Keywords: Clinical Decision Support Systems, Artificial Intelligence, Medical Imaging
Abstract: Background/Aim: Neuroblastoma, the most common solid cancer in early childhood, exhibits significant heterogeneity, ranging from benign to aggressive forms. This paper presents a comprehensive study of a multimodal clinical predictive model for childhood neuroblastoma. Materials and Methods: The dataset, collected within the PRIMAGE project, includes 1056 patients with neuroblastoma, featuring clinical, molecular, genetic, and imaging data at diagnosis and follow-up. Radiomics features were extracted from T1/T2-weighted MR and CT images. Diffusion-weighted imaging provided apparent diffusion coefficient maps used to assess malignancy. Dynamic contrast-enhanced MR imaging was used to calculate semi-quantitative parameters for lesion characterization. Customized multiscale tumor growth models integrated initial cell density distribution, tumor vasculature, and geometry from clinical MR imaging. Results: A multimodal clinical predictive model was developed. The integration of clinical, molecular, genetic, radiological, radiomics, and multiscale tumor growth data led to a remarkable 15% improvement in predictive accuracy and a 20% boost in precision compared to models using individual data sources. Conclusions: The proposed multimodal methodology offers a more nuanced and personalized understanding of childhood neuroblastoma, promising improved outcomes and advancing research in the field of pediatric oncology.

Keywords: Medical Imaging, Artificial Intelligence, Data Science
Abstract: The ProCAncer-I project, an important initiative in the field of prostate cancer research, is exploiting the power of Artificial Intelligence (AI) to advance prostate cancer diagnosis, prognosis, and treatment. The success of AI in prostate cancer imaging within this context relies on meticulous data preparation, especially when dealing with clinical data that need to be integrated with imaging data. This paper explores the importance of data preparation to implement AI tools for prostate cancer imaging, and outlines the approach followed and challenges needed to be addressed in the context of ProCAncer-I, for ensuring data quality, privacy, and reliability.

Keywords: Medical Imaging, Health Data Science, Data Curation 
Abstract: Abstract— In Cancer Imaging research, data collection, integration, and utilization to generate multicentric data repositories pose a series of challenges such as data harmonization, quality, and suitability. This work presents the INCISIVE project approach towards assessing the quality of cancer imaging data and clinical (meta)data to serve as a map ensuring high quality data repository – a crucial factor for trustworthy AI-services development.

Clinical Relevance— The overall approach facilitates the generation of reliable, and well-harmonized cancer imaging repositories.

Keywords: Artificial Intelligence
Abstract: With the fast-growing applications of artificial intelligence (AI) in healthcare, it is essential to keep track of their credibility and reliability. We conducted a survey with healthcare practitioners to obtain requirements for developing reliable AI tools for breast cancer. We share our findings with the healthcare community, hoping that our work serves as a resource to build extensively validated and trustworthy solutions.

Keywords: Artificial Intelligence, Medical Imaging, Deep Learning
Abstract: A responsible approach to artificial intelligence and machine learning technologies, grounded in sound scientific foundations, technical robustness, rigorous testing and validation, risk-based continuous monitoring and alignment with human values is imperative to guarantee their favorable impact and prevent any adverse effects they may have on individuals and communities. An essential aspect of responsible development is transparency, which constitutes a fundamental principle of the European approach towards artificial intelligence. Transparency can be achieved at different levels, such as data origin and use, system development, operation and usage. In this paper, we present the techniques implemented and delivered in the EU H2020 ProCAncer-I project to meet the transparency requirements at the different levels required.

Clinical Relevance—This paper examines the primary transparency hurdles in artificial intelligence for medical imaging diagnostics, and presents the approaches that the EU H2020 project ProCAncer-I is taking to address them.

Keywords: Healthcare Informatics, Health Data, Data Curation 
Abstract: We present the application of HL7-FHIR to standardize a retrospective heterogeneous dataset, enhancing human/machine readability and interoperability. Clinical Relevance— The adopted strategy enables secondary use of clinical data in scientific medical research.

Keywords: Bioinformatics, Artificial Intelligence, Data Analytics
Abstract: Leveraging the untapped potential of depth information in RGB-D images, this study introduces a deep neural network classifier for advanced body shape classification. Going beyond traditional RGB image analysis, our method innovatively employs multi-task learning, simultaneously performing body shape classification, posture estimation, and body part segmentation to achieve superior accuracy. This approach promises to revolutionize personalization avenues in healthcare, fashion, and entertainment industries, establishing a new benchmark in body shape analysis.

Keywords: Artificial Intelligence, Cognitive Informatics, Generative AI
Abstract: In vision, the brain is a feature extractor that works from images. We hypothesize that fMRI can mimic the latent space of a classifier, and employ deep diffusion models with BOLD data from the occipital cortex to generate images which are plausible and semantically close to the visual stimuli administered during fMRI. To this end, we mapped BOLD signals onto the latent space of a pretrained classifier and used its gradients to condition a generative model to reconstruct images. The semantic fidelity of our BOLD response to visual stimulus reconstruction model is superior to the state of the art.

Keywords: Biomarker Discovery, Deep Learning, Bioinformatics
Abstract: Α novel feature selection approach is presented for the identification of potential biomarkers in auto-inflammatory disorders. It involves training a deep neural network for binary classification, and subsequently selecting features based on the most significant shifts in neuron weight contributions. The findings demonstrate enhanced performance against the standard false discovery rate analysis, as evidenced by the improved classification metrics (~11% increase in sensitivity) when the selected features are utilized.

Keywords: Data Visualization, Computer Aided Surgery, Big Data
Abstract: Kidney surgeries present significant complexity due to the kidneys’ susceptibility to ischemic injury and bleeding. Ensuring precise reconstruction of the renal vascular tree (RVT) is critical for patient safety and surgical outcomes. The analysis encompassed 33 corrosive endocasts as well as 12 3D artificial models, generated, 3D printed, and subjected to micro-CT, to provide both visual comparison and quantitative evaluation of the reconstruction method viability. The efficiency and accuracy of two RVT reconstruction methods, adapting circle rolling and convolution, were compared. The study also investigated the potential enhancement of combining these two methods. Results show that the convolution method offers high accuracy and short computation times. The accuracy of the reconstruction quantitatively evaluated on 3D artificial models provided promising results, i.e. dice coefficient equal to 0.8609 (std=0.0017), and accuracy equal to 0.99677 (std=0.00092). The combined method exhibited the fastest execution time among the three while, the sphere rolling method demonstrated the slowest performance. Moreover, the sphere rolling method dura- tion for the larger data exceeded the feasible time limit. Future work should focus on optimizing the convolution method.

Keywords: Artificial Intelligence, Health Data Science, Healthcare Data Integration
Abstract: Medical history forms, often lacking standardization, pose challenges for healthcare professionals. This study employs an AI-based model to digitize and categorize these forms, enhancing accessibility. Results show the model achieved 79.65% accuracy with nine documents, rising to 88.74% with 99 documents, demonstrating the potential for improved medical history documentation.

Keywords: Electronic Health Records, Wearable Devices, Evidence-Based Healthcare
Abstract: In the era of digital health, Electronic Health Records (EHRs) are becoming the norm for storing and analyzing patient data. Data not only aids in clinical decision-making but also has the potential to help in achieving better clinical outcomes during patient recovery. Recently, wearables are seen to play a crucial role in creating EHRs, especially in a home-based care setting, where traditional bulky machines cannot be deployed for continuous monitoring of patients. Pheezee®, an USFDA Cleared [ 510(k) exempt] novel prognostic app-based wearable device for the rehab sector, has established itself to have huge potential in generating patient specific ailment records in a scientific manner, thereby opening up the concept of EHRs in a Clinical Rehab setting. It consists of range of motion (ROM) sensors and surface electromyography (sEMG) bio-amplifier that can be used for assessment, monitoring and tracking of recovery of patients via shareable scientific reports. The objective is to integrate the multi-sensory information from the device into an EHR system through remote monitoring. Pheezee® acquires the movement and muscle contraction information in real-time, pushes the data to the cloud for post processing, thereby establishing a novel approach to introducing EHR in Physiotherapy. The data is encrypted, stored in a secured format and maintained through an effective database management system.

Keywords: Wearable Devices, Artificial Intelligence, Medical Imaging
Abstract: This research was designed to evaluate if it is viable to awaken children with urinary incontinence at the pre-void phase using a smart wearable device and enable them to control incontinence with fine-tuned individual parameters determined by the device intelligently. To address this research question, a miniaturised wearable smart device was built in this multidisciplinary research to monitor the non-linear behaviours of the bladder during its expansion with urine intake. The device, with its customisable abilities, sets an individual alarm point to awaken the child with incontinence before voiding. Safety parameters, aesthetics and ergonomic use of the device were investigated through hospital trials with children and the device was improved based on the obtained feedback from these trials.

Clinical Relevance: The device will help children learn how to control their incontinence over time.

Keywords: Artificial Intelligence, Medical Imaging, Health Monitoring
Abstract: Carotid-to-femoral pulse wave velocity (cf-PWV) is a crucial biomarker, essential for cardiovascular disease diagnosis and prediction. However, the standard measuring of cf-PWV is highly complex making it prone to errors and inaccuracies. In this paper, a deep learning model based on visibility graph representation obtained from the non-invasive easily measured photoplethysmogram (PPG) waveform is proposed. The obtained results illustrate the feasibility and robustness of visibility graph for image based data-driven cf-PWV estimation from non-invasive PPG signals.

Keywords: Medical Imaging, Deep Learning, Medical Diagnostics
Abstract: In this research study, we compare the predictive performance of two advanced deep learning-based models in order to provide a solution to TACE (Transarterial Chemoembolization) response prediction in HCC (Hepatocellular Carcinoma) patients. Using entire abdominal CT scans enabled a broader perspective available for the model, eliminating the need for segmentation during the pre-processing. Making use of both single-phase and multi-phase CT imaging, we have used DenseNet121 and have obtained an accuracy of 80% for the multi-phase model.

Keywords: Forecasting, Computational Biology, Deep Learning
Abstract: The development of motor brain-computer interface (BCI) critically relies on neural time series decoding algorithms. Recent advances in deep learning architectures allow for automatic feature selection to approximate higher-order dependencies in data. This article presents the FingerFlex model – a convolutional encoder-decoder architecture adapted for finger movement regression on electrocorticographic (ECoG) brain data. State-of-the-art performance was achieved on the publicly available BCI Competition IV dataset 4, with a correlation coefficient of up to 0.74 between true and predicted trajectories, significantly outperforming existing solutions. The presented method provides the opportunity for developing high-precision cortical motor brain-computer interfaces.

Keywords: Health Data Privacy And Security, Health Information Technology
Abstract: We have developed a trust data distribution platform for the health and medical data. In the health and medical fields, data quality, transparency, and reliability are important to data distribution. Especially, trustworthy data is essential for accurate predictions and diagnostic assistance using AI and machine learning. Our proposed system is based on non-fungible tokens (NFTs), which can enhance the reliability and value of data. Increased value of NFTs leads to increased value of data, which in turn becomes a valuable asset for patients and healthcare professionals. The system is designed to facilitate each stakeholder's efforts to improve data reliability.

Keywords: Diagnostics, Health Data Science, Machine Learning
Abstract: This work investigates to what extent it is possible to detect different gait restrictions compared to normal gait using a capacitive sensory floor. For this purpose, several gait parameters and a classification using Random Decision Forest (RDF) are calculated. Furthermore, the importance of the individual features for the different classes is analyzed using Recursive Feature Elimination (RFE). In this paper, different results are visible for the classification of single gaits, but results with an accuracy of up to 90.28% have been achieved.

Keywords: Deep Learning, Health Data, Clinical Decision Support Systems
Abstract: Human gait refers to the walking patterns of individuals and abnormal gait may reveal the progression of various diseases. Here, we presented the gait parameter-based deep neural network for detecting the presence of Alzheimer’s disease. Initially, the raw gait data of sixty-nine participants were recorded using pressure sensors during a free walking test on the walkway system mat. The twelve spatiotemporal features and five ratio features of temporal gait parameters were then extracted. These features were used to construct the final sequential samples by concatenating every four consecutive strides, serving as the input for four classification models: Recurrent Neural Network, Long-Short Term Memory, Bidirectional Long-Shot Term Memory, and Gated Recurrent Unit. The best performance indicated the RNN model, which was shown in the F1-score of 0.909. This study demonstrated the feasibility of utilizing gait parameters-based Deep learning models as a wide-scale screening tool for Alzheimer’s disease, complementing the conventional cognitive screening instruments, while also accelerating the integration of Artificial Intelligence in global healthcare.

Keywords: Health Monitoring, Personalized Healthcare, Machine Learning
Abstract: A continuous multiple-timestep blood glucose level prediction system based on stacked long short-term memory (LSTM) is proposed for high-accuracy hypoglycemia alerts in smart contact lenses. A deeper representation of input data is provided by using stacked LSTM rather than conventional vanilla LSTM, enabling higher accuracy. Python programming with the Keras library was implemented to evaluate the proposed system. The results showed a 16% improvement compared to the conventional approach.

Keywords: Deep Learning, Medical Diagnostics, Neural Networks
Abstract: Electrocardiogram (ECG) abnormality classification is to detect various types of clinical abnormalities from ECG. This paper proposes a deep neural network (DNN)-based ECG abnormality classification architecture where ResNet and DenseNet are cascaded. ResNet in the proposed architecture comprises residual U-shaped (ResU) blocks that effectively capture multi-scale feature maps without significantly increasing neural parameters. In addition, we use a multi-head self-attention (MHSA) to ensure that the model focuses on essential features in the given ECG. Experimental results show that our proposed model has superior ECG abnormality classification performance compared to other recently proposed DNN-based models.

Keywords: Wearable Devices, Machine Learning, Data Science
Abstract: In this study, we investigate the relationship between blood glucose levels and visual acuity during simulated driving using smart glasses and machine learning. We found compelling correlations between blood glucose levels and visual performance, with a classification accuracy of 98.70%. These results highlight the importance of blood glucose regulation for road safety and human well-being, and the potential of wearable sensor technology to advance our understanding of physiological influences on human performance.

Keywords: Artificial Intelligence, Biomarker Discovery, Bioinformatics
Abstract: Abstract— The development of novel therapeutic approaches that target gene alterations in melanoma cells remains an area of active research. This study prioritized ten genes out of 76 genes - associated with melanoma through a transcriptome-wide association study (TWAS). We investigated the potential of the identified ten genes in prognosis of metastasis in melanoma patients.

Clinical Relevance— Our findings have clinical implications for melanoma treatment strategies. Targeted therapies that specifically address the dysregulation of our 10 genes may hold promise in preventing or slowing melanoma metastasis.

Keywords: Clinical Decision Support Systems, Artificial Intelligence, Decision Support
Abstract: We developed a deep neural network (DNN) model to detect sample misidentification error using previous and current clinical chemistry test results as input and compared the detection performance with conventional methods such as DPC and absDPC. DNN models were developed to detect sample misidentification errors for each of the five tumor markers; AFP, CA19-9, CA125, CEA, and PSA. As a result of 1,000 times of repeated simulations of 1% random sample misidentification test, the DNN model achieved an AUC of ≥0.802. Moreover, the accuracy of DNN model was ≥0.732, which is higher than the DPC (≥0.646) and absDPC (≥0.509).

Keywords: Health Monitoring, Deep Learning, Healthcare
Abstract: Skin cancer is one of the deadliest types of cancers that need to be detected at an early stage as it can widely spread affecting other organs of the body. Thus, our objective is to develop an efficient mobile application dedicated to the early detection of skin cancer in primary healthcare at a low cost. In the design of an Android-based mobile application, EfficientNetV2 deep learning model achieved the highest accuracy and feasibility of deployment as an offline model compared to other CNN pretrained models. The designed deep learning system classifies three classes: healthy, benign, and malignant skin lesions, and further classification of malignant lesions into Basal Cell Carcinoma, Melanoma, or Nevus (the most common skin cancer malignancies) is performed. The developed two-stage classification system has achieved the highest accuracy with 91% and 90% using the EfficientNetV2-model in the designed application

Keywords: Artificial Intelligence, Machine Learning, Medical Imaging
Abstract: This work proposes a machine learning (ML) framework to classify the hair type of dermatoscopy images into four classes using the imaging features taken from the binary hair contour masks. Furthermore, the optimal kernel of the black-hat hair removal algorithm is then examined through the structural similarity index measure (SSIM) between the original and the pre-processed image. The best performance of the classification model in terms of ACC and AUC was obtained by the SVM classifier, achieving 80% and 79.8%, respectively. A kernel size of up to 20 by 20 is proposed for image filtering without significant loss of texture information in the lesion.

Keywords: Data Curation , Health Analytics, Artificial Intelligence
Abstract: The goal of pose estimation in computer vision is to detect and classify the joints of the human body to describe the pose of a person. During running, pose estimation can provide valuable insights into the running form and identify abnormalities or areas for improvement. Most existing datasets are image-based and contain at most 17 keypoints and miss important joints, such as heels and toes, which are essential for properly analysing the running form. This paper proposes a new video dataset annotated with 26 body key points to facilitate the benchmarking of pose estimation in this area and promote new opportunities for further research on the running form.

Keywords: Artificial Intelligence, Big Data Analytics, Diagnostics
Abstract: This study aims to develop a machine learning model to diagnose atrial fibrillation(AF) using only medical information that can be obtained during routine care or medical examinations. We obtained the age, sex, body mass index, and presence of hypertension, stroke, or coronary artery disease of a total of 404,898 subjects, including 6,661 AF patients, from UK Biobank data and developed and validated an XGBoost-based model to diagnose AF. The developed model showed an average AUROC of 0.785 for diagnosing AF, with the following factors having the greatest impact on the results, in this order: age, sex, hypertension, body mass index, coronary artery disease, and stroke.

Keywords: Artificial Intelligence, Bioinformatics, Machine Learning
Abstract: A methodology for the identification of non-small cell lung cancer from blood samples, combining feature selection methods followed by Graph Convolutional Networks (GCN) with Genetic Algorithm (GA) optimization is presented. The methodology was tested on RNA-seq data from the GSE207586 dataset. The evaluation results showed that mixing the top 100 features from different feature selections and modeling with GCN offered the highest identification performance among all evaluated setups.

Keywords: Machine Learning
Abstract: Carotid artery disease refers to the build up of atherosclerotic plaques into the carotid arteries. This pathological condition is responsible for cerebral events and stroke. The diagnosis of the disease is based on imaging techniques (carotid ultrasound, computed tomography angiography and magnetic resonance angiography) and the disease is either symptomatic or asymptomatic. The aim of the presented study is to early identify individuals of high risk for carotid artery disease through utilizing data driven techniques. More specifically, we utilize a dataset of 135 individuals with carotid artery disease and healthy ones, as it is defined with the use of carotid ultrasound. The dataset includes both typical medical records (demographics, risk factors) and biomarkers (biochemical, hematological) and is trained into a gradient boosting, aiming to detect the healthy participants and the individuals of high risk. The pipeline includes the pre-processing of the dataset, the handling of the class imbalance, the implementation of feature ranking techniques and finally the evaluation of the proposed pipeline. The model achieved an accuracy of 0.82 and a sensitivity and specificity of 0.68 and 0.92, respectively.

Keywords: Medical Imaging, Big Data, Artificial Intelligence
Abstract: Thorax disease analysis in large-scale, multi-centre, and multi-scanner settings is often limited by strict privacy policies. Federated learning (FL) offers a potential solution, while traditional parameter-based FL can be limited by issues such as high communication costs, data leakage, and heterogeneity. Distillation-based FL can improve efficiency, but it relies on a proxy dataset, which is often impractical in clinical practice. To address these challenges, we introduce a data-free distillation-based FL approach FedKDF. In FedKDF, the server employs a lightweight generator to aggregate knowledge from different clients without requiring access to their private data or a proxy dataset. FedKDF combines the predictors from clients into a single, unified predictor, which is further optimized using the learned knowledge in the lightweight generator. Our empirical experiments demonstrate that FedKDF offers a robust solution for efficient, privacy-preserving federated thorax disease analysis.

Keywords: Medical Imaging, Neural Networks, Deep Learning
Abstract: Deep Learning networks can be used to rapidly estimate temperature in order to perform real-time safety assessment in MRI. In this work, we have developed two Deep Learning networks that, using as input 5 thermal parameters maps, can estimate the spatial distribution of the baseline temperature of the patient, which corresponds to the temperature before the beginning of the MRI scan. One network is based on the analysis of 2D matrices, and another on 3D matrices. The 2D network could predict the temperature with a percent MSE between 8.2% and 15.3%, while the 3D network with a percent MSE between 5.2% and 8.0%. The 2D network could predict accurately the temperature in the head, while the 3D network also in the shoulders of the body model.

Keywords: Machine Learning
Abstract: A stroke, also known as brain attack, occurs when blood supply to your brain is interrupted. Primary prevention relies on prompt prediction of a stroke. While currently there are several clinical risk scores, machine learning (ML) models seems to be more suitable tools for accurate prediction of stroke events. Therefore, this work focuses on the prediction of stroke within 7 years follow-up in patients who have not suffered from a stroke or TIA event at baseline. LightGBM (LGBM), Extreme Grading Boosting (XGBoost), Support Vector Machine (SVM) and Decision Tree were employed in the getABI dataset, which includes 5,897 participants. The performance of models was calculated by Accuracy (ACC), Sensitivity (SENS), Specificity (SPE) and area under the receiver operating characteristic curve (AUC) of each model. According to the comparison analysis's results, LGBM has been shown to be the most trustworthy algorithm, with accuracy 68 %. Moreover, sex, age, status of peripheral artery disease (PAD), history of myocardial infarction, angina pectoris, amputation and diabetes and pulse status of different arteries can be used as a simple and cost-effective way to predict stroke.