changes associated with altered elasticity compared to that of the normal cerebral organoids. Furthermore, mitochondrial drugs modulated the elasticity in SCZ which was found related to the changes in the spectral exponent. Therefore, we developed an electro-mechanical model that related the microtubular-actin tensegrity structure to the elasticity and the 1/f noise. Model-based analysis showed that a decrease in the number and length of the constitutive elements in the tensegrity structure decreased its elasticity and made the spectral exponent more negative while thermal white noise will make . Based on the microtubular-actin model and the cross-talk in structural (elasticity) and functional (electrophysiology) response, aberrant mitochondrial dynamics in SCZ are postulated to be related to the deficits in mitochondrial-cytoskeletal interactions for long-range transport of mitochondria to support synaptic activity for E/I balance.

Keywords: Computational modeling - Analysis of high-throughput systems biology data
Abstract: The carotid artery disease is one of the leading causes of mortality worldwide, as it leads to the progressive arterial stenosis that may result to stroke. To address this issue, the scientific community is attempting not only to enrich our knowledge on the underlying atherosclerotic mechanisms, but also to enable the prediction of the atherosclerotic progression. This study investigates the role of T-cells in the atherosclerotic plaque growth process through the implementation of a computational model in realistic geometries of carotid arteries. T-cells mediate in the inflammatory process by secreting interferon-γ that enhances the activation of macrophages. In this analysis, we used 5 realistic human carotid arterial segments as input to the model. In particular, magnetic resonance imaging data, as well as, clinical data were collected from the patients at two time points. Using the baseline data, plaque growth was predicted and correlated to the follow-up arterial geometries. The results exhibited a very good agreement between them, presenting a high coefficient of determination R2=0.64.

Keywords: Organ modeling, Models of organ physiology
Abstract: The gastro-esophageal junction (GEJ) regulates the entry of food into the stomach and prevents reflux of acidic gastric contents into the lower esophagus. This is achieved through multiple mechanisms and the maintenance of a localized high-pressure zone. Diseases of the GEJ typically involve impairments to its muscular functions and often are accompanied by symptoms of reflux, heartburn, and dysphagia. This study aimed to develop a finite element-based model from a unique human cadaver GEJ data reconstructed from an ultramill imaging setup. A pipeline was developed to generate a mesh from an input stack of images. The anatomy of the model was compared to an existing Visible Human finite element GEJ model. Biomechanical simulations were also performed on both models using loading steps of differing levels of calcium to model different levels of contraction. It was found that the ultramill GEJ model is shorter than the Visible Human model (31 vs 48.3 mm), as well as producing lower pressure (1.35 vs 4.36 kPa). The model will be used to investigate detailed pressure development in the GEJ during swallowing under realistic loading conditions.

Keywords: Data-driven modeling
Abstract: Recent reports have highlighted the potential challenges skin pigmentation can have in the accurate estimation of arterial oxygen saturation when using a pulse oximeter. Pulse oximeters work on the principle of photoplethysmography (PPG), an optical technique used for the assessment of volumetric changes in vascular tissue. The primary aim of this research is to investigate the effect of melanin on tissue when utilising the technique of PPG. To address this, a Monte Carlo (MC) light-tissue interaction model is presented to explore the behaviour of melanin in the visible range in the epidermis. A key novelty in this paper is the ability to model the Modified Beer Lambert Law (MBLL) through a fully functional three-dimensional (3D) model in reflective optical geometry. Maximum photon penetration depth was achieved by red light, however limited bio-optical information was retrieved by moderately and darkly pigmented skin at source-detector separations of less than 3 mm. The current MC model can be modified to provide a more realistic representation of absorption and scattering processes in skin.

Keywords: Modeling of cell, tissue, and regenerative medicine - 2d and 3d cell modeling, Modeling of cell, tissue, and regenerative medicine - Cells, Computational modeling - Biological networks
Abstract: Modeling biological neural networks has been a field opening to major advances in our understanding of the mechanisms governing the functioning of the brain in normal and pathological conditions. The emergence of real-time neuromorphic platforms has been leading to a rising significance of bio-hybrid experiments as part of the development of neuromorphic biomedical devices such as neuroprosthesis. To provide a new tool for the neurological disorder characterization, we design real-time single and multicompartmental Hodgkin-Huxley neurons on FPGA. These neurons allow biological neural network emulation featuring improved accuracy through compartment modeling and show integration in bio-hybrid system thanks to its real-time dynamics.

Keywords: Computational modeling - Biological networks, Models of organ physiology, Model building - Sensitivity analysis
Abstract: Spinal cord stimulation (SCS) could be used to restore control of the bladder after spinal cord injury, but substantial development is still required to tailor this technology for bladder function. Computational models could be utilized to accelerate these efforts enabling in-silico optimization of stimulation parameters. However, no model of the spinal pudendo-vesical reflex can simulate the effect of stimulation amplitude on neuron recruitment. This limitation hinders accurate prediction of bladder pressure changes for different stimulation configurations. Here, we implemented an open-source realistic spiking neural network model of the pudendo-vesical reflex enabling exploration of the impact of stimulation amplitude and frequency on bladder pressure changes. We used the o2S2PARC platform to design a parallel implementation of the bladder reflex circuits with NEURON. Our model successfully reproduced and expanded previous studies, producing a decrease in bladder pressure at low stimulation frequency (10 Hz) and excitation at high stimulation frequency (≥33 Hz) in isovolumetric experiments. We then explored the effect of mixed nerve recruitment, simulating a common case of poorly selective spinal cord stimulation. We found that high recruitments of pudendal nerve axons are necessary to maintain this bi-modal behavior, regardless of stimulation specificity. Our framework is fully open-source and can be used to simulate any type of axon stimulations such as SCS and peripheral nerve stimulation.

Keywords: General and theoretical informatics - Machine learning, Sensor Informatics - Wearable systems and sensors, General and theoretical informatics - Supervised learning method
Abstract: Exercise testing has been available for more than a half-century and is a remarkably versatile tool for diagnostic and prognostic information of patients for a range of diseases, especially cardiovascular and pulmonary. With rapid advancements in technology, wearables, and learning algorithm in the last decade, its scope has evolved. Specifically, Cardiopulmonary exercise testing (CPX) is one of the most commonly used laboratory tests for objective evaluation of exercise capacity and performance levels in patients. CPX provides a non-invasive, integrative assessment of the pulmonary, cardiovascular, and skeletal muscle systems involving the measurement of gas exchanges. However, its assessment is challenging, requiring the individual to process multiple time series data points, leading to simplification to peak values and slopes. But this simplification can discard the valuable trend information present in these time series. In this work, we encode the time series as images using the Gramian Angular Field and Markov Transition Field and use it with a convolutional neural network and attention pooling approach for the classification of heart failure and metabolic syndrome patients. Using GradCAMs, we highlight the discriminative features identified by the model.

Keywords: General and theoretical informatics - Natural language processing, General and theoretical informatics - Supervised learning method, General and theoretical informatics - Big data analytics
Abstract: While there has been recent progress in abstractive summarization as applied to different domains including news articles, scientific articles, and blog posts, the application of these techniques to clinical text summarization has been limited.This is primarily due to the lack of large-scale training data and the messy/unstructured nature of clinical notes as opposed to other domains where massive training data come in structured or semi-structured form. Further, one of the least explored and critical components of clinical text summarization is factual accuracy of clinical summaries. This is specifically crucial in the healthcare domain, cardiology in particular, where an accurate summary generation that preserves the facts in the source notes is critical to the well-being of a patient. In this study, we propose a framework for improving the factual accuracy of abstractive summarization of clinical text using knowledge-guided multi-objective optimization. We propose to jointly optimize three cost functions in our proposed architecture during training: generative loss, entity loss and knowledge loss and evaluate the proposed architecture on 1) clinical notes of patients with heart failure (HF), which we collect for this study; and 2) two benchmark datasets, Indiana University Chest X-ray collection (IU X-Ray), and MIMIC-CXR, that are publicly available. We experiment with three transformer encoder-decoder architectures and demonstrate that optimizing different loss functions leads to improved performance in terms of entity-level factual accuracy.

Keywords: General and theoretical informatics - Pattern recognition, General and theoretical informatics - Machine learning, General and theoretical informatics - Deep learning and big data to knowledge
Abstract: Early diagnosis and intervention are clinically considered the paramount part of treating cerebral palsy (CP), so it is essential to design an efficient and interpretable automatic prediction system for CP. We highlight a significant difference between CP infants' frequency of human movement and that of the healthy group, which improves prediction performance. However, the existing deep learning-based methods did not use the frequency information of infants' movement for CP prediction. This paper proposes a frequency attention informed graph convolutional network and validates it on two consumer-grade RGB video datasets, namely MINI-RGBD and RVI-38 datasets. Our proposed frequency attention module aids in improving both classification performance and system interpretability. In addition, we design a frequency-binning method that retains the critical frequency of the human joint position data while filtering the noise. Our prediction performance achieves state-of-the-art research on both datasets. Our work demonstrates the effectiveness of frequency information in supporting the prediction of CP non-intrusively and provides a way for supporting the early diagnosis of CP in the resource-limited regions where the clinical resources are not abundant.

Keywords: General and theoretical informatics - Predictive analytics, General and theoretical informatics - Machine learning
Abstract: Accessing aortic remodeling status through regular follow-ups is essential for acute type A aortic dissection patients undergone surgical treatment. Aortic remodeling status was usually determined using diameter or area measurements of the true and false lumen in specific anatomical slices of medical images. However, these indicators only represent partial information about the aorta and can hardly characterize the overall aorta situation. In this study, we included two types of morphology features collected from computed tomography angiography images to predict the aortic remodeling. One type is the volumetric measurements of the true and false lumen, which provide a better overall description of the aorta, and the other type is the volumetric measurements of the thrombus in false lumen and the patent false lumen, which present more detailed information of the dissection. Through progressively incorporating these measurements into the construction of the remodeling prediction model, we investigated the importance of the features that describe the overall situation and that characterize aortic internal details in remodeling prediction, especially the effect of quantitative thrombosis features. The results showed that with the inclusion of the two types of volume features, the prediction accuracy of the model increased, which proves that volumetric measurements of aortic dissection, especially the volume of thrombus, are of significant value in aortic remodeling prediction, and should be paid more attention on in clinical practice and research areas.

Keywords: General and theoretical informatics - Statistical data analysis, Health Informatics - Disease profiling and personalized treatment, Health Informatics - Informatics for chronic disease management
Abstract: Tinnitus is the perception of sound when no actual external noise is present. Tinnitus is highly prevalent, with more than 1 in 7 adults in the EU having tinnitus, and it causes negative effects on quality of life for many individuals. However, there is currently no cure for tinnitus and its pathophysiology and genesis are unknown. Auditory evoked potentials (AEPs) provide a non-invasive means by which the electrical signals evoked by the brain can be recorded, and constitute a useful indicator for the evaluation of auditory disorders such as tinnitus and hearing loss. The present study analyzed a total of 98 auditory middle evoked potential (AMLR) waveforms, a subtype of AEPs, from 49 participants with subjective tinnitus, attempting to identify differences in AMLR parameters between sufferers with and without tinnitus distress. The waveforms were divided into three categories according to the ear’s hearing level, and comparisons were made between sufferers in the same hearing level category. The results of the analysis indicated some statistically significant differences in AMLR latencies and amplitudes between the compared groups.

Keywords: General and theoretical informatics - Unsupervised learning method, Bioinformatics - Bioinformatics for health monitoring, Health Informatics - Behavioral health informatics
Abstract: Since the mutation in SARS-COV2 poses new challenges in designing vaccines, it is imperative to develop advanced tools for visualizing the genetic information. Specially, it remains challenging to address the patient-to-patient variability and control severe conditions. In this endeavor we analyze the large-scale RNA-sequencing data collected from bronchoalveolar fluid. In this work, we have used PCA and tSNE for the dimension-reduction. The novelty of the current work is to depict a detailed comparison of k-means, HDBSAN and neuro-fuzzy method in visualization of high-dimension data on gene expression. Clinical Relevance— The subpopulation profiling can be used to study the patient-to patient variability when infected by SARS-COV-2 and its variants. The distribution of cell types can be relevant in designing new drugs that are targeted to control the distribution of epithelial cells, T cells and macrophages.

Keywords: Bioinformatics - High throughput –omics (genomics, proteomics, metabolomics, lipidomics, and metagenomics) data analytics for precision health, General and theoretical informatics - Machine learning
Abstract: Multi-omics data integration is key for cancer prediction as it captures different aspects of molecular mechanisms. Nevertheless, the high-dimensionality of multi-omics data with a relatively small number of patients presents a challenge for the cancer prediction tasks. While feature selection techniques have been widely used to tackle the curse of dimensionality of multi-omics data, most existing methods have been applied to each type of omics data separately. In this paper, we propose a multi-agent architecture for feature selection, called MAgentOmics, to consider all omics data together. MAgentOmics extends the ant colony optimization algorithm to multi-omics data, which iteratively builds candidate solutions and evaluates them. Moreover, a new fitness function is introduced to assess the candidate feature subsets without using prediction target such as survival time of patients. Therefore, it can be considered as an unsupervised method. We evaluate the performance of MAgentOmics on the TCGA ovarian cancer multi-omics data from 176 patients using a 5-fold cross-validation. The results demonstrate that the integration power of MAgentOmics is relatively better than the state-of-the-art supervised multi-view method. The code is publicly available at https://github.com/SinaTabakhi/MAgentOmics.

Keywords: General and theoretical informatics - Deep learning and big data to knowledge, General and theoretical informatics - Machine learning, General and theoretical informatics - Artificial Intelligence
Abstract: Recent years have shown a growth in the application of deep learning architectures such as convolutional neural networks (CNNs), to electrophysiology analysis. However, using neural networks with raw time-series data makes explainability a significant challenge. Multiple explainability approaches have been developed for insight into the spectral features learned by CNNs from EEG. However, across electrophysiology modalities, and even within EEG, there are many unique waveforms of clinical relevance. Existing methods that provide insight into waveforms learned by CNNs are of questionable utility. In this study, we present a novel model visualization-based approach that analyzes the filters in the first convolutional layer of the network. To our knowledge, this is the first method focused on extracting explainable information from EEG waveforms learned by CNNs while also providing insight into the learned spectral features. We demonstrate the viability of our approach within the context of automated sleep stage classification, a well-characterized domain that can help validate our approach. We identify 3 subgroups of filters with distinct spectral properties, determine the relative importance of each group of filters, and identify several unique waveforms learned by the classifier that were vital to the classifier performance. Our approach represents a significant step forward in explainability for electrophysiology classifiers, which we also hope will be useful for providing insights in future studies.

Keywords: General and theoretical informatics - Deep learning and big data to knowledge, Bioinformatics - Bioinformatics databases, General and theoretical informatics - Big data analytics
Abstract: Cellular Thermal Shift Assay (CETSA) has been widely used in drug discovery, cancer cell biology, immunology, etc. One of the barriers for CETSA applications is that CETSA experiments have to be conducted on various cell lines, which is extremely time-consuming and costly. In this study, we make an effort to explore the translation of CETSA features cross cell lines, i.e., known CETSA feature of a given protein in one cell line, can we automatically predict the CETSA feature of this protein in another cell line, and vice versa? Inspired by pix2pix and CycleGAN, which perform well on image-toimage translation cross various domains in computer vision, we propose a novel deep neural network model called CycleDNN for CETSA feature translation cross cell lines. Given cell lines A and B, the proposed CycleDNN consists of two auto-encoders, the first one encodes the CETSA feature from cell line A into Z in the latent space Z, then decodes Z into the CETSA feature in cell line B. Similarly, the second one translates the CETSA feature from cell line B to cell line A through the latent space Z ′ . In such a way, the two auto-encoders form a cyclic feature translation between cell lines. The reconstructed loss, cycleconsistency loss, and latent vector regularization loss are used to guide the training of the model. The experimental results on a public CETSA dataset demonstrate the effectiveness of the proposed approach.

Keywords: General and theoretical informatics - Deep learning and big data to knowledge, Sensor Informatics - Multi-sensor data fusion, Health Informatics - Quality of service, trust, security
Abstract: Wearable Photoplethysmography (PPG) has gained prominence as a low cost, unobtrusive and continuous method for physiological monitoring. The quality of the collected PPG signals is affected by several sources of interference, predominantly due to physical motion. Many methods for estimating heart rate (HR) from PPG signals have been proposed with Deep Neural Networks (DNNs) gaining popularity in recent years. However, the “black-box” and complex nature of DNNs has caused a lack of trust in the predicted values. This paper contributes DeepPulse, an uncertainty-aware DNN method for estimating HR from PPG and accelerometer signals, with aims of increasing the reliability, usability and interpretability of the predicted HR values. To the best of the authors’ knowledge no PPG HR estimation method has considered aleatoric and epistemic uncertainty metrics. The results show DeepPulse is the most accurate method for DNNs with smaller network sizes. Finally, recommendations are given to reduce epistemic uncertainty, validate uncertainty estimates, improve the accuracy of DeepPulse as well as reduce the model size for resource-constrained edge devices.

Keywords: General and theoretical informatics - Deep learning and big data to knowledge, Health Informatics - internet of things in healthcare, Sensor Informatics - Wearable systems and sensors
Abstract: Atrial Fibrillation (AF) is a kind of arrhythmia, which is a major morbidity factor, and AF can lead to stroke, heart failure and other cardiovascular complications. Electrocardiogram (ECG) is the basic marker to test the condition of heart and it can effectively detect AF condition. Single lead ECG has the practical advantage for being small form factor and it is easy to deploy. With the sophistication of the current deep learning (DL) models, researchers have been able to construct cardiologist-level models to detect different arrhythmias including AF condition detection from single lead short-time ECG signals. However, such models are computationally expensive and require huge memory size for deployment (more than 100 MB to deploy state-of-the-art 34-layer convolutional neural network-based ECG classification model). Such models need to be significantly trimmed with insignificantly loss of its classification performance for deployment in practical applications like single lead ECG classification in wearable and implantable devices. We have found that classical deep learning model compression techniques like pruning, quantization are not capable of substantial model size reduction without compromising on the model performance. In this paper, we propose LTH-ECG, which is our novel goal-driven winning lottery ticket discovery method, where lottery ticket hypothesis (LTH)-based iterative model pruning is used with the aim of over-pruning avoidance. LTH-ECG reduces the model size by 142x times with insignificant loss of classification performance (less than 1% test F1-score penalty).

Keywords: General and theoretical informatics - Machine learning, General and theoretical informatics - Big data analytics, Bioinformatics - Bioinformatics databases
Abstract: The Cellular Thermal Shift Assay (CETSA) is a biophysical assay based on the principle of ligand-induced thermal stabilization of target proteins. This technology has revolutionized cell-based target engagement studies and has been used as guidance for drug design. Although many applications of CETSA data have been explored, the correlations between CETSA data and protein-protein interactions (PPI) have barely been touched. In this study, we conduct the first exploration study applying CETSA data for PPI prediction. We use a machine learning method, Decision Tree, to predict PPI scores using proteins’ CETSA features. It shows promising results that the predicted PPI scores closely match the groundtruth PPI scores. Furthermore, for a small number of protein pairs, whose PPI score predictions mismatch the ground truth, we use iterative clustering strategy to gradually reduce the number of these pairs. At the end of iterative clustering, the remaining protein pairs may have some unusual properties and are of scientific value for further biological investigation. Our study has demonstrated that PPI is a brand-new application of CETSA data. At the same time, it also manifests that CETSA data can be used as a new data source for PPI exploration study.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Optical imaging and microscopy - Microscopy
Abstract: Automatic surgical phase recognition plays a key role in surgical workflow intelligent analysis and overall optimization in clinical work. In the complicated surgical procedures, similar inter-class appearance and drastic variability in phase duration make this still a challenging task. In this paper, a spatio-temporal transformer is proposed for online surgical phase recognition with different granularity. To extract richer spatial information, a spatial transformer is used to model global spatial dependencies of each time index. To overcome the variability in phase duration, a temporal transformer captures the multi-scale temporal context of different time indexes with a dual pyramid pattern. Our method is thoroughly validated on the public Cholec80 dataset with 7 coarse-grained phases and CATARACTS2020 dataset with 19 fine-grained phases, outperforming state-of-the-art approaches with 91.4% and 84.2% accuracy, taking only 24.5M parameters.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, X-ray imaging applications, Image enhancement
Abstract: Breast cancer remains the leading cause of cancer deaths and the second highest cause of death, in general, among women worldwide. Fortunately, over the last few decades, with the introduction of mammography, the mortality rate of breast cancer has significantly decreased. However, accurate classification of breast masses in mammograms is especially challenging. Various Computer-Aided Diagnosis (CAD) systems are being developed to assist radiologists with the accurate classification of breast abnormalities. In this study, classification of benign and malignant masses, based on the subtraction of temporally sequential digital mammograms and machine learning, is proposed. The performance of the algorithm was evaluated on a dataset created for the purposes of this study. In total, 196 images from 49 patients, with precisely annotated mass locations and biopsy confirmed malignant cases, were included. Ninety-six features were extracted and five feature selection algorithms were employed to identify the most important features. Ten classifiers were tested using leave-one-patient-out and 7-fold cross-validation. Neural Networks, achieved the highest classification performance with 90.85% accuracy and 0.91 AUC, an improvement compared to the state-of-the-art. These results demonstrate the effectiveness of the subtraction of temporally consecutive mammograms for the classification of breast masses as benign or malignant.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Image feature extraction, Image classification
Abstract: Colorectal cancer (CRC) was responsible during 2020 for about one million deaths worldwide. Polyps are protuberance masses, observed in routine colonoscopies, that constitute the main CRC biomarker. Nonetheless, one of the best alternatives to the polyp malignancy classification is the vascular pattern analysis, typically observed from specialized narrow-band images (NBI). Even worst, these patterns are only characterized from gastroenterologist observations, introducing subjectivity and being prone to diagnostic errors, with misclassifications ranging from 59.5% to 84.2%. This work introduces a non-aligned and bi-directional deep projection between optical colonoscopy (OC) and NBI sequences, to recover enhanced OC sequences, integrating vascular patterns, that allow better discrimination among adenomas, hyperplastic and serrated polyps. This self-supervised representation help with misclassification in standard OC observations. The validation was performed on a total of 76 OC and 76 NBI sequences, achieving a gain of 22.34% w.r.t descriptors computed from raw OC.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Image classification, Ultrasound imaging - Other organs
Abstract: Lung ultrasound (LUS) as a diagnostic tool is gaining support for its role in the diagnosis and management of COVID-19 and a number of other lung pathologies. B-lines are a predominant feature in COVID-19, however LUS requires a skilled clinician to interpret findings. To facilitate the interpretation, our main objective was to develop automated methods to classify B-lines as pathologic vs. normal. We developed transfer learning models based on ResNet networks to classify B-lines as pathologic (at least 3 B-lines per lung field) vs. normal using COVID-19 LUS data. Assessment of B-line severity on a 0-4 multi-class scale was also explored. For binary B-line classification, at the frame-level, all ResNet models pretrained with ImageNet yielded higher performance than the baseline nonpretrained ResNet-18. Pretrained ResNet-18 has the best Equal Error Rate (EER) of 9.1% vs the baseline of 11.9%. At the clip-level, all pretrained network models resulted in better Cohen’s kappa agreement (linear-weighted) and clip score accuracy, with the pretrained ResNet-18 having the best Cohen’s kappa of 0.815 [95% CI: 0.804-0.826], and ResNet-101 the best clip scoring accuracy of 93.6%. Similar results were shown for multi-class scoring, where pretrained network models outperformed the baseline model. A class activation map is also presented to guide clinicians in interpreting LUS findings. Future work aims to further improve the multi-class assessment for severity of B-lines with a more diverse LUS dataset.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Multimodal image fusion, Magnetic resonance imaging - Other organs
Abstract: Clinically significant regions (CSR), captured over multi-parametric MRI (mp-MRI) images, have emerged as a potential screening test for early prostate cancer detection and characterization. These sequences are able to quantify morphology, micro-circulation, and cellular density patterns that might be related to cancer disease. Nonetheless, this evaluation is mainly carried out by expert radiologists, introducing inter-reader variability in the diagnosis. Therefore, different deep learning models were proposed to support the diagnosis, but a proper representation of prostate lesions remains limited due to the non-alignment among sequences and the dependency of considerable amounts of labeled data for learning. The main limitation of such representation lies in the cross-entropy minimization that only exploits inter-class variation, being insufficient data augmentation and transfer learning strategies. This work introduces a Supervised Contrastive Learning (SCL) strategy that fully exploits the inter and intra-class variability of prostate lesions to robustly represent MRI regions. This strategy extracts lesion sample tuples, with positive and negative labels, regarding a query lesion. Such tuples are involved into an easy-positive, and semi-hard negative mining to project samples that better update the deep representation. The proposed learning strategy achieved an average ROC-AUC of 0.82, to characterize prostate cancer in MRI, using only the 60% of the available annotated data. newline

indent textit{Clinical relevance}— A robust learning scheme that properly finds representations in limited data scenarios to classify clinically significant MRI regions on prostate cancer.

Keywords: Magnetic resonance imaging - Cardiac imaging, Image feature extraction, Machine learning / Deep learning approaches
Abstract: The presence of abnormalities when the left ventricle is deformed is related to the patients’ prognosis after a first myocardial infarction. These deformations can be detected by performing a cardiac magnetic resonance (CMR) study. Currently, late gadolinium enhancement (LGE) is considered to be the gold standard when performing CMR imaging. However, CMR with LGE overestimates infarct size and underestimates recovery of dysfunctional segments after myocardial infarction. Based on this statement, the objective is to detect, characterize, and quantify the extent of myocardial infarction in patients with cardiac pathologies, using parameters derived from CMR, in order to obtain greater precision in patients’ recovery predictions than when only studying LGE images. For this purpose, we studied the infarct presence and extension from a total of 105 images from 35 patients, and calculated myocardium strain and torsion to characterize and quantify the affected tissue. A total of twenty-one parameters were selected to create predictive models. Moreover, we compared two feature extraction methods, and the performance of five machine learning algorithms. Results show that both temporal and strain parameters are the most relevant to detect and characterize the extent of myocardial infarction. The use of imaging techniques and machine learning algorithms have great potential and show promising results when it comes to detecting the presence and extent of myocardial infarction.

Keywords: Magnetic resonance imaging - Cardiac imaging, Magnetic resonance imaging - Pulse sequence, Cardiac imaging and image analysis
Abstract: Magnetic Resonance Imaging (MRI) is the clinical gold standard for assessment of myocardial viability, but requires injection of exogenous Gadolinium-based contrast agents. Recently, T1ρ-mapping has been proposed as a fully non-invasive alternative for imaging myocardial fibrosis without the need for contrast agent injection. However its applicability at high fields is hindered by susceptibility to MRI system imperfections, such as inhomogeneities in the B0and B+1 fields. In this work we propose a single breath-hold ECG-triggered single-shot bSSFP sequence to enable T1ρ-mapping in-vivo at 3T. Adiabatic T1ρ preparations are evaluated to reduce B0 and B+1 sensitivity in comparison with conventional spin-lock modules. Numerical Bloch simulations were performed to identify optimal parameters for the adiabatic pulses. Experiments yield T1ρ values in the myocardium equal to 148.13 ± 54.08 ms for the best adiabatic preparation and 16.01 ± 20.75 for the reference non-adiabatic spin-lock, with 26.91% against 89.74% relative difference in T1ρ values across two shimming conditions. Both phantom and in vivo measurements show increased myocardium/blood contrast and improved resilience against system imperfections compared with non-adiabatic T1ρ preparations, enabling the use at 3T. Adiabatically-preparedT1ρ-mapping sequences form a promising candidate for non-contrast evaluation of ischemic and non-ischemic cardiomyopathies at 3T.

Keywords: Magnetic resonance imaging - Cardiac imaging, Magnetic resonance imaging - Pulse sequence, Image reconstruction and enhancement - Parametric image reconstruction
Abstract: Ischemic heart disease (IHD) is one of the leading causes of death worldwide. Myocardial infarction (MI) makes up a third of all IHD cases, and cardiac magnetic resonance imaging (MRI) is often used to assess its damage to myocardial viability. Late gadolinium enhancement (LGE) is the current gold standard, but the use of gadolinium-based agents limits the clinical applicability in some patients. Spin-lock (SL) dispersion has recently been proposed as a promising non-contrast biomarker for the assessment of MI. However, at 3T, the required range of SL preparations acquired at different amplitudes suffers from specific absorption rate (SAR) limitations and off-resonance artifacts. Relaxation Along a Fictitious Field (RAFF) is an alternative to SL preparations with lower SAR requirements, while still sampling relaxation in the rotating frame. In this study, a single breath-hold simultaneous TRAFF2 and T2 mapping sequence is proposed for SL dispersion mapping at 3T. Excellent reproducibility (coefficient of variations lower than 10%) was achieved in phantom experiments, indicating good intrascan repeatability. The average myocardial TRAFF2, T2, and SL dispersion obtained with the proposed sequence (68.0 ± 10.7 ms, 44.0 ± 4.0 ms, and 0.4 ± 0.2 x 10-4 s2, respectively) were comparable to the reference methods (62.7 ± 11.7 ms, 41.2 ± 2.4 ms, and 0.3 ± 0.2 x 10-4 s2, respectively). High visual map quality, free of B0 and B1 related artifacts, for T2, TRAFF2, and SL dispersion maps was obtained in phantoms and in vivo, suggesting promise in clinical use at 3T.

Keywords: Magnetic resonance imaging - Cardiac imaging, Image segmentation, Cardiac imaging and image analysis
Abstract: Cardiac magnetic resonance imaging (CMRI) improves the diagnosis of cardiovascular diseases by providing images at high spatio-temporal resolution helping physicians in providing correct treatment plans. Segmentation and identification of various substructures of the heart at different cardiac phases of end-systole and end-diastole helps in the extraction of ventricular function information such as stroke volume, ejection fraction, myocardium thickness, etc. Manual delineation of the substructures is tedious, time-consuming, and error-prone. We have implemented a 3D GAN that includes 3D contextual information capable of segmenting and identifying the substructures at different cardiac phases with improved accuracy. Our method is evaluated on the ACDC dataset (4 pathologies, 1 healthy group) to show that the proposed method outperforms other methods in literature with less amount of data. Also, the proposed provided a better Dice score in segmentation surpassing other methods on a blind-tested M&Ms dataset.

Keywords: Magnetic resonance imaging - Cardiac imaging, Iterative image reconstruction, Regularized image Reconstruction
Abstract: Cardiac magnetic resonance (CMR) imaging is the one of the gold standard imaging modalities for the diagnosis and characterization of cardiovascular diseases. The clinical cine protocol of the CMR typically generates high-resolution 2D images of heart tissues in a finite number of separated and independent 2D planes, which are appropriate for the 3D reconstruction of biventricular heart surfaces. However, they are usually inadequate for the whole-heart reconstruction, specifically for both atria. In this regard, the paper presents a novel approach for automated patient-specific 3D whole-heart mesh reconstruction from limited number of 2D cine CMR slices with the help of a statistical shape model (SSM). After extracting the heart contours from 2D cine slices, the SSM is first optimally fitted over the sparse heart contours in 3D space to provide the initial representation of the 3D whole-heart mesh, which is further deformed to minimize the distance from the heart contours for generating the final reconstructed mesh. The reconstruction performance of the proposed approach is evaluated on a cohort of 30 subjects randomly selected from the UK Biobank study, demonstrating the generation of high-quality 3D whole-heart meshes with average contours to surface distance less than the underlying image resolution and the clinical metrics within acceptable ranges reported in previous literature.

Keywords: Magnetic resonance imaging - Cardiac imaging, Image reconstruction and enhancement - Machine learning / Deep learning approaches
Abstract: In this paper, we describe a 3D convolutional neural network (CNN) framework to compute and generate super-resolution late gadolinium enhanced (LGE) cardiac magnetic resonance imaging (MRI) images. The proposed CNN framework consists of two branches: a super-resolution branch with a 3D dense deep back-projection network (DBPN) as the backbone to learn the mapping of low-resolution LGE cardiac volumes to high-resolution LGE cardiac volumes, and a gradient branch that learns the mapping of the gradient map of low resolution LGE cardiac volumes to the gradient map of their high-resolution counterparts. The gradient branch of the CNN provides additional cardiac structure information to the super-resolution branch to generate structurally more accurate super-resolution LGE MRI images. We conducted our experiments on the 2018 atrial segmentation challenge dataset. The proposed CNN framework achieved a mean peak signal-to-noise ratio (PSNR) of 30.91 and 25.66 and a mean structural similarity index measure (SSIM) of 0.91 and 0.75 on training the model on low-resolution images downsampled by a scale factor of 2 and 4, respectively.

Keywords: Sensory neuroprostheses, Sensory neuroprostheses - Visual, Sensory neuroprostheses - Auditory
Abstract: An important brain re-wiring, the so-called cross-modal plasticity, occurs during progression of retinal degenerative diseases to compensate for lack of visual input. The visual cortex does not go ‘unused’, instead it is devoted to processing other sensory modalities. In this study we recorded, in the visual cortex, visual- and auditory-evoked potentials in an anesthetized murine model of retinal degeneration. The latency to the first peak of the recorded local field potentials was used to assess the speed of the response. Visual responses occurred significantly faster in the control group. Conversely, auditory responses appeared significantly faster in animals with retinal degeneration. This suggests the compensatory neural rewiring is optimizing the performance of other sensory modalities, hearing in this case. This phenomenon may play an important role in visual neuro-rehabilitation. Whether or not it can promote or deter the interpretation of artificially encoded neural signals from a visual prosthesis remains to be studied.

Keywords: Neural stimulation - Deep brain, Brain physiology and modeling, Neurological disorders - Psychiatric disorders
Abstract: This study models and investigates whether temporally interfering electric fields (TI EFs) could function as an effective non-invasive brain stimulation (NIBS) method for deep brain structure targeting in humans, relevant for psychiatric applications. Here, electric fields off- and on-target are modelled and compared with other common NIBS modalities (tACS, TMS). Additionally, local effects of the field strength are modelled on single-compartment neuronal models. While TI EFs are able to effectively reach deep brain targets, the ratio of off- to on-target stimulation remains high and comparable to other NIBS and may result in off-target neural blocks.

Clinical Relevance— This study builds on earlier work and demonstrates some of the challenges –such as off-target conduction blocks– of applying TI EFs for targeting deep brain structures important in understanding the potential of treating neuropsychiatric conditions in the future.

Keywords: Neural stimulation - Deep brain, Brain physiology and modeling, Neurological disorders
Abstract: Abstract— Finite Element Method (FEM) simulations of the electric field is a useful tool to estimate the activated tissue around Deep Brain Stimulation (DBS) electrodes. Based on our previous research, a two-part software package named DBSim and ELMA is presented. ELMA is used to classify brain tissue into grey matter, white matter, blood, and cerebrospinal fluid and assign electric conductivities accordingly. This data is then used in DBSim to generate patient-specific simulations of the electric field around currently implemented leads Medtronic 3387 and 3389, and Abbott 6180 and 6181. The software is available for free download at https://liu.se/en/article/ne-downloads

Clinical Relevance— This is a tool meant for research and educational purposes for e.g., studies on optimal target areas for DBS.

Keywords: Neural stimulation - Deep brain, Neural stimulation
Abstract: Deep brain stimulation (DBS) offers therapeutic benefits to patients suffering from treatment-resistant movement and neurological disorders. The newest generation of DBS devices utilizes segmented electrodes to direct current asymmetrically to the surrounding neuronal tissue. Since segmented electrodes offer a larger number of contacts for individualized patient-specific programming, it is becoming more critical to ensure that the surgically intended and actual orientation of the lead match to avoid adverse side effects caused by stimulation of specific nuclei. Postoperative image analysis algorithms, such as DiODe in Lead-DBS, are commonly used to determine DBS leads' actual orientation. However, there has yet to be a comparison of the deviation between intended and actual orientations across the most commonly implanted directional DBS systems: Boston Scientific Cartesia™ and St. Jude Medical Infinity. This study is the first to statistically analyze the rotation of 126 leads from both DBS systems.

Keywords: Neural stimulation, Neural stimulation - Deep brain, Smart neural implants - Neurostimulation
Abstract: Deep brain stimulation (DBS) is becoming a fundamental tool for the treatment and study of neurological and psychiatric diseases and disorders. Recently developed DBS devices and electrodes have allowed for more flexible and precise stimulation. Densely packed stimulation contacts can be independently stimulated to shape the electric field, targeting pathways of interest, and avoiding those that may cause side-effects. However, this flexibility comes at a cost. Each additional stimulation setting causes an exponential increase in the number of potential stimulation settings. Recent works have addressed this problem using Bayesian optimization. However, this approach has a limited ability to learn from multiple subjects to improve performance. In this study we extend a recently developed meta-Bayesian optimization algorithm to the DBS domain. We evaluated this approach compared to classical Bayesian optimization and a random search using data collected from a nonhuman primate during stimulation of the subthalamic nucleus while recording evoked potentials in the motor cortex and locally within the subthalamic nucleus. On the task of finding the stimulation setting that maximized the evoked potential across a distribution of generated objective functions, meta-Bayesian optimization significantly outperformed the other approaches with a cumulative reward of 8.93±0.70, compared to 7.17±1.64 for Bayesian optimization (p < 10-9) and 6.89±1.56 for the random search (p < 10-9). Moreover, the algorithm outperformed Bayesian optimization when tested on an objective function not used during training. These results demonstrate that meta-Bayesian optimization can take advantage of the structure underlying a distribution of objective function and learn an optimal search strategy that can generalize beyond the objective functions that were not part of the training data.

Keywords: Neural stimulation - Deep brain, Neurorehabilitation, Neural signals - Machine learning & Classification
Abstract: Tuning the parameters of controllers to attain the best performance is a challenging task in designing effective closed-loop neuromodulation systems. In this paper, we present a distributed architecture for automated tuning and adaptation of closed-loop neuromodulation control systems. We use this approach for the automated parameter tuning of a Proportional-Integral (PI) neuromodulation controller using Bayesian optimization. We use a biophysically-grounded mean-field model of neural populations under electrical stimulation as a simulation environment for testing and prototyping the proposed framework and characterizing its performance. Our results demonstrate the feasibility of using Bayesian optimization for performance-based automated tuning of a PI controller in closed-loop set-point neuromodulation control tasks.

Keywords: Time-frequency and time-scale analysis - Wavelets, Signal pattern classification, Neural networks and support vector machines in biosignal processing and classification
Abstract: Hypertension is a health issue whose late diagnosis could lead to renal, cerebral, and cardiac events. In this work, it is proposed to use the wavelet scattering transform (WST) as a feature extraction technique applying classical machine learning techniques using photoplethysmography (PPG) signals as input to detect elevated blood pressure and compare its performance with transfer learning applied through fine-tuned convolutional neural networks. The results show that the features obtained by applying the WST and training a logistic regression and support vector machine produced similar results in terms of accuracy compared to fine-tuned convolutional neural networks, with the advantage that the WST could be used to generate a white-box model, which is better suited for a potential medical diagnosis application.

Keywords: Data mining and big data methods - Patient outcome and risk analysis, Physiological systems modeling - Signal processing in physiological systems, Neural networks and support vector machines in biosignal processing and classification
Abstract: Intracranial hypertension (IH) is associated with poor outcome in traumatic brain injury (TBI) patients and must be avoided to prevent secondary brain injury. In clinical practice the most common method of IH detection is the calculation of the mean value of intracranial pressure (ICP) and the therapeutic intervention is usually introduced when the mean exceeds a certain threshold. This threshold, however, is rather individual for each patient than universal for all. It is well known that impaired cerebrovascular reactivity and reduced intracranial compliance are associated with raised ICP. This work explores a new definition of life-threatening hypertension (LTH) which accounts for the state of cerebral compliance. In the proposed method, changes in compliance are analysed through identification of likely pathological and/or pathological shapes of ICP pulse waveforms using a neural network. In terms of predictive power for mortality in TBI, detection of both shape clasess of ICP pulse waveforms during raised ICP offers similar results to previously proposed LTH definition accounting for the state of cerebrovascular reactivity (77.8% vs 76.9% accuracy, respectively). On the other hand, the fully pathological shapes of ICP pulses are present during ICP rises almost only in recordings of patients who died: out of 216 analysed patients only 6% of surviving and as many as 42% of deceased patients developed this type of LTH event. The stricter definition of LTH events including only pathological shape of ICP pulses presents the highest accuracy among the analysed approaches for mortality prediction (87.9%).

Keywords: Signal pattern classification, Physiological systems modeling - Signal processing in physiological systems
Abstract: Hemorrhage is the leading cause of preventable death from trauma. Traditionally, vital signs have been used to detect blood loss and possible hemorrhagic shock. However, vital signs are not sensitive for early detection because of physiological mechanisms that compensate for blood loss. As an alternative, machine learning algorithms that operate on a noninvasive arterial blood pressure (ABP) waveform acquired via photoplethysmography have been shown to provide an effective early indicator. However, these machine learning approaches lack physiological interpretability. In this paper, we evaluate the importance of nine ABP-derived features that provide physiological insight, using a database of 40 human subjects from a lower-body negative pressure model of progressive central hypovolemia. One feature was found to be considerably more important than any other. That feature, the half-rise to dicrotic notch (HRDN), measures an approximate time delay between the ABP ejected and reflected wave components. This delay is an indication of compensatory mechanisms such as reduced arterial compliance and vasoconstriction. For a scale of 0% to 100%, with 100% representing normovolemia and 0% representing decompensation, linear regression of the HRDN feature results in root-mean-squared error of 16.9%, R2 of 0.72, and an area under the receiver operating curve for detecting decompensation of 0.88. These results are comparable to previously reported results from the more complex black box machine learning models.

Keywords: Physiological systems modeling - Signal processing in physiological systems, Directionality, Multivariate methods
Abstract: Traumatic Brain Injury (TBI) patients present high levels of physical stress, which in some situations can manifest as Plateau Wave (PW) episodes. This intense stress phenomenon can be evidenced by Heart Rate Variability (HRV). Thus, the multivariate and simultaneous analysis of cardiocerebrovascular oscillations, involving the RR intervals, mean arterial pressure (MAP) and the amplitude of intracranial pressure (AMP), will be useful to understand the interconnections between body signals, allowing the interpretation of the combined activity of pathophysiological mechanisms. In this work, the multiscale representation of the Transfer Entropy (TE) and of its decomposition in the network of these three interacting processes is obtained, based on a Vector AutoRegressive Fractionally Integrated (VARFI) framework for Gaussian processes. This method allows to assess directed interactions and to quantify the information flow accounting for the simultaneous presence of short-term dynamics and long range correlations. The results show that the baseline RR, but not MAP can provide information about the possibility of a PW arising. During PW, the long-term correlations highlight synergistic interactions between MAP and AMP processes on RR. The multiscale decomposition of the information along with the incorporation of the long term correlations allowed a better description of HRV during PW, highlighting the fact that the HRV mirrors this cerebrovascular phenomena.

Keywords: Signal pattern classification, Neural networks and support vector machines in biosignal processing and classification, Data mining and big data methods - Machine learning and deep learning methods
Abstract: Bioelectronic medicine is a new approach for developing closed-loop neuromodulation protocols on the peripheral nervous system (PNS) to treat a wide range of disorders currently treated with pharmacological approaches. Algorithms need to have low computational cost in order to acquire, process and model data for the modulation of the PNS in real time. Here, we present a fast learning-based decoding algorithm for the classification of cardiovascular and respiratory functional alterations (i.e., challenges) by using neural signals recorded from intraneural electrodes implanted in the vagus nerve of 5 pigs. Our algorithm relies on 9 handcrafted features, extracted following signal temporal windowing, and a multi-layer perceptron (MLP) for feature classification. We achieved fast and accurate classification of the challenges, with a computational time for feature extraction and prediction lower than 1.5 ms. The MLP achieved a balanced accuracy higher than 80% for all recordings. Our algorithm could represent a step towards the development of a closed-loop system based on a single intraneural interface with both the potential of real time classification and selective modulation of the PNS.

Keywords: Signal pattern classification, Physiological systems modeling - Signal processing in physiological systems, Adaptive filtering
Abstract: We propose an accurate chronic stress estimation system that utilizes personalized models based on correlation maximization between physiological features and ground truth, which helps determine physiological features effective for the estimation. The personalized models are trained using features respectively found for each individual classes among which the relationships between features and ground truth differ. Which class a new user belongs to can be estimated from the results of a personality questionnaire, as well as by means of conventional methods. W.r.t. evaluation data, with the cooperation of 168 subjects, 599 sets of 1-month wearable-sensor data and ground-truth Perceived Stress Scale (PSS) data were collected, along with the Big Five Personality Traits for each subject. In chronic stress estimation evaluations using this above data, we have confirmed that the proposed classification system achieved 69.1% estimation accuracy in terms of increase/decrease in PSS, as compared to 59.3% and 56.8% achieved, respectively, with two conventional methods, one employing no classification and the other employing k-means clustering.

Keywords: Wearable body sensor networks and telemetric systems, Wearable sensor systems - User centered design and applications, Health monitoring applications
Abstract: Wearables are objective tools for human activity recognition (HAR). Advances in wearables enable synchronized multi-sensing within a single device. This has resulted in studies investigating the use of single or multiple wearable sensor modalities for HAR. Some studies use inertial data, others use surface electromyography (sEMG) from multiple muscles and different post-processing approaches. Yet, questions remain about accuracies relating to e.g., multi-modal approaches, and sEMG post-processing. Here, we explored how inertial and sEMG could be efficiently combined with machine learning and used with post-processing methods for better HAR. This study aims recognition of four basic daily life activities; walking, standing, stair ascent and descent. Firstly, we created a new feature vector based on the domain knowledge gained from previous mobility studies. Then, a feature level data fusion approach was used to combine inertial and sEMG data. Finally, two supervised learning classifiers (Support Vector Machine, SVM, and the k-Nearest Neighbors, kNN) were tested with 5-fold cross-validation. Results show the use of inertial data with sEMG increased overall accuracy by 3.5% (SVM) and 6.3% (kNN). Extracting features from linear envelopes instead of bandpass filtered sEMG improves overall HAR accuracy in both classifiers.

Keywords: Wearable sensor systems - User centered design and applications, Modeling and analysis
Abstract: The objective of this work focuses on multiple independent user profiles that capture behavioral, emotional, medical, and physical patterns in the working and living environment resulting in one general user profile. Depending on the user’s current activity (e.g. walking, eating, etc.), medical history, and other influential factors, the developed framework acts as a supplemental assistant to the user by providing not only the ability to enable supportive functionalities (e.g. image filtering, magnification, etc.) but also informative recommendations (e.g. diet, alcohol, etc.). The personalization of such a profile lies within the user’s past preferences using human activity recognition as a base, and it is achieved through a statistical model, the Bayesian belief network. Training and real-time methodological pipelines are introduced and validated. The employed deep learning techniques for identifying human activities are presented and validated in publicly available and in-house datasets. The overall accuracy of human activity recognition reaches up to 86.96%.

Keywords: IoT sensors for health monitoring, Health monitoring applications, Physiological monitoring - Novel methods
Abstract: Parkinson's disease (PD) affects 1% of the population over the age of 60, and its prevalence increases with age. The disease progresses over time, and the condition can vary significantly in a day, which makes it difficult for precise diagnosis and medication based on short clinical sessions. Therefore, home health monitoring can play an important role in improving the healthcare of the PD patients. In this study, we proposed a method to detect, classify, and quantify daily movements and motor symptoms of PD by using the wireless sensing technology. With the presence of human movements in a space with the Wi-Fi coverage, the channel state information (CSI) of the wireless signal was transformed into images. The images were used to train a deep learning model to distinguish between different daily movements and simulated tremor. The results showed that our method obtained 99.59% and 100% accuracy of recognizing the tremor with modified VGG19 and modified Resnet152, respectively. In addition, the tremor movement was then successfully segmented out and quantified for the frequency and duration.

Keywords: Modeling and analysis, Novel methods
Abstract: Maintaining adequate hydration is important for health. Inadequate liquid intake can cause dehydration problems. Despite the increasing development of liquid intake monitoring, there are still open challenges in drinking detection under free-living conditions. This paper proposes an automatic liquid intake monitoring system comprised of wrist-worn Inertial Measurement Units (IMUs) to recognize drinking gesture in free-living environments. We build an end-to-end approach for drinking gesture detection by employing a novel multi-stage temporal convolutional network (MS-TCN). Two datasets are collected in this research, one contains 8.9 hours data from 13 participants in semi-controlled environments, the other one contains 45.2 hours data from 7 participants in free-living environments. The Leave-One-Subject-Out (LOSO) evaluation shows that this method achieves a segmental F1-score of 0.943 and 0.900 in the semi-controlled and free-living datasets, respectively. The results also indicate that our approach outperforms the convolutional neural network and long-short-term-memory network combined model (CNN-LSTM) on our datasets. The dataset used in this paper is available at https://github.com/Pituohai/drinking-gesture-dataset/.

Keywords: IoT sensors for health monitoring, Sensor systems and Instrumentation, Modeling and analysis
Abstract: In this paper, an ensemble gentle boost decision tree classification algorithm is trained to classify handwashing from similar activities such as applying lotion to hands. Data is collected using a 3-axis accelerometer and gyroscope worn on the wrist. First, the data collection procedure is described. Then, feature identification is discussed. Once the feature matrix was created, the MATLAB classification learner app was used to classify the data based on the identified features. The overall classification rate achieved was 91.6% using an optimized boosted ensemble classifier.

Keywords: New sensing techniques, Health monitoring applications, Physiological monitoring - Modeling and analysis
Abstract: Detailed assessment of smoking topography (puffing and post-puffing metrics) can lead to a better understanding of factors that influence tobacco use. Research suggests that portable mouthpiece-based devices used for puff topography measurement may alter natural smoking behavior. This paper evaluated the impact of a portable puff topography device (CReSS Pocket) on puffing & post-puffing topography using a wearable system, the Personal Automatic Cigarette Tracker v2 (PACT 2.0) as a reference measurement. Data from 45 smokers who smoked one cigarette in the lab and an unrestricted number of cigarettes under free-living conditions over 4 consecutive days were used for analysis. PACT 2.0 was worn on all four days. A puff topography instrument (CReSS pocket) was used for cigarette smoking on two random days during the four days of study in the laboratory and free-living conditions. Smoke inhalations were automatically detected using PACT2.0 signals. Respiratory smoke exposure metrics (i.e., puff count, duration of cigarette, puff duration, inhale-exhale duration, inhale-exhale volume, volume over time, smoke hold duration, inter-puff interval) were computed for each puff/smoke inhalation. Analysis comparing respiratory smoke exposure metrics during CReSS days and days without CReSS revealed a significant difference in puff duration, inhale-exhale duration and volume, smoke hold duration, inter-puff interval, and volume over time. However, the number of cigarettes per day and number of puffs per cigarette were statistically the same irrespective of the use of the CReSS device. The results suggested that the use of mouthpiece-based puff topography devices may influence measures of smoking topography with corresponding changes in smoking behavior and smoke exposure.

Keywords: Biomechanics and robotics - Clinical evaluation in rehabilitation and orthopedics, Mechanics of locomotion and balance, New technologies and methodologies in human movement analysis
Abstract: A key problem in human balance recovery lies in understanding the mechanism of balance behavior with redundant bio-mechanical motors. Motor synergy has been known as an efficient tool to analyze characteristics of motion behavior and reconstruct control command. In this paper, motor synergy analysis for different control strategies is proposed to analyze different balance motion coordination for various levels of pushing force, and understand the coordination of human multiple joints regarding balance recovery. The spatial synergy of specific joint angles for different pushing levels is computed with the principal component analysis (PCA) to evaluate the adaptive balance motion response patterns and illustrate the improvement of balance robustness through the transition of joint coordination. Therefore, coordination transitions over multiple joints in balance recovery movements were analyzed to better understand the mechanism of balance strategy generation in this study.

Keywords: Mechanics of locomotion and balance, New technologies and methodologies in human movement analysis, Applied tissue and organ models and motion analysis
Abstract: Forefoot pain, hallux valgus, shoe sore, flat foot, and calluses are among the common foot problems encountered by high heel wearers. This study aimed to investigate the external forces associated with shoe sore and callus while wearing formal heel shoes. The external force on the 1st, 2nd, and 5th metatarsal heads and heel center was measured using the ShokacChip. Women were asked to wear pumps with four heel heights (10, 30, 55, and 80 mm) and walk 15 m twice. Thirty-five women were included. The data of two participants were excluded due to sensor fault. With higher heels, normal stress (pressure) was significantly stronger on the inside of the forefoot and significantly weaker on the outside. Shear stress did not always increase or decrease proportionally with respect to heel height. SPR-i of the forefoot associated with callus formation was minimal in the 30-mm heel.

Keywords: New technologies and methodologies in human movement analysis, Rehabilitation robotics and biomechanics - Integrated diagnostic and therapeutic systems
Abstract: In recent years, markerless motion capture using a depth camera or RGB camera without any restriction on the subject has been attracting attention. Especially, depth cameras such as Kinect and RealSense allow instantaneous motion capture even at home outside lab environment, which is attractive for rehabilitation usage. However, single depth camera can capture steadily skeleton only when the subject stands facing to camera for the limited range, thus it is hard to apply to track skeletons while walking. Multiple depth cameras setting may allow to expand the range, but it can involve non-practical calibration process and can affect instantaneous capture advantage of depth camera. In this study, we propose a systematic method to integrate the motion information of skeletal models obtained from multiple depth cameras. The proposed method can perform a quick calibration using skeletal models instead of external reference objects, and estimate the spatial relationship of the sensors that allows the depth camera to move. The result demonstrates stable skeleton tracking free from occlusion problem keeping instantaneous capture capability of depth cameras.

Keywords: Optimization in musculoskeletal biomechanics, Modeling and simulation in musculoskeletal biomechanics, Joint biomechanics
Abstract: Several biomedical contexts such as diagnosis, rehabilitation, and ergonomics require an accurate estimate of human upper limbs kinematics. Wearable inertial measurement units (IMUs) represent a suitable solution because of their unobtrusiveness, portability, and low-cost. However, the time-integration of the gyroscope angular velocity leads to an unbounded orientation drift affecting both angular and linear displacements over long observation interval. In this work, a Denavit-Hartenberg model of the upper limb was defined in accordance with the guidelines of the International Society of Biomechanics and exploited to design an optimization kinematics process. This procedure estimated the joint angles by minimizing the difference between the modelled and IMU-driven orientation of upper arm and forearm. In addition, reasonable constraints were added to limit the drift influence on the final joint kinematics accuracy. The validity of the procedure was tested on synthetic and experimental data acquired with a robotic arm over 20 minutes. Average rms errors amounted to 2.8 deg and 1.1 for synthetic and robot data, respectively.

Keywords: Robotics - Orthotics and Exoskeletons, Wearable robotic systems - Orthotics and Exoskeletons, Exoskeleton applications
Abstract: In operational settings, lower-limb active exoskeletons may experience errors, where an actuation that should be present is missed. These missed actuations may impact users' trust in the system and the adapted human-exoskeleton coordination strategies. In this study, we introduced pseudo-random catch trials, in which an assistive exoskeleton torque was not applied, to understand the immediate responses to missed actuations and how users' internal models to an exoskeleton adapt upon repeated exposure to missed actuations. Participants (N = 15) were instructed to complete a stepping task while wearing a bilateral powered ankle exoskeleton. Human-exoskeleton coordination and trust were inferred from task performance (step accuracy), step characteristics (step length and width), and joint kinematics at selected peak locations of the lower limb. Step characteristics and task accuracy were not impacted by the loss of exoskeleton torque as hip flexion was modulated to support completing the stepping task during catch trials, which supports an impacted human-exoskeleton coordination. Reductions in ankle plantarflexion during catch trials suggest user adaptation to the exoskeleton. Trust was not impacted by catch trials, as there were no significant differences in task performance or gait characteristics between earlier and later strides. Understanding the interactions between human-exoskeleton coordination, task accuracy, and step characteristics will support development of exoskeleton controllers for non-ideal operational settings.

Keywords: Joint biomechanics, New technologies and methodologies in human movement analysis, Biomechanics and robotics - Clinical evaluation in rehabilitation and orthopedics
Abstract: Open-access databases can facilitate data sharing among researchers and provide normative data for objective clinical assessment development, robotic design, and biomechanical modeling. However, most existing databases focus on gait, balance, and hand gestures without providing elbow and shoulder kinematics that are required in activities of daily living. Furthermore, the few existing upper limb datasets include small sample sizes without consistent data collection protocols, which hinder robotic engineers’ ability to design robotic devices that accommodate the general population. To address the literature gap, an open-access upper limb kinematic database was proposed. Due to the impact of COVID-19 on human research, only data from 16 participants were collected.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches
Abstract: Automatic diagnosis of eye diseases from retinal fundus images is quite challenging. Common public datasets include images of subjects with multiple diseases with uneven distribution of labels. Rare diseases are especially challenging due to their under-representation in such datasets. In this paper, we propose a training pipeline for the multi-labeled classification with uneven distribution of the sample size and sample difficulty. First, we guide the training of the initial model by weighing the training loss using an inverse-frequency for each class. This will balance the training on over-represented and under-represented samples. We then adjust the class weights using the aggregated loss for each class, and train for more iterations. In this way, the model at each iteration will focus more on difficult samples and cover the shortcomings of the previous model. Finally, we ensemble together all the models and apply post-processing for improving multi-label predictions. Experiments on the Retinal Fundus Multi-Disease Image Dataset (RFMiD) prove the effectiveness of our pipeline. Furthermore, our method provides new ideas for the training of deep learning for the multi-label classification task and imbalanced datasets.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Multivariate image analysis, MEG imaging
Abstract: Major depressive disorder (MDD) and bipolar disorder (BD) are two major mood disorders with partly overlapped symptoms but different treatments. However, their misdiagnosis and mistreatment are common based on the DSM-V criteria, lacking objective and quantitative indicators. This study aimed to develop a novel approach that accurately classifies MDD and BD based on their resting-state magnetoencephalography (MEG) signals during euthymic phases. A revisited 3D CNN model, Semi-CNN, that could automatically detect brainwave patterns in spatial, temporal, and frequency domains was implemented to classify wavelet-transformed MEG signals of normal controls and MDD and BD patients. The model achieved 96.05% testing accuracy and an average of 95.71% accuracy for 5-fold cross-validation. Furthermore, saliency maps of the model were estimated using Grad-CAM++ for visual explanation of the proposed classification model and highlighting the disease-specific brain regions and frequencies.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Image classification, Image segmentation
Abstract: Extravasation occurs secondary to the leakage of medication from blood vessels into the surrounding tissue during intravenous administration resulting in significant soft tissue injury and necrosis. If treatment is delayed, invasive management such as surgical debridement, skin grafting and even amputation may be required. Thus, it is imperative to develop a develop a smartphone application for predicting level of extravasation severity from skin image. Two Deep Neural Network (DNN) architectures, U-Net and DenseNet-121, were used to segment skin and lesion, and to classify extravasation severity. Sensitivity and specificity for predicting binary classification between asymptomatic and abnormal cases were 77.78 and 90.24%. For severe extravasation attained the highest accuracy of 90.83%, followed by mild extravasation of 85.32%, and moderate extravasation of 77.98%. The accuracy of moderate-to-severe extravasation classification can be improved by using an ensemble model for multi-class classification. These findings proposed a novel and feasible DNN approach for screening extravasation from skin images. The implementation of DNN-based applications on mobile devices has a strong potential for clinical application in low-resource countries.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, X-ray radiography
Abstract: For physicians to make rapid clinical decisions for patients with congestive heart failure, the assessment of pulmonary edema severity in chest radiographs is vital. Although deep learning has shown promise in detecting the presence or absence or discrete grades of severity, of such edema, prediction of continuous-valued severity yet remains a challenge. Here, we propose PENet: Siamese convolutional neural networks to assess the continuous spectrum of severity of lung edema from chest radiographs. We present different modes of implementing this network and demonstrate that our best model outperforms that of earlier work (mean AUC of 0.91 over 0.87), while using only 1/16-th the dimension of input images and 1/69-th the size of training data, thus also saving expensive computation.

Keywords: Image analysis and classification - Machine learning / Deep learning approaches, Image segmentation, Optical imaging - Coherence tomography
Abstract: The retinal vascular system adapts and reacts rapidly to ocular diseases, such as glaucoma, diabetic retinopa- thy and age-related macular degeneration. Here we present a combination of methods to further extract vascular information from 12x12mm wide-field optical coherence tomography angiog- raphy (OCTA). An integrated U-Net for the segmentation and classification of arteries and veins reached a segmentation IoU of 0.7095 0.0224, and classification IoU of 0.8793 0.1049 and 0.8928 0.0929 respectively. A correcting algorithm which uses topological information was created to correct the misclassification and connectivity of the vessels, which showed an average increase of 8.29% in IoU. Finally, the vessel morphometry of branch orders was extracted, where this allows the direct comparison of artery/vein, arterioles/venules and capillaries.

Keywords: Image classification, Image analysis and classification - Machine learning / Deep learning approaches, Image feature extraction
Abstract: Computer-aided skin lesion classification using dermoscopy is essential for early detection of melanoma, which is the most effective means to reduce the mortality rate. Although many deep learning models have been designed for this task, skin lesion classification remains challenging due to the small sample size, inter-class similarity, intra-class inconsistency, and class imbalance. In this paper, we propose a hybrid deep residual network and Fisher vector (ResNet-FV) algorithm for skin lesion classification, aiming to boost the performances of ResNet using the Fisher vector encoding scheme. The proposed algorithm has been evaluated on the 2018 Skin Lesion Analysis Towards Melanoma Detection Challenge (ISIC-skin 2018) dataset and achieved a balanced multi-class accuracy of 0.798, outperforming several existing solutions.

Keywords: Magnetic resonance imaging - MR neuroimaging, Image enhancement - Denoising, Brain imaging and image analysis
Abstract: NOise Reduction with DIstribution Corrected (NORDIC) principal component analysis (PCA) has been shown to selectively suppress thermal noise and improve temporal signal-to-noise ratio (tSNR) in human functional magnetic resonance imaging (fMRI). However, the feasibility to improve rodent fMRI using NORDIC PCA has not been explored. In this study, we developed a rodent fMRI preprocessing pipeline by incorporating NORDIC and evaluated its performance in a range of rodent fMRI applications from resting-state fMRI to task-evoked fMRI using optogenetics. In resting-state fMRI, we demonstrated a significant increase in tSNR by more than 3 times after NORDIC correction with reduced variance and improved task-free relative cerebrovascular reactivity across cortical depth. In optogenetic fMRI, apart from tSNR increase, more activated voxels and a significant decrease in the variance of activated brain signals were observed after NORDIC correction without apparent change in brain morphology. Taken together, our results signified the values of NORDIC correction for better detection of brain activities in rodent fMRI.

Keywords: Brain imaging and image analysis, Magnetic resonance imaging - MR neuroimaging
Abstract: Autism spectrum disorder (ASD) is a lifelong neurodevelopmental condition characterized by social communication, language and behavior impairments. Leveraging deep learning to automatically predict ASD has attracted more and more attention in the medical and machine learning communities. However, how to select effective measure signals for deep learning prediction is still a challenging problem. In this paper, we studied two kinds of measure signals, i.e., regional homogeneity (ReHo) and Craddock 200 (CC200), which both represents homogeneous functional activity, in the framework of deep learning, and designed a new mechanism to effectively joint them for deep learning based ASD prediction. Extensive experiments on the ABIDE dataset provide empirical evidence in support of effectiveness of our method. In particular, we obtained 79% in terms of accuracy by effectively fusing these two kinds of signals, much better than any single-measure model (ReHo SM-model: ~69% and CC200 SM-model: ~70%). These results suggest that leveraging multi-measure signals together are effective for ASD prediction.

Keywords: Magnetic resonance imaging - MR neuroimaging, Machine learning / Deep learning approaches, Brain imaging and image analysis
Abstract: There remains an open question about whether and in what context brain function varies in adolescence and adulthood. In this work, we systematically study the functional brain networks of adolescents and adults, outlining the significant differences in the developing brain detected via time-resolved functional network connectivity (trFNC) derived from a fully automated independent component analysis pipeline applied to resting-state fMRI data in over 50K individuals. We then statistically analyze the transient, recurrent, and robust brain state profiles in both groups. We confirmed the results in independent replication datasets for both groups. Our findings indicate a strengthening of a state reflecting functional coupling within the visual, motor, and auditory domains and anticorrelation with all other domains in a unique adult state profile, a pattern consistently less modular in adolescents. This new insight into possible integration, strengthening, and modularization of resting-state brain connections beyond childhood convergently indicates that the highlighted temporal dynamics likely reflect robust differences in brain function in adolescents versus adults.

Keywords: Magnetic resonance imaging - MR neuroimaging, Machine learning / Deep learning approaches, Image analysis and classification - Machine learning / Deep learning approaches
Abstract: Multi-site collaboration, which gathers together samples from multiple sites, is a powerful way to overcome the small-sample problem in the neuroimaging field and has the potential to discover more robust and reproducible biomarkers. However, confounds among the datasets caused by various site-specific factors may dramatically reduce the cross-site reproducibility performance. To properly remove confounds while improving cross-site task performances, we propose a maximum classifier discrepancy generative adversarial network (MCD-GAN) that combines the advantages of generative models and maximum discrepancy theory. The mechanisms of MCD-GAN and how it harmonizes the dataset are visualized using simulated data. The performance of MCD-GAN was also compared with state-of-the-art methods (e.g., ComBat, cycle-GAN) within Adolescent Brain Cognitive Development (ABCD) dataset. Result demonstrates that the proposed MCD-GAN can effectively improve the cross-site gender classification performance by harmonizing site effects. Our proposed framework is also suitable for various classification/prediction tasks and is promising to facilitate the cross-site reproducibility of neuroimaging studies.

Keywords: Magnetic resonance imaging - MR neuroimaging, Image segmentation, Brain imaging and image analysis
Abstract: Deep brain stimulation (DBS) is an established yet growing treatment for a range of neurological and psychiatric disorders. Over the last decade, numerous studies have underscored the effect of electrode placement on the clinical outcome of DBS. As a result, imaging is now extensively used for DBS electrode localization, even though the accuracy of different modalities in determining the true coordinates of DBS electrodes is less explored. Postoperative magnetic resonance imaging (MRI) is the gold standard method for DBS electrode localization, however, the geometrical distortion induced by the lead's artifact could limit the accuracy. In this work, we investigated to what degree the difference between the true location of the lead's tip and the location of the tip estimated from the MRI artifact varies depending on the MRI sequence parameters, acquisition plane, phase encoding direction, and the implant's extracranial trajectory.

Keywords: Magnetic resonance imaging - MR neuroimaging, Brain imaging and image analysis, Functional image analysis
Abstract: Resting-state functional network connectivity (rsFNC) has shown utility for identifying characteristic functional brain patterns in individuals with psychiatric and mood disorders, providing a promising avenue for biomarker development. However, several factors have precluded widespread clinical adoption of rsFNC diagnostics, namely the lack of standardized approaches for capturing comparable and reproducible imaging markers across individuals, as well as the disagreement on the amount of data required to robustly detect intrinsic connectivity networks (ICNs) and diagnostically relevant patterns of rsFNC. Here, we investigate the robustness of (1) subject-specific ICNs standardized to an a priori network template via spatially constrained ICA (scICA), and (2) rsFNC differences between schizophrenia and control groups with respect to the length of the fMRI. Our results suggest clinical rsFMRI scans, when decomposed with scICA, could potentially be shortened to just 2-4 minutes without significant loss of individual rsFNC information or classification performance of longer scan lengths.

Keywords: Optical imaging, Multimodal image fusion, Image registration, segmentation, compression and visualization - Volume rendering
Abstract: In laparoscopic surgery image-guided navigation systems could support the surgeon by providing subsurface information such as the positions of tumors and vessels. For this purpose, one option is to perform a reliable registration of preoperative 3D data and a surface patch from laparoscopic video data. A robust and automatic 3D-3D registration pipeline for the application during laparoscopic surgery has not yet been found due to application-specific challenges. To gain a better insight, we propose a framework enabling a qualitative and quantitative comparison of different registration approaches. The introduced framework is able to evaluate 3D feature descriptors and registration algorithms by generating and modifying synthetic data from clinical examples. Different confounding factors are considered and thus the reality can be reflected in any simplified or more complex way. Two exemplary experiments with a liver model, using the RANSAC algorithm, showed an increasing registration error for a decreasing size of the surface patch size and after introducing modifications. Moreover, the registration accuracy was dependent on the position and structure of the surface patch. The framework helps to quantitatively assess and optimize the registration pipeline, and hereby suggests future software improvements even with only few clinical examples.

Keywords: Optical imaging, Machine learning / Deep learning approaches, Image classification
Abstract: Keratoconus is a severe eye disease that leads to deformation of the cornea. It impacts people aged 10-25 years and is the leading cause of blindness in that demography. Corneal topography is the gold standard for keratoconus diagnosis. It is a non-invasive process performed using expensive and bulky medical devices called corneal topographers. This makes it inaccessible to large populations, especially in the Global South. Low-cost smartphone-based corneal topographers, such as SmartKC, have been proposed to make keratoconus diagnosis accessible. Similar to medical-grade topographers, SmartKC outputs curvature heatmaps and quantitative metrics that need to be evaluated by doctors for keratoconus diagnosis. An automatic scheme for evaluation of these heatmaps and quantitative values can play a crucial role in screening keratoconus in areas where doctors are not available. In this work, we propose a dual-head convolutional neural network (CNN) for classifying keratoconus on the heatmaps generated by SmartKC. Since SmartKC is a new device and only had a small dataset (114 samples), we developed a 2-stage transfer learning strategy—using historical data collected from a medical-grade topographer and a subset of SmartKC data—to satisfactorily train our network. This, combined with our domain-specific data augmentations, achieved a sensitivity of 91.3% and a specificity of 94.2%.

Keywords: Optical imaging
Abstract: Abstract: Corneal visualization Scheimpflug technology (Corvis ST) has the potential to indirectly provide information on corneal structure via the analysis of statistical properties of the backscatter. The aim of this work was to ascertain whether there are age-related changes in the dynamics of corneal backscatter during an air-puffed induced corneal deformation. Retrospective data from Corvis ST measurements of 151 young subjects (19–30 years) and 82 older subjects (50–87 years) were considered. Each measurement consisted of 140 frames (sampling frequency: 4330 fps). For every frame the cornea was first segmented, then regions of interest, encompassing temporal, central and nasal parts of cornea were selected, to which the parameters of Weibull distribution (scale and shape) were fitted, leading to time series of the estimated parameters. Apparent differences were found between the parameters of Weibull distribution between the two considered groups that manifest themselves mostly in the nasal region of the cornea. However, those differences cannot be attributed to the age alone. For this, a normalization method is proposed that leads to a much better separation between the groups in all considered regions. Clinical relevance: The parameters of the corneal backscatter are widely used to asses corneal clarity (so-called corneal densitometry). Recently, the parameters of Weibull distribution, fitted to the corneal backscatter data, have been used to support diagnosis of keratoconus. This work contributes to the assessment of corneal clarity by identifying the apparent age-related differences in those parameters when dynamic raw data is considered, highlighting the need for such parameters to be appropriately normalized. Further, it shown that the shape parameter of Weibull distribution, unlike the scale parameter, carries that information already for the raw, non-normalized data.

Keywords: Optical imaging and microscopy - Fluorescence microscopy, Image reconstruction and enhancement - Machine learning / Deep learning approaches
Abstract: Convolutional neural networks (CNN) have revealed exceptional performance for fluorescence lifetime imaging (FLIM). However, redundant parameters and complicated topologies make it challenging to implement such networks on embedded hardware to achieve real-time processing. We report a lightweight, quantized neural architecture that can offer fast FLIM imaging. The forward-propagation is significantly simplified by replacing matrix multiplications in each convolution layer with additions and data quantization using a low bit-width. We first used synthetic 3-D lifetime data with given lifetime ranges and photon counts to assure correct average lifetimes can be obtained. Afterwards, human prostatic cancer cells incubated with gold nanoprobes were utilized to validate the feasibility of the network for real-world data. The quantized network yielded a 37.8% compression ratio without performance degradation.

Keywords: Optical imaging and microscopy - Fluorescence microscopy, Image analysis and classification - Machine learning / Deep learning approaches
Abstract: Wide-field fluorescence lifetime imaging (FLIM) is a promising technique for biomedical and clinic applications. Integrating with CMOS single-photon avalanche diode (SPAD) sensor arrays can lead to cheaper and portable real-time FLIM systems. However, the FLIM data obtained by such sensor systems often have sophisticated noise features. There is still a lack of fast tools to efficiently recover lifetime parameters from highly noise-corrupted fluorescence signals. This paper proposes a smart wide-field FLIM system containing a 192 × 128 COMS SPAD sensor, and a field-programmable gate array (FPGA) embedded deep learning (DL) processor. The processor adopts a hardware-friendly and light-weighted neural network for fluorescence lifetime analysis, showing the advantages of high accuracy against noise, fast speed, and low power consumption. Experimental results demonstrate the proposed system's superior and robust performances, promising for many FLIM applications such as FLIM-guided clinical surgeries, cancer diagnosis, and biomedical imaging.

Keywords: Optical imaging and microscopy - Fluorescence microscopy, Image analysis and classification - Digital Pathology, Image analysis and classification - Machine learning / Deep learning approaches
Abstract: Breast conserving surgery aims at the complete removal of malignant lesions while minimizing healthy tissue loss. To ensure the balance between complete resection of the cancer and conservation of healthy tissue, intra-operative margin assessment is necessary. Deep ultra violet (DUV) fluorescence scanning microscope provides fast whole-surface-imaging (WSI) of excised tissue with contrast between malignant and normal tissues. Then, an automated breast cancer classification method on DUV images is required for intra-operative margin assessment. Deep learning shows the promising results in breast cancer classification, but limited DUV image dataset poses overfitting challenge to train the robust network. To tackle this challenge, we partition the DUV WSI image into small patches and extract pathological features for each patch from a pre-trained network using a transfer learning approach. We feed pathological features into a decision-tree-based classifier and fuse patch-level classification results based on regional importance to determine malignant or benign WSI. Experimental results on 60 DUV images show that our proposed method outperforms the standard deep learning classification in terms of improving the classification performance and identifying cancerous regions.

Keywords: Data mining and big data methods - Biosignal classification, Time-frequency and time-scale analysis - Time-frequency analysis, Data mining and big data methods - Machine learning and deep learning methods
Abstract: The interest in development of methods and tools for recognizing human emotions has increased continuously. Using physiological information, especially the peripheral physiological signals, to identify emotions is an important direction for this area. This paper proposes an approach for emotion recognition based on energy-related features extracted from peripheral physiological signals. Three emotions： calm, happiness and fear, were elicited in 54 volunteers using video clips while three peripheral physiological signals were recorded: Electrocardiography (ECG), Photoplethysmography (PPG) and Respiration. Given that energy-related features of physiological signals are closely related to autonomic nervous systems activities, nine energy-related features were extracted from the recorded physiological signals. To find the optimal feature subset to represent the target emotions, the correlation between features and emotion state, as well as the discrimination ability of feature for emotion recognition were both analyzed. Four optimal features were then selected for further classification. Moreover, models based on Decision Tree (DT) were built to evaluate the performance of these features for purpose of recognition of emotion states of calm, happiness, and fear. The results show that the DT models based on these four optimal features could distinguish fear from calm (AUC=0.879, Accuracy=87.8%), happiness from calm (AUC=0.915, Accuracy=91.8%), and fear from happiness (AUC=0.822, Accuracy=81.8%), with a global recognition accuracy of 70.8%. These results indicate that energy-related features of peripheral physiological signals can reliably identify emotions, especially intense emotions.

Keywords: Neural networks and support vector machines in biosignal processing and classification
Abstract: The electrocardiogram (ECG) is a vital diagnostic tool used in many health applications. In practice, interference by muscle artifacts is very common and may significantly complicate interpretation of the ECG waveform. In this work we investigate the removal of muscle artifacts in single-channel ECG signals using neural network models. To this end, we compare two neural network architectures which were previously used for ECG denoising and propose a novel third method based on the ConvTasNet. The neural networks are trained on simulated data using artificial mixtures of single-channel ECG (lead II) and surface EMG signals taken from publicly available datasets. ECG data were sampled from the MIT-BIH Arrhythmia and the PTB-XL database. The former provides recordings of ambulatory ECGs while the latter contains a large variety of cardiac pathologies recorded in a clinical setting. The muscle artifacts were sampled from the MIT-BIH Noise Stress Test database and the TaiChi database. In the past, most denoising methods were only tested on the two smaller MIT-BIH datasets. In this work, we report performances on larger datasets and thus provide stronger evidence for a clinical use-case. We also report out-of distribution performance of the three methods by switching the ECG dataset between training and test. The herein investigated variant of the ConvTasNet substantially reduces interference by muscle artifacts, outperforms state-of-the-art methods and thus, may support clinical decision making.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Physiological systems modeling - Signal processing in physiological systems, Data mining and big data methods - Machine learning and deep learning methods
Abstract: Continuous monitoring of blood pressure (BP) can help individuals manage their chronic diseases such as hypertension, requiring non-invasive measurement methods in free-living conditions. Recent approaches fuse Photoplethysmograph (PPG) and electrocardiographic (ECG) signals using different machine and deep learning approaches to non-invasively estimate BP; however, they fail to reconstruct the complete signal, leading to less accurate models. In this paper, we propose a cycle generative adversarial network (CycleGAN) based approach to extract a BP signal known as ambulatory blood pressure (ABP) from a clean PPG signal. Our approach uses a cycle generative adversarial network that extends the GAN architecture for domain translation, and outperforms state-of-the-art approaches by up to 2X in BP estimation.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Data mining and big data methods - Machine learning and deep learning methods, Data mining and big data methods - Biosignal classification
Abstract: For semantic segmentation, U-Net provides an end-to-end trainable framework to detect multiple class objects from background. Due to its great achievements in computer vision tasks, U-Net has broadened its application to biomedical signal processing, especially, segmentation of waveforms in ECG signal. Despite its superior performance for QRS complex detection to other traditional signal processing methods, direct application of the U-Net to R peak detection has limitation since the U-Net structures tend to predict high probability around true peak. Such multiple detection results require additional process to determine a unique peak location in each QRS complex. In this study, we use a regression process to detect R peak instead of pixel-wise classification. Such regression process guarantees a unique peak location prediction. We collect data from resting ECG systems and wearable ECG devices as well as public ECG databases and the proposed model is trained on various combinations of the data sources. Especially, we investigate the robustness of the model for input data from the wearable devices when the model is trained by data from heterogeneous devices.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Data mining and big data methods - Machine learning and deep learning methods, Data mining and big data methods - Biosignal classification
Abstract: In this study, a lightweight CNN-based Electrocardiogram (ECG) classification model is implemented to operate it on a wearable device for real-time arrhythmia detection by efficiently reducing the number of parameters of the model. Ten second-windowed ECGs from three different public ECG databases were used to learn and classify them into four classes: normal sinus rhythm, atrial fibrillation, atrial premature contraction, and ventricular premature contraction. The model implemented in the workstation environment was converted using the TensorFlow Lite framework and then imported into an ARM Cortex-M4 architecture-based nRF52840 microprocessor. The proposed model shows high performance (97.7% accuracy and 97.4% F1 score) with reasonable execution time: 298ms and current consumption: 3.55mA at optimized for speed and execution time: 480ms and current consumption: 3.82mA at optimized for size, respectively.

Keywords: Neural networks and support vector machines in biosignal processing and classification, Signal pattern classification, Data mining and big data methods - Machine learning and deep learning methods
Abstract: Ballistography(BSG) is a non-intrusive and low-cost alternative to electrocardiography (ECG) for heart rate (HR) monitoring in infant. Due to the inter-patient variance and susceptibility to noise, heartbeat detection in the BSG waveform remains a challenge. The aim of this study was to estimate HR from a bed-based pressure mat BSG signal using a deep learning approach. We trained a U-Net-based deep neural network through supervised learning by deriving ground truth heartbeat time window targets from simultaneously recorded ECG signals after peak matching. For improved generalization, we modified an existing U-Net to include an IC-layer. A predictive performance of 80 % was achieved using the U-Net without the IC-layer. The inclusion of the IC-layer, while improving the generalization ability of the model to detect heartbeats, did not improve the HR estimation performance.

Keywords: Nonlinear dynamic analysis - Biomedical signals, Coupling and synchronization - Coherence in biomedical signal processing
Abstract: Prolonged use of mechanical ventilation (MV) can lead to greater complications for a patient. In clinical practice, it is important to identify patients who could fail in the extubation process. However, accurately predicting the outcome of this process remains a challenge. The diaphragm muscle is one of the most active elements in the breathing process. On the other hand, there are several techniques to derive respiratory information from the ECG signal. Signals derived from diaphragmatic activity and from the ECG, such as the envelope of the surface diaphragm electromyographic signal (sEMGi) and the respiratory signal derived from the electrocardiogram (ECG) could contribute to analyze the respiratory response in patients assisted by MV. This work proposes the analysis of the coherence between sEMGi and EDR signals to determine possible differences in the respiratory pattern between successful and failed patients undergoing weaning. 40 patients with MV, candidates for weaning trial process and underwent a spontaneous breathing test were analyzed, classified into: a successful group (SG: 19 patients) that maintained spontaneous breathing after the test, and a failed group (FG: 21 patients) that required reconnection to the MV. The cross correlation, power spectral density and magnitude squared coherence (MSC) of the sEMGi and the EDR signals were estimated. According to the results, the MSC parameters such as area under the curve and mean coherence value presented statistically significance differences between the two groups of patients (p = 0.024). Our results suggest that both sEMGi and EDR signals could provide information about the behavior of the respiratory system in these patients.

Clinical Relevance— This study analyzes the correlation and the coherence between the envelope of the surface electromyographic signal, and the respiratory signal derived from the ECG to characterize the respiratory pattern of successful and failed patients on weaning process.

Keywords: Nonlinear dynamic analysis - Biomedical signals, Signal pattern classification, Data mining and big data methods - Biosignal classification
Abstract: Ventricular arrhythmias are the primary arrhythmias that cause sudden cardiac death. In current clinical and preclinical research, the discovery of new therapies and their translation is hampered by the lack of consistency in diagnostic criteria for distinguishing between ventricular tachycardia (VT) and ventricular fibrillation (VF). This study develops a new set of features, similarity maps, for discrimination between VT and VF using deep neural network architectures. The similarity maps are designed to capture the similarity and the regularity within an ECG trace. Our experiments show that the similarity maps lead to a substantial improvement in distinguishing VT and VF.

Keywords: Physiological systems modeling - Multivariate signal processing, Signal pattern classification, Neural networks and support vector machines in biosignal processing and classification
Abstract: Pain assessment represents the first fundamental stage for proper pain management, but currently, methods applied in clinical practice often lack in providing a satisfying characterization of the pain experience. Automatic methods based on the analysis of physiological signals (e.g., photoplethysmography, electrodermal activity) promise to overcome these limitations, also providing the possibility to record these signals through wearable devices, thus capturing the physiological response in everyday life. After applying pre-processing, feature extraction and feature selection methods, we tested several machine learning algorithms to develop an automatic classifier fed with physiological signals recorded in real-world contexts and pain ratings from 21 cancer patients. The best algorithm achieved up to 72% accuracy. Although performance can be improved by enlarging the dataset, preliminary results proved the feasibility of assessing pain by using physiological signals recorded in real-world contexts.

Keywords: Physiological systems modeling - Signal processing in physiological systems, Nonlinear dynamic analysis - Biomedical signals, Data mining and big data methods - Machine learning and deep learning methods
Abstract: This work proposes a novel beat scoring system for quantifying the effects of exhalation and inhalation on the seismocardiogram (SCG) signals in rest and physiologically modulated conditions. Data from 19 subjects during rest, listening to classical music and recovery states were used. First, the SCG and electrocardiogram (ECG) signals were segmented into exhalation and inhalation phases using the respiration signal; and a representative SCG beat for each exhale and inhale phase was constructed using the ECG R-peak locations. Second, the significant differences across the exhalation- and inhalation-induced SCG beats were detected and extracted using the Teager-Kaiser energy operator. Finally, a gradient-based beat scoring system was developed using extreme gradient boosted trees and monotonic mapping. For the rest, classical music and recovery sessions, the area under the receiver operating characteristic curve was found to be 0.978, 0.874, 0.985, respectively. On the other hand, the kernel density estimation distributions of the inhalation and exhalation scores had an overlap of 14.2%, 41.2%, 10.6%, respectively. Overall, our results show that different physiological modulations directly change the effect of respiration on the SCG morphology, thus standardization across the beats should be studied for achieving more reliable and accurate investigation of cardiovascular parameters.

Keywords: Signal pattern classification, Data mining and big data methods - Machine learning and deep learning methods, Physiological systems modeling - Signal processing in physiological systems
Abstract: Long-term acquisition of respiratory and heart signals is useful in a variety of applications, including sleep analysis, monitoring of respiratory and heart disorders, and so on. Ballistocardiography (BCG), a non-invasive technique that measures micro-body vibrations caused by cardiac contractions as well as motion caused by breathing, snoring, and body movements, would be ideal for long-term vital parameter acquisition. Turtle Shell Technologies Pvt. Ltd.'s Dozee device, which is based on BCG, is a contactless continuous vital parameters monitoring system. It is designed to measure Heart Rate (HR) and Respiratory Rate (RR) continuously and without contact in a hospital setting or at home. A validation study for HR and RR was conducted using Dozee by comparing it to the vitals obtained from the FDA-approved Patient Monitor. This was done in a sleep laboratory setting over 110 nights in 51 subjects to evaluate HR and over 20 nights in 17 subjects to evaluate RR at the National Institute of Mental Health and Neurosciences (NIMHANS), Bangalore, India. Approximately 789 hours data for HR and approximately 112 hours data for RR was collected. Dozee was able to achieve a mean absolute error of 1.72 bpm for HR compared to the gold standard ECG. A mean absolute error of ~1.24 breaths/min was obtained in determining RR compared to currently used methods. Dozee is ideal for long-term contactless monitoring of vital parameters due to its low mean absolute errors in measuring both HR and RR.