ABSTRACT. We investigate fine-tuning Vision-Language Models (VLMs) for multi-task medical image understanding, focusing on detection, localization, and counting of findings in medical images. Using MedMultiPoints, a multimodal dataset combining annotations from endoscopy (polyps and instruments) and microscopy (sperm cells), we reformulate these tasks as instruction-based prompts suitable for vision-language reasoning. We fine-tune Qwen2.5-VL-7B-Instruct using Low-Rank Adaptation (LoRA) across multiple task combinations. Results show that multi-task training improves accuracy and robustness for example, reducing Count Mean Absolute Error (MAE) and increasing Matching Accuracy in the Counting + Pointing task. However, we also observe trade-offs, such as increased zero-case point predictions despite improved average accuracy—indicating reduced reliability on edge cases as a trade-off for better performance on common samples. Our study highlights the potential of adapting general-purpose VLMs to specialized medical tasks via prompt-driven fine-tuning, achieving improved object counting and localization performance when combining tasks during training. The model retains interpretable, structured outputs, making it a promising step toward explainable, versatile medical AI. Code, model weights and evaluation protocols will be released for reproducibility.

ABSTRACT. According to the World Health Organization's (WHO) 2023 statistics, approximately 5% of people worldwide experience depression. Early diagnosis is crucial. However, misdiagnosis and delayed diagnosis are common because of professional subjectivity and reliance on patient responses. To address this issue, audio- and text-based methods for depression prediction have been a focus of recent research. However, previous methods are limited in generalizability, adaptability to new data, and prediction accuracy because they cannot fully reflect an individual’s speech habits and symptom levels. To overcome this problem, this study proposes a liquid feed-forward neural network-enhanced multimodal former (LEFORMER) that incorporates an individual's symptom scores, along with learnable and dynamic parameters, into the transformer block. The LEFORMER consists of two main blocks: the symptom prediction block, which predicts patients' symptoms and incorporates third-party assessments, and the audio-text interaction block, which captures depression-related speech patterns while accounting for individual speech habits. In the depression score prediction experiment based on DAIC-WOZ, the LEFORMER achieved an MAE of 2.87 and an RMSE of 4.12, demonstrating an improvement of 0.4 in MAE and 0.71 in RMSE compared to previous studies.

ABSTRACT. As the aging population grows, effective healthcare communication becomes increasingly critical, particularly for older adults managing multimorbidity (multiple chronic conditions). Traditional methods often fail to engage this demographic, leading to misunderstandings, inefficient care coordination, and increased provider workload. This paper presents a generative AI-driven solution integrating a multimodal chatbot and an AI-enhanced provider dashboard to bridge this gap. The chatbot employs a hybrid architecture combining intent-driven logic with large language model (LLM)-powered natural language understanding (NLU) for safe, context-aware interactions, while the dashboard synthesizes patient–chatbot dialogues, extracting key insights like sentiment trends, discussion topics, and tone analysis to aid clinical decision-making. Additionally, an LLM-assisted virtual meeting room enables real-time transcription, patient history summarization, interactive querying of past interactions, and streamlining consultations. By leveraging conversational AI, real-time analytics, and AI-assisted care coordination, this scalable solution enhances accessibility, promotes independent living, and improves provider efficiency, offering a transformative approach to patient-centered healthcare for aging populations.

ABSTRACT. Vision-Language Models (VLMs) excel in multimodal tasks, yet their effectiveness in specialized medical applications remains underexplored. Accurate interpretation of medical images and text is crucial for clinical decision support, particularly in multiple-choice question answering (MCQA). To address the lack of benchmarks in this domain, we introduce the Multilingual Multimodal Medical Exam Dataset (MMMED), designed to assess VLMs' ability to integrate visual and textual information for medical reasoning. MMMED includes 582 MCQA pairs from Spanish medical residency exams (MIR), with multilingual support (Spanish, English, Italian) and paired medical images. We benchmark state-of-the-art VLMs, analyzing their strengths and limitations across languages and modalities. The dataset is publicly available on Hugging Face (https://huggingface.co/datasets/praiselab-picuslab/MMMED), with experimental code on GitHub (https://github.com/PRAISELab-PicusLab/MMMED).

ABSTRACT. Creation of models to generate synthetic data has represented a significant advancement across diverse scientific fields, but this technology also brings important privacy considerations for users. This work focuses on enhancing a non-parametric copula-based synthetic data generation model, DP-NPC, by incorporating Differential Privacy through an Enhanced Fourier Perturbation method. The model generates synthetic data for mixed tabular databases while preserving privacy. We compare DP-NPC with three other models (PrivBayes, DP-Copula, and DP-Histogram) across three public datasets (income with sociodemographic information, criminal defendant’s likelihood data, and Hospital data considering utilization, charity, and admission data), evaluating privacy, utility, and execution time. DP-NPC outperforms others in modeling multivariate dependencies, maintaining privacy for small $\epsilon$ values, and reducing training times. However, limitations include the need to assess the model's performance with different encoding methods for categorical variables and consider additional privacy attacks. Future research should address these areas to enhance privacy-preserving synthetic data generation.

ABSTRACT. The growing use of patient data in research underscores its value (for instance, in training AI). It also highlights the need for strong anonymization when health datasets are released publicly due to the risk of de-anonymization attacks. Electrocardiograms (ECG) are widely used, and real patient data have been openly released anonymously. However, ECGs are susceptible to linkage attacks, raising concerns around privacy, non-compliance with regulations such as the General Data Protection Regulation (GDPR), and loss of trust in digital healthcare. In this paper, we present a novel lightweight de-anonymization linkage attack on ECGs, and discuss benchmarking routes and an inclusive privacy protection framework that can be used in mitigating de-anonymization risks. The proposed matching attack leverages Convolutional Neural Networks (CNN)-based and ECG-specific features, and was tested on three open datasets: ECG-ID, MIT-BIH and MIMIC-IV. Unlike authentication-focused works, our study evaluates re-identification from an adversarial perspective, quantifying the risk on anonymized datasets based on metrics that establish a benchmarking baseline. Experimental results demonstrate an average matching accuracy of 97.22%, and nearly 100% for the best result, on the MIT-BIH dataset, for which previous results exist in the literature. Our results are substantially higher than the previous best-performing attack, which achieved an 81.9% accuracy. Consistent results on other datasets demonstrate the generality of our approach. The attacks emphasize evaluating de-anonymization risks before publicly releasing datasets. Based on our findings, we formalize recommendations into a new privacy-by-design framework resilient against real-world de-anonymization attacks, including inclusive processes to guide stakeholders in assessing requirements and offering insights into privacy metrics and improvement axes.

ABSTRACT. Health institutions play a critical role in providing essential healthcare services. However, despite the numerous benefits of benchmarking, many healthcare institutions harbor reservations about openly sharing these metrics. One predominant concern is the potential for retaliatory actions, be it from regulatory bodies, competitors, or the public. In this paper, we propose a new methodology that allows institutions to compare performance metrics without disclosing the actual values. The method is based on clustering, which involves grouping health institutions’ outcomes into a known number of clusters, allowing institutions to position themselves in a range of clusters without sharing the true means of their target data. The proposed method uses the K-means and K-modes clustering algorithms and was tested on data from real Electronic health records and public datasets. This approach provides a valid benchmark of hospital metrics and performances while protecting the privacy of participating institutions. By leveraging new approaches to benchmarking, hospitals can continuously improve their operations and ensure they provide the highest quality of care possible.

ABSTRACT. The increasing use of mobile technologies in healthcare has driven significant advancements in patient care management while raising critical concerns about data security and privacy. This study addresses the challenges and solutions for ensuring security and privacy in mobile health applications, emphasizing the importance of integrating Privacy by Design (PbD) principles from the development phase. By implementing a framework in Flutter, the proposed approach focuses on safeguarding sensitive data through measures such as screenshot prevention and data encryption, ensuring strict compliance with regulations like the General Data Protection Law (LGPD). The results demonstrate that adopting PbD not only meets legal requirements but also strengthens user trust by effectively protecting personal data. The proposed framework establishes a new standard for developing mobile health applications, ensuring that security and privacy are integral throughout the entire design and operational process.

ABSTRACT. Virtual Reality (VR) serious games offer significant potential for augmenting existing methods in medical education and therapy. Despite their advantages, widespread adoption in healthcare remains limited due to the substantial expertise, time, and financial investment required for development. In this paper, we introduce VIA-VR, a novel platform that empowers healthcare professionals to create immersive VR serious games by enabling collaborative workflows where software engineers develop reusable building blocks that healthcare experts can assemble into tailored VR experiences. Our work presents both the platform’s concept, including the idea, workflow, and architecture, and the initial implementation that can be built upon. In particular, VIA-VR targets applications where a trainee or patient is immersed in a virtual environment (VE) using a consumer-grade head-mounted display (HMD), and a supervisor monitors the session—often the same individual who has authored or adapted the VR experience. To minimize development efforts and ensure long-term extensibility of the provided functionalities, we combine established software solutions with new domain-specific modules. VIA-VR’s features include a VE editor, a story editor, a polygon reduction tool for geometry optimization, a behavior editor based on building blocks (eliminating the need for scripting), a photogrammetry-driven virtual character creation tool, and a physiological data measurement engine. We present two minimal demos that have been created with the first iteration of the platform: a VR exposure therapy application for pyrophobia and a gait rehabilitation exercise application. As an open-source platform, VIA-VR strives to contribute to open standards for developing medical VR serious games, aiming to accelerate innovation and improve accessibility in healthcare.

ABSTRACT. Parkinson’s disease is a neurodegenerative disorder that significantly affects the motor function of those affected, with bradykinesia being one of its central manifestations. Accurate modeling of this cardinal symptom is essential for diagnosing and monitoring the disease. The Movement Disorder Societysponsored revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), considered the gold standard in clinical settings, presents several associated limitations, such as subjectivity and susceptibility to interobserver variability. This study proposes a novel pipeline for performing an automated analysis of lower limb bradykinesia in videos from complete routine clinical consultations in which the whole MDS-UPDRS is analyzed, utilizing computer vision techniques. The videos are segmented into temporal actions to identify tasks on the scale intended to assess lower limb bradykinesia, which are subjected to a process of person detection and subsequent pose estimation to extract kinematic features that facilitate modeling each task. The algorithm was evaluated using videos comprising PD patients and control subjects, demonstrating significant motor differences between the two groups, particularly regarding the toe-tapping test. This proposed approach enables an objective and automatic assessment of lower limb bradykinesia, showing promise for remote monitoring and longitudinal follow-up of motor symptoms associated with this pathology

ABSTRACT. Using simulation for medical education improves clinical practice and creates a safe environment for students to develop skills. High-fidelity patient simulators, as one kind of simulation tool, use advanced life-like manikins. A high-fidelity simulation session includes pre-briefing, simulation scenarios, and debriefing. While pre-briefing and debriefing are widely recognized, the reflection-in-action during simulation is often overlooked despite its importance. Reflection-in-action is the ability to reflect on the student and adapt during the simulation. One major issue regarding reflection-in-action is the limited number of instructors at the university compared to the large number of students. In addition, the intensive time required for the simulations further constrains the capacity of the instructors. To address this critical gap, we propose using a proactive computing approach to provide real-time support for decisions in high-fidelity medical simulations. By implementing the proactive system and providing immediate adaptive feedback to students based on sensor data, we hope to assist teachers and improve learning outcomes.

ABSTRACT. This study evaluates the impact of immersive (Head-Mounted Display, HMD) and non-immersive (2D conventional monitor) visual displays on user performance in a haptic-enabled virtual reality (VR) platform. Fourteen healthy participants completed three interactive tasks—Pick and place (P&P), Canvas, and Box and Block Test (BBT)—under both display conditions. A counterbalanced design was employed to minimize order effects. Statistical analysis revealed that display type significantly influenced performance in the P&P task (p = 0.003), where users performed better with the HMD, likely due to its enhanced depth perception capabilities. However, no significant differences were observed in Canvas (p = 0.258) or BBT (p = 0.634), where depth perception plays a less critical role. Despite its potential effectiveness for rehabilitation, the platform does have limitations, such as the constrained workspace of the TouchTM haptic device. Future studies should involve clinical populations to evaluate its therapeutic potential. These findings underscore the influence of display modality on VR-based tasks and its implications for human-computer interaction, particularly in rehabilitation contexts.

ABSTRACT. Current models for predicting waitlist mortality among liver transplant (LT) candidates primarily rely on conventional statistical regression-based approaches, which are typically developed separately for hepatocellular carcinoma (HCC) and non-HCC patients. These linear models may fail to capture complex, nonlinear relationships in the data, limiting their predictive performance. In this study, we evaluate and compare existing clinical scoring systems against various machine learning (ML) models, including both linear and nonlinear approaches, with a particular focus on a Multi-Task Learning (MTL) framework. Our results demonstrate that MTL outperforms both conventional methods and single-task learners across HCC and non-HCC groups. Furthermore, by leveraging SHapley Additive Explanations (SHAP), we provide deeper insights into the MTL model’s decision-making process, offering both global and local explanations while pinpointing key risk factors for waitlist mortality in both patient groups. This study highlights the potential of advanced ML methodologies to enhance LT organ allocation and underscores the need for their broader adoption in clinical practice.

ABSTRACT. —Radiomic analysis of 3D medical images has gained traction as a non-invasive approach to tumor characterization and outcome prediction. However, anisotropic voxel geometry remains a critical source of non-biological variability in radiomic features. Isotropic resampling is widely used to address this issue, but may introduce interpolation artifacts that compromise features fidelity. This study evaluates four preprocessing strategies—no resampling, isotropic resampling, voxel spacing correction, and voxel spacing weighting—for radiomic feature extraction from 685 pulmonary nodules obtained from the LIDC-IDRI computed tomography (CT) dataset. A total of 94 original features (firstorder, shape, and texture) were extracted using PyRadiomics. We assessed feature significance, stability, and predictive power using statistical tests, reproducibility metrics (CV, ICC), and machine learning classifiers (Logistic Regression, Random Forest, XGBoost, LightGBM). SHAP analysis was performed for model interpretability. Spacing-aware methods yielded better feature reproducibility and stronger statistical separability in 23 of 94 tested features. The best overall performance was achieved by Logistic Regression with voxel spacing correction, reaching a composite score of 1.50. SHAP results confirmed alignment between statistical robustness and model relevance. These findings support voxel spacing–aware preprocessing as a robust alternative to isotropic resampling, particularly in heterogeneous CT datasets where spatial fidelity and reproducibility are critical.

ABSTRACT. This study presents the implementation and evaluation of YOLOv11x, a deep learning object detection algorithm, for automated renal stone detection on kidney-ureter-bladder (KUB) radiographs. A dataset comprising 1,081 KUB radiographs (523 containing renal stones, 558 without) from a Thai population cohort was partitioned into training (n=881), validation (n=100), and testing (n=100) sets. The model was developed using transfer learning from COCO-pretrained weights with hyperparameters optimized specifically for urolithiasis detection. Quantitative performance assessment demonstrated precision of 90.82%, sensitivity of 84.76%, and an F1-score of 87.68% at an Intersection over Union threshold of 0.35. The detection-based approach exhibited superior performance compared to conventional patch-based classification methods, particularly in precision metrics. Notably, this represents the first implementation of YOLOv11x for renal stone detection in KUB radiographs within a Thai population. The results establish the efficacy of object detection frameworks for clinical screening applications in resource-limited settings where radiation exposure and cost constraints render computed tomography suboptimal as a primary diagnostic modality.

ABSTRACT. Skin cancer cases has increased in the last years and melanoma cases have a high rank in the incidence of mortality. However class imbalance in melanoma dataset produces a challenge for researcher to tackle a classification model. This work proposes the use of a dynamic ratio in dataset imabalance with the addition of synthetic image generation by Progressive Growth of Generative Adversarial Networks (PGGAN) and the assessment of this ratio problem. Our methodology involved the preprocessing of the ISIC 2019 challenge dataset, selection between PGGAN and Wasserstein GAN with Gradient Penalty (WGAN-GP) for image generation, dynamic ratio addition in original dataset, use of Resnet 50 pretrained model and evaluation of classification metrics. Our study demonstrated selected ratios that improved precision , recall and f1 score metrics on melanoma detection against original dataset with traditional image transformations.

ABSTRACT. The development of new drugs is a costly and time-consuming process, often spanning over a decade. Drug repurposing has emerged as a promising alternative, leveraging existing compounds to identify novel therapeutic uses. In this context, DRIVE (Data-dRiven platform for dIsease Visualization and Drug rEpurposing) is a platform designed to integrate and exploit heterogeneous biomedical data to support hypothesis generation in drug repurposing. Developed at the Medical Data Analytics Laboratory (MEDAL) of Universidad Politécnica de Madrid, DRIVE combines disease-centered network visualizations and six complementary computational methods, ranging from datadriven-based approaches to graph neural network models. The platform is powered by data from DISNET and integrates phenotypic, molecular, and pharmacological layers of knowledge. Users can interactively explore disease mechanisms, visualize multi-layer disease networks, and obtain ranked repurposing candidates through a web interface. This poster presents the platform architecture, discusses the methodologies implemented, and illustrates the capabilities of DRIVE through representative use cases.

ABSTRACT. A third of America’s elderly die of Alzheimer’s or dementia every year. Treatments like MRI scans and injections can cost thousands of dollars and fail to accommodate individual patient needs. Moreover, prior literature has prioritized disease detection without insight into its spread, making treatment less effective over time. While Alzheimer’s produces physical changes in the brain, it also causes behavioral changes such as increased repetitiveness or irritability. These changes manifest themselves in writings, a form of emotional expression. The researcher harnessed this phenomenon to develop MindMap, the first web-based tool that maps the progression of Alzheimer’s disease in the brain based on the psychological features exhibited in diary entries of early-stage patients. The researcher assessed around 1000 diary entries across a set of patients (n=10) and evaluated the presence of various psychological features such as positivity and sensitivity over time. These behavioral biomarkers were used to develop personalized forecasting models to map the progression of the psychological features within each patient. Numerous models (ARIMA, SARIMAX, and ExponentialSmoothing) were trained and hyperparameter-tuned for this task. The best performing models achieved average MAE, RMSE, and MAPE of about 0.609, 0.775, and 25.4% respectively on patient test data, indicating high predictive accuracy for each psychological feature. Moreover, the Mann-Whitney U test revealed a high average correlation between forecasts and true progression (p=0.881). These insights contributed to the development of the MindMap tool, which successfully forecasts Alzheimer’s through human expression and ultimately helps with targeted drug therapy to the affected geographic brain regions.

ABSTRACT. The present study has identified 6 polymorphisms associated with variability in response to ovarian stimulation in young women with good ovarian reserve indicators using machine learning. Women who carry these genetic variants may be suitable candidates for personalised ovarian stimulation treatments to help prevent inadequate responses.

ABSTRACT. Surrogate outcomes are increasingly used in clinical trials to accelerate drug approvals, yet the reporting of their use as primary endpoints in study protocols remains inconsistent. In a clinical trial, the primary endpoint is the most significant measure used to evaluate the primary objective of a study. The primary endpoint is used to establish treatment efficacy. The primary endpoint is determined before the start of a study and is important for the interpretation of results and making regulatory decisions. A surrogate outcome is a substitute measure for a patient final outcome, that is, a direct measure of how a person feels, functions and/or survives. Surrogate outcomes are typically a biomarker. An intermediate outcome can be defined as a standardised functionality measure. Any surrogate, patient final or intermediate outcome could act as a primary endpoint measure in a clinical trial. Our study investigates the application of machine learning approaches to determine, using a multiclass classification approach, where a surrogate outcome has been used as the primary endpoint in nervous system clinical trial protocols. We leveraged a manually labelled dataset of European Union Clinical Trials Nervous System protocols to train and evaluate supervised learning classifiers using leave-one-out cross-validation (LOOCV). The best-performing model that we evaluated was Complement Naïve Bayes with TF-IDF embeddings, achieving the highest recall for surrogate outcome detection. Furthermore, we applied the trained classifier to an unlabelled dataset of nervous system Health Research Authority research ethics committee forms and applied an active learning framework to iteratively refine predictions whilst lessening annotation effort. Our results conclude that whilst CNB can effectively identify surrogate outcome usage, active learning displayed fluctuating recall, highlighting the challenges in identifying surrogate endpoint usage with the text data included in protocols. This study serves as a proof of concept for the potential of machine learning to automate the detection of primary surrogate endpoint use in clinical trial protocols, thus addressing gaps in reporting.

ABSTRACT. This study investigates patient’s perspective on the use of AI in healthcare and the role of Explainable AI in this context. Through a co-creative workshop with six participants from diverse disciplines, we investigated the impact of transparency on trust. The findings highlight parallels between AI and doctors as "black boxes," the complexity of informed consent and the importance of emotional safety. This work serves as a starting point for ongoing research that engages diverse stakeholder groups, to ensure the development of user-centered XAI solutions that can be effectively implemented in clinical practice.

ABSTRACT. Despite the widespread adoption of electronic medical records (EMR), the variations in format and terminology across institutions hinder inter-institutional comparisons and feature extraction. This paper proposes a demonstration of extracting frequent disease-specific instruction sequences and efficiently visualizing them with statistical insights, e.g. statistical trends, and abnormal inspection result rates from real multi-institutional EMR data. The utility of the developed visualization tool was presented for decision support and improving clinical processes.

ABSTRACT. The molecular classification of adult-type diffuse gliomas is essential for determining appropriate therapeutic strategies, but genomic sequencing remains costly. Recent advances in digital pathology and deep learning have led to several studies exploring molecular classification using multiple instance learning (MIL) on whole slide images (WSIs). However, achieving optimal classification performance using only histological slides is challenging due to the lack of guidance from genomic data. In this study, we propose a teacher-student distillation framework for glioma molecular classification using WSIs. Our method leverages a pretrained self-normalizing neural network (SNN) as the genomic teacher model, which selects genes based on survival analysis-driven criteria to guide the MIL-based student model in learning effective histological representations. During training, both genomic and pathological data are utilized, while inference relies solely on WSIs. Experimental validation on the TCGA GBM-LGG datasets shows that our approach outperforms state-of-the-art (SOTA) MIL models, highlighting its effectiveness in glioma diagnostic subtyping using WSIs.

ABSTRACT. Diagnosing fractures from medical images is challenging due to the limited availability of large, annotated datasets and the inherent variability across imaging modalities, such as CT and X-ray. Generating synthetic images that combine complementary information from different sources has the potential to enhance diagnostic accuracy. In this work, we propose a GAN-based multimodal image fusion framework to generate synthetic X-ray images from CT scans. The generated images were evaluated both qualitatively and quantitatively by comparing them with real X-rays using metrics such as MSE, PSNR, and SSIM. To evaluate the effectiveness of the fused data, we trained a ResNet-18 classifier to differentiate between fractured and non-fractured knees, incrementally augmenting the original image with additional fused channels. The results showed a clear improvement in classification performance when fused modalities were included, particularly when two or three fusion outputs were combined. This approach demonstrates significant potential for advancing diagnostic tools in medical imaging, particularly when multi-modal data is limited or unpaired.

ABSTRACT. Generating positron emission tomography (PET) images from computed tomography (CT) scans via deep learning offers a promising pathway to reduce radiation exposure and costs associated with PET imaging, improving patient care and accessibility to functional imaging. Whole-body image translation presents challenges due to anatomical heterogeneity, often limiting generalized models. We propose a framework that segments whole-body CT images into four regions-head, trunk, arms, and legs-and uses district-specific Generative Adversarial Networks (GANs) for tailored CT-to-PET translation. Synthetic PET images from each region are stitched together to reconstruct the whole-body scan. Comparisons with a baseline non-segmented GAN and experiments with Pix2Pix and CycleGAN architectures tested paired and unpaired scenarios. Quantitative evaluations at district, whole-body, and lesion levels demonstrated significant improvements with our district-specific GANs. Pix2Pix yielded superior metrics, ensuring precise, high-quality image synthesis. By addressing anatomical heterogeneity, this approach achieves state-of-the-art results in whole-body CT-to-PET translation. This methodology supports healthcare Digital Twins by enabling accurate virtual PET scans from CT data, creating virtual imaging representations to monitor, predict, and optimize health outcomes.

ABSTRACT. Glioblastoma multiforme (GBM) is a high-grade glioma with poor overall survival and high local recurrence rates. Standard treatment includes surgical removal of the tumor followed by radiotherapy (RT) and adjuvant chemotherapy. In RT planning, general target volume is manually defined using contrast-enhanced T1-weighted (T1-ce) and T2-FLAIR MR images to locate the residual enhancing tumor region and extend contours on visible perifocal edema, respectively. However, planning process requires strong knowledge and long times to clearly define tumor contours. This study aims to find the best candidate between several open-source segmentation networks to build an automatic tumor segmentation tool for the enhancing tumor delineation in RT planning using T1-ce MR sequences. For that purpose, four deep learning models including nnU-Net, TransUNet, SwinUNETR and MedNeXt, were trained and tested under different hyperparameter’s conditions using BraTS public dataset and an additional private cohort of 43 patients as external test. nnU-Net and MedNeXt showed best performances reaching same Dice Score (DSC) in both validation (0.89) and external test (0.81) sets. A peritumoral edema segmentation model on T2- FLAIR images built on nnU-Net’s best configuration was also implemented with success.

ABSTRACT. Antimicrobial Resistance (AMR) poses a significant global public health challenge, necessitating early detection strategies to enable timely clinical interventions. Electronic Health Records (EHRs) offer extensive real-world clinical data but present challenges due to their irregularly sampled, heterogeneous, and multivariate temporal structure. This paper investigates graph-based learning models to predict AMR in Intensive Care Unit patients by systematically modeling spatial and temporal dependencies within EHR data represented as Multivariate Time Series. We propose and evaluate a novel Spatio-Temporal Graph Convolutional Neural Network architecture, demonstrating its superior predictive performance by achieving a Receiver Operating Characteristic Area Under the Curve of 80.00%, surpassing baseline models by approximately 6%. Furthermore, our analysis of the learned graph structures highlights critical clinical interactions, notably emphasizing catheter-related variables as central nodes, aligning well with established clinical knowledge. By combining high predictive performance with enhanced interpretability, our approach presents a robust and transparent framework, well-suited for clinical applications aimed at improving AMR risk assessment and patient care management.

ABSTRACT. This work proposes a novel approach to predicting previously unobserved side effects in drug-drug interactions (DDIs) by identifying them as Out of Distribution (OoD) samples. The rise of polypharmacy has increased both the number of drug combinations administered to patients and the risk of DDIs, many such interactions remain undocumented in clinical trials or medical records, posing a significant challenge for patient safety. While traditional DDI prediction is typically framed as a supervised machine learning problem, novel side effects lack labeled data. To address this, we introduce a hybrid approach that combines diverse drug representations with Out of Distribution (OoD) detection to capture unseen interactions. Our method integrates multiple drug embeddings generated from diverse characteristics, including biomedical texts, molecular structures, knowledge graphs and similarity profiles to generate rich latent representations that enable us to predict the presence of novel DDIs as OoD samples.

ABSTRACT. Cardiovascular disease (CVD), particularly heart failure (HF), remains a significant public health challenge in Europe, necessitating new therapeutic strategies. Current treat- ments have limited efficacy, especially for heart failure with preserved ejection fraction (HFpEF) and acute HF. Addition- ally, cardiovascular events induced by cancer therapies further complicate treatment outcomes. To address these challenges, we developed a comprehensive structure-based virtual screening (SBVS) pipeline integrating artificial intelligence-driven protein structure prediction, extensive ligand databases, high-throughput ligand screening, and robust molecular dynamics simulations. Our computational framework leverages high-performance com- puting resources and state-of-the-art deep learning techniques to enhance the accuracy and efficiency of candidate identification and validation. Moreover, a user-friendly web application ensures broad accessibility, facilitating the translation of complex bioin- formatics results into actionable therapeutic discoveries. This integrated approach significantly accelerates the identification of promising new therapies for heart failure.

ABSTRACT. Antimicrobial resistance (AMR) is a growing global health challenge that necessitates accurate computational methods for predicting resistant phenotypes from genomic data. In this study, we evaluate the performance of traditional machine learning (ML) models—Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbors (k-NN)—alongside deep learning (DL) approaches, including Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN). We also introduce TabTransformer as an additional model for AMR classification. Our models leverage k-mer-based features, resistance gene presence/absence, and SNP data across three feature sets. Our results show that LR, RF, and SVM consistently achieved an over 90% accuracy, while k-NN underperformed (77.78%) due to sensitivity to high-dimensional feature spaces. Among deep learning models, RNN demonstrated superior performance in high-dimensional settings (97.22% accuracy on 1024 features), while CNN and MLP experienced performance declines. TabTransformer also exhibited robust accuracy across different feature sets, outperforming CNN and MLP. Increasing the number of selected features beyond 256 did not significantly improve performance, emphasizing the importance of feature selection. Future work will explore DNABERT and further integration of biological metadata to enhance model interpretability and prediction accuracy.

ABSTRACT. Substance use disorders (SUDs) and inflammatory bowel disease (IBD) are increasingly linked to gut microbiota alterations, which influence metabolism, immune function, and neurological health. This study employs advanced computational bioinformatics techniques, including statistical validation, clustering algorithms, and dimensionality reduction methods, to analyze gut microbiota composition in individuals with alcohol use disorder (AUD), opioid use disorder (OUD), cannabis use disorder (CUD), and IBD. Using 16S rRNA sequencing open-source raw data, we assessed microbial shifts across groups. Firmicutes and Bacteroidota emerged as the dominant phyla, with significant variations: the CUD group showed a higher Bacteroidota proportion, while IBD displayed increased Firmicutes representation. The AUD group formed a distinct microbiota cluster, suggesting unique alterations. Principal Component Analysis (PCA) and hierarchical clustering revealed overlapping microbial profiles among SUD and IBD groups, supporting the hypothesis of shared dysbiosis patterns. Effect size calculations and machine learning-based classification models identified key phyla contributing to microbial imbalances. The computational integration of microbiome data enabled the identification of distinct clustering patterns and demonstrated the potential for predictive modeling in microbiota-based diagnostics. These findings underscore the power of bioinformatics in deciphering gut microbiota complexities in SUDs and IBD, offering insights for precision medicine applications and microbiome-targeted therapeutic interventions.

ABSTRACT. Telerehabilitation is a promising solution to ensure continuity of care for people with disabilities by delivering rehabilitation services directly in their homes or other familiar environments. This paper introduces a method for assessing movement execution during telerehabilitation sessions, offering effective support to clinicians during remote monitoring. This approach uses the 3D coordinates of the body joints provided by Microsoft Kinect to evaluate movement based on joint positions and angles. The proposed method has been validated using the IntelliRehabDS dataset, which includes rehabilitation exercises designed for different pathologies. Preliminary results demonstrate the system’s ability to detect deviations and assist in performance evaluation. Future developments aim to integrate this approach, supported by AI-driven evaluations, within a Cloud/Edge architecture to provide telerehabilitation services. In this way, patients can receive real-time feedback on the Edge, and clinicians can receive a detailed assessment on the Cloud, leading to improved rehabilitation outcomes and greater patient involvement.

ABSTRACT. This study examines the ecological validity of a Living Lab (LL) experimentation protocol by comparing it with in-the-wild data collection from older adults, focusing on walking speed as a measure of indoor ambulation. Data were collected in two settings: a Living Lab testbed resembling a home space and real world environment (homes). Three older adult cases were included in the current study. A case-series exploratory approach was used to analyze walking speed patterns across the two environments – the living lab and the real-life one. Statistical features were calculated, and statistical tests were applied to assess differences and agreement between the experimental conditions. The results revealed higher variability and lower values in walking speed in the home environment compared to the Living Lab, suggesting that the suggested experimental setting still lacks ecological validity. Given the small sample size, results are considered as hypothesis generation rather than definitive conclusions.

ABSTRACT. This paper introduces a robotic rehabilitation system for upper limb therapy, utilizing a manipulator controlled through variable admittance control. The proposed approach allows dynamic modulation of the patient’s mobility within a predefined spatial tolerance region, which is set by the therapist at the beginning of each session based on the patient’s capabilities. The system adjusts movement constraints based on the applied force, promoting guidance along the intended path. Specifically, it supports motion along the path tangent while increasing resistance to deviations, but only when the applied force pushes the patient away from the trajectory. Conversely, when the movement is directed toward realignment, the system permits greater freedom in all directions that facilitate a return to the predefined path. To validate the proposed approach, an experimental study was conducted to analyze the system’s behavior in practice. The results demonstrate the system’s effectiveness in modulating movement assistance and resistance based on patient interaction, ensuring both guidance and adaptability within the rehabilitation process.

ABSTRACT. Autism spectrum disorder (ASD) affects many children and limits their social interaction, communication, and behavioral skills. Regular follow-ups with qualified professionals assist in the patients’ development through sessions and progress evaluations. This progress is manually recorded by professionals, which can lead to errors in analysis. To assist with these annotations, an automation process is proposed using pose estimation and object detection techniques in computer vision to detect interaction between participants in therapy sessions with children with ASD. For this purpose, the Yolov8 object detector and Yolov8-Pose estimator are applied. Subsequently, interaction detection is performed using the results obtained from the previous predictions. Heuristics were introduced to compare these techniques. To enhance performance, solutions were explored to address the limitations encountered in object detection and pose estimation predictions. The results demonstrated a 15.8% improvement in precision for the pose estimation heuristic compared to the bounding box heuristic, achieving satisfactory performance for the proposed approach.

ABSTRACT. When artificial intelligence (AI) and machine learning (ML) models are applied in healthcare, the ability to understand and explain model decisions is an important aspect. Methods in the field of explainable AI (XAI) have been developed to create explanations for such decisions, which provides transparency and trust to the prediction model. However, the use of model explanations with the purpose of improving prediction performance remains unexplored. Our proposed Explanation Supported Learning (XSL) framework can improve classification performance for ML models used in medical imaging systems, while also providing a new understanding of how medical images are processed by deep learning (DL) models. The XSL framework consists of novel methods to achieve knowledge transfer from one or several teacher models to a student model. The novelty lies in using explanations from the teacher models, obtained from XAI techniques, as added features when training the student model. This approach enables flexible knowledge transfer between models of different architecture types. We further demonstrate how the XSL framework can be used as a new metric for measuring the quality of the explanations provided by XAI methods. The achievement of increased performance in this framework requires that the chosen XAI technique contains useful information based on the learned understanding of the input data by the teacher models. By testing XSL on the HyperKvasir gastrointestinal image dataset, we achieved significant increases in most of the measured classification metrics, and exceeded most benchmark scores of the HyperKvasir paper. A link to our code repository will be provided upon acceptance.

ABSTRACT. Tuberculosis is one of the deadliest diseases in the world, despite being treatable, the number of new cases increases yearly. The situation is even more concerning due to antimicrobial resistance (AMR). To avoid undetected cases and breaking transmission, the research community is using artificial intelligence (AI) to develop new and fast diagnosis tools. Nevertheless, the creation of good quality and well-balanced datasets for training AI tools is challenging. It is especially difficult to collect data on asymptomatic patients who do not go to the professional as they do not feel the symptoms, such as the one who suffered latent tuberculosis. In this work, we propose an explainable deep learning ensemble of convolutional neural networks (CNNs) to classify tuberculosis chest X-ray (CXRs) images. The ensemble was designed to alleviate being influenced by unbalanced data due to the small number of latent patients included in the dataset. The system was trained to detect tuberculosis against healthy patients and other diseases with CXRs as the only input, so as to inform tuberculosis patients of the stage of the disease (active or latent). The model was trained by applying two parallel CNNs in an ensemble that used random forest (RF) to overcome the imbalance. The system reported a performance of 96% accuracy, showing that the ensemble can improve the performance of the underrepresented class. Further, the RF completed the interpretability of the system provided by grad-CAM heatmaps, supporting the evolution and integration of the system in clinical environments.

ABSTRACT. Residual inhibition (RI) is a phenomenon observed in many tinnitus patients, where tinnitus remains temporarily suppressed for a short duration—typically less than a minute—after the cessation of an appropriate masking stimulus. Despite decades of clinical interest in RI, machine learning (ML)-based, feature-driven classification approaches remain scarce.

In this study, we investigate the potential of ML models to classify RI by developing a dedicated data analysis pipeline. Given the heterogeneous nature of the features—including numerical, binary, and ordinal variables—we apply feature importance techniques tailored for mixed-type data to ensure a comprehensive evaluation and improve interpretability. Our results demonstrate a clear separation between RI classes, highlighting the relevance of specific clinical and audiological factors in distinguishing them.

Building on this, we assess the predictive power of RI classifications with high confidence within a supervised learning framework to determine their relevance for treatment outcome prediction. While our findings confirm that RI can be effectively classified, they also suggest that RI alone is not sufficient to serve as a reliable predictor for treatment outcomes.

ABSTRACT. Emergency departments (EDs) face increasing demands that require efficient methods to quickly identify patients at risk for critical outcomes (i.e., inpatient death or admission to the ICU within 12 hours). This study aimed to develop interpretable rule-based models for predicting critical patient outcomes using machine learning. We used data from the MIMIC-IV-ED database, applying Gradient Boosting (GB) and Logistic Regression (LR) models to identify patients at risk based on 13 readily available initial triage variables. Gradient boosting achieved better performance for the test set (Accuracy: 78.21%, AUROC: 0.887, AUPRC: 0.445) compared to Logistic Regression (Accuracy: 77.27%, AUROC: 0.863, AUPRC: 0.370). Using the Te2Rules method, we extracted 43 clinically interpretable rules from the GB model, achieving high overall fidelity (98.90%). Furthermore, categorizing extracted rules using Rule Coverage Index (RCI) into ”high,” ”medium,” ”low” improved their clinical applicability. Our method aims to provide a practical balance between predictive accuracy and interpretability, potentially assisting clinicians in promptly identifying critically ill patients during the early stages of assessment.

ABSTRACT. Cancer is a major health concern in Brazil and globally and is characterized by its high incidence and mortality rates. Diagnosis typically involves the preparation and microscopic analysis of tissue samples, which are often stained with hematoxylin and eosin (H&E). However, color variation in these images poses a significant challenge for computer-aided diagnosis systems. This study explored dictionary learning techniques for H\&E stain color normalization by utilizing public histological image datasets with varying colors for performance comparisons. The findings revealed that the non-negative matrix factorization techniques outperformed existing methods in the literature, particularly in feature preservation, achieving maximum FSIM, PSNR, QSSIM, and SSIM values of approximately 0.82, 40.21, 0.84, and 0.93, respectively. Furthermore, the impact of normalization on nuclear segmentation highlighted that the visual quality of the normalized images did not directly correlate with the quantitative segmentation results. Therefore, this study raises important open questions for the development of future research in this area.

ABSTRACT. Prostate cancer, a growing global health concern, necessitates precise diagnostic tools, with Magnetic Resonance Imaging (MRI) offering high-resolution soft tissue imaging that significantly enhances diagnostic accuracy. Recent advancements in explainable AI and representation learning have significantly improved prostate cancer diagnosis by enabling automated and precise lesion classification. However, existing explainable AI methods, particularly those based on frameworks like GAN, are predominantly developed for natural image generation, and their application to medical imaging often leads to suboptimal performance due to the unique characteristics and complexity of medical image. To address these challenges, our paper introduces three key contributions. First, we propose ProjectedEx, a generative framework that provides interpretable, multi-attribute explanations, effectively linking medical image features to classifier decisions. Second, we enhance the encoder module by incorporating feature pyramids, which enables multi-scale feedback to refine the latent space and improves the quality of generated explanations. Additionally, we conduct comprehensive experiments on both the generator and classifier, demonstrating the clinical relevance and effectiveness of ProjectedEx in enhancing interpretability and supporting the adoption of AI in medical settings.

ABSTRACT. The respiratory system can be significantly affected by thoracic injuries, which can lead to complications such as lung dysfunction. Therefore, immediate diagnosis, along with the precise location of these injuries is crucial, as it allows targeted medical interventions, reduces unnecessary treatments, and accelerates patient recovery. Respiratory function is dependent on the diaphragm and intercostal muscles, which work in sync to produce individual breathing motion. These motions are highly individual and can be influenced by injuries and respiratory therapy. In this paper, we employ an embedded sensor network to record human breathing patterns and present a novel approach utilizing Temporal Graph Neural Networks (TGCNs) to develop radiation-free, non-invasive techniques for breathing motion monitoring. We modeled the sensor network into a graph structure in which the nodes represent the sensors and the edges represent the correlation among them, enabling our method to effectively capture both spatial and temporal relationships between sensor measurements. We simulate a network of standard sensors to monitor human breathing movements, generating a synthetic dataset of different breathing motions using real data for oversampling, including the degree of abnormalities such as mild, moderate, and severe. We perform node-level classification using the Graph Convolutional Gated Recurrent Unit (GConvGRU) model to identify the abnormal breathing pattern along with the location and severity level of the injuries. Our results demonstrate that TGCNs can accurately localize breathing abnormalities through a graphical sensor network representation, facilitating the location of the potential severity of the injury, and improving remote diagnosis, particularly in post-injury rehabilitation.

ABSTRACT. Mammographic breast density classification is essential for cancer risk assessment but remains challenging due to subjective interpretation and inter-observer variability. This study compares multimodal and CNN-based methods for automated classification using the BI-RADS system, evaluating BioMedCLIP and ConvNeXt across three learning scenarios: zero-shot classification, linear probing with textual descriptions, and fine-tuning with numerical labels. Results show that zero-shot classification achieved modest performance, while the fine-tuned ConvNeXt model outperformed the BioMedCLIP linear probe. Although linear probing demonstrated potential with pretrained embeddings, it was less effective than full fine-tuning. These findings suggest that despite the promise of multimodal learning, CNN-based models with end-to-end fine-tuning provide stronger performance for specialized medical imaging. The study underscores the need for more detailed textual representations and domain-specific adaptations in future radiology applications.

ABSTRACT. Interoperability in healthcare data remains a challenge, especially when exploring complex relationships within data modeled as Fast Healthcare Interoperability Resources (FHIR). While FHIR standardizes data exchange, traditional visualization methods often fail to capture its semantic richness. This paper introduces \emph{FHIR Lens}, an interactive graph-based platform that enhances FHIR exploration through RDF. It leverages SPARQL for dynamic query execution and employs recursive graph expansion to traverse structured healthcare data, addressing challenges like blank nodes and inconsistent references. By visualizing FHIR resources as interconnected nodes and edges, FHIR Lens supports advanced data analysis, relationship discovery, and decision support, contributing to semantic interoperability in healthcare by bridging data standards with graph-based modeling. We describe the platform's functionality and underlying algorithms and evaluate its performance on increasingly large FHIR datasets.

ABSTRACT. Biomedical ontologies are crucial for organizing domain-specific knowledge, yet traditional alignment methods relying on lexical matching often fail to capture complex semantic relationships. To address this limitation, we propose a novel approach leveraging siamese neural networks and transformer-based models to enhance ontology alignment within the biomedical domain. Our method applies self-supervised contrastive learning to biomedical literature, optimizing the prediction of semantic similarities between concepts in the UMLS Metathesaurus. The results demonstrate that this approach surpasses lexical-based techniques by identifying contextual relationships and uncovering new interconnections among UMLS terminologies. This highlights the potential of our models in improving ontology alignment and enriching biomedical knowledge integration.

ABSTRACT. Event-Based Surveillance Systems (EBS) are crucial for detecting emerging public health threats. However, these systems face significant challenges, including overreliance on manual expert intervention, limited handling of heterogeneous textual data, etc. The Description-Detection Framework (DDF) addresses some of these limitations by leveraging PropaPhen (Core Propagation Phenomenon Ontology), UMLS, and OpenStreetMaps to detect suspicious health-related cases using spatiotemporal and textual data. However, DDF is restricted to detection and lacks the ability to classify the detected observations into meaningful categories.

To adress this limitation, we propose to enhance DDF by incorporating a clustering-based classification process. This enhancement employs BioSTransformers, a pretrained biomedical language model built on Sentence Transformers trained on PubMed data, to compute semantic similarity between observations. By capturing domain-specific semantic relationships, BioSTransformers enables clustering that integrates biological semantics with spatiotemporal context, outperforming traditional methods from the literature in observation classification. Our proposed approach reduces the dependency on manual expert effort, improves the system's ability to process heterogeneous data, and enhances the accuracy and contextual relevance of case classification. The results demonstrate the potential of this method to advance EBS systems, providing a scalable and automated solution to public health surveillance challenges.

ABSTRACT. Lung cancer is one of the deadliest types of cancer and poses a significant public health challenge. Despite numerous studies identifying various risk factors associated with this disease, further research remains essential, particularly in the biological domain. Currently, multiple data sources compile biological information on various diseases, including lung cancer and its subtypes. However, these sources often differ in structure and format, making data extraction and efficient use in artificial intelligence (AI) models more challenging. Ontologies, semantic technologies, and data reuse strategies play a crucial role in addressing this issue. By leveraging these approaches, it is possible to build a knowledge graph that integrates heterogeneous data sources into a unified format, facilitating interoperability and data extraction. Lung-CABO is an ontology specifically designed for lung cancer and its subtypes, whose effectiveness has been evaluated. Through this ontology, a knowledge graph has been developed to explore, extract, and utilize information both to identify risk factors and as input for AI models. Additionally, Lung-CABO is reusable and can be expanded by incorporating association classes that integrate other relevant data related to the disease, such as environmental factors, further enhancing its scope and applicability.

ABSTRACT. The integration of heterogeneous clinical datasets represents a fundamental challenge in contemporary biomedical research, particularly when reconciling multi-language and multi-institution data sources. The challenge of this procedure lies in the effort required to map the original concepts with their standard definitions. Various automated mapping solutions can assist researchers in this process, but the complexity grows when handling multi-language datasets, resulting in substantial manual work for translation and mapping. In this paper, we proposed a novel framework for clinical concept harmonisation that leverages vector-based embeddings and semantic search methodologies to enhance interoperability in multi-cohort studies. The methodology incorporates comprehensive data profiling, ontology-driven concept alignment, and machine learning-based vector search within a unified architecture. We demonstrate the efficacy of this approach through practical application to Alzheimer's disease~(AD) research datasets from distinct institutions with different languages, achieving effective cross-lingual concept mapping while maintaining compatibility with established standardisation frameworks.

ABSTRACT. Hallucination in deep learning (DL) classification, where DL models yield confidently erroneous predictions remains a pressing concern. This study investigates whether binary classifiers are truly learning disease-specific features when distinguishing overlapping radiological presentations among pneumonia subtypes on chest X-ray (CXR) images. Specifically, we evaluate if uncertainty measure is a valuable tool in classifying signs of different pathogen-specific subtypes of pneumonia. We evaluated two binary classifiers to classify bacterial pneumonia and viral pneumonia, respectively, from normal CXRs. A third classifier explored the ability to distinguish bacterial from viral pneumonia presentation to highlight our concern regarding the observed hallucinations in the former cases. Our comprehensive analysis computes the Matthews Correlation Coefficient and prediction entropy metrics on a pediatric CXR dataset and reveals that the normal/bacterial and normal/viral classifiers consistently and confidently misclassify the unseen pneumonia subtype to their respective disease class. These findings expose a critical limitation concerning the tendency of binary classifiers to hallucinate by relying on general pneumonia indicators rather than pathogen-specific patterns, thereby challenging their utility in clinical workflows.

ABSTRACT. Chest X-rays are a crucial diagnostic tool, but their interpretation can be time-consuming. Deep learning offers a promising solution, but requires large, well-annotated datasets, which are often limited by privacy issues and data access restrictions. Federated learning (FL) allows hospitals to collaboratively train deep learning models without data sharing, addressing these limitations. In this paper, we propose an FL model based on a lightweight ConvNeXt architecture and Federated Averaging to analyze both performance and interpretability as the number of clients increases, while maintaining the same total amount of data for chest X-ray imaging. The results demonstrate superior performance of the FL model compared to the local one, and the interpretability analysis shows that FL-trained models produce better alignment with ground truth compared to local models. However, both performance and interpretability decline as the number of clients increases. These findings highlight the need for strategies to mitigate performance degradation in high-client scenarios to improve the clinical applicability of FL-based systems.

ABSTRACT. Accurately localizing pulmonary nodules relative to other anatomical structures is crucial for disease management, guiding biopsies, and formulating effective treatment strategies. This study introduces a fully automated approach for classifying nodules detected in computed tomography (CT) images as pleural (near the pleura as (N_p)) or non-pleural (distant from the pleura as (D_p)). We propose a combination of Principal Component Analysis (PCA) and deep learning approach to determine a threshold for the optimal correlation between the original image and the latent PCA-based image reconstruction used to classify lung nodule as (N_p) or (D_p). Applying our methodology to the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset, we found that the PCA-based approach demonstrated substantial agreement with human evaluations of the proximity of the lung nodule to the pleura, achieving a Cohen's Kappa value of 0.76, outperforming an intensity-based baseline model with a Cohen’s Kappa value of 0.55. Further analysis revealed significant variability in radiologists' semantic ratings for (N_p) versus (D_p), with the highest variability observed in the texture feature. These findings demonstrate that our approach has the potential to enhance the classification of lung nodule localization when integrated into computer-aided diagnosis (CAD) systems.

ABSTRACT. Emphysema is a hallmark of Chronic Obstructive Pulmonary Disease and an independent risk factor for lung cancer. Computed Tomography (CT) is the main diagnostic platform for identifying emphysema. In clinical practice, the quantitative assessment identifies emphysema as low attenuation areas under a specific cut-off threshold set to -950 Hounsfield Unit. Despite its wide adoption, this method lacks consensus on an optimal cut-off threshold and is prone to measurement variation, asking for new solutions that encompass this limitation. We propose a hybrid deep learning approach for emphysema classification that combines convolutional neural networks for local feature extraction with MAMBA’s capability to model long-range dependencies. This fusion ensures a complementary feature representation, capturing both fine-grained and global contextual information. Furthermore, we demonstrate the effectiveness of self-supervised pretraining in domain-specific data, refining the weight configuration of the model to better align with the target distribution and improve its performance during supervised training. The results show on the SCAPIS public cohort that our hybrid model not only outperforms the traditional LAV950 method for emphysema quantification but also surpasses two well-established deep learning architectures.

ABSTRACT. Diabetic retinopathy (DR) is a severe illness which can be detected from retinal images. In this work we address the problem of its machine learning-based prediction from clinical records and retinal images that suffer from an underrepresentation of the DR cases in the clinical dataset. Several classification models are compared in the task of predicting DR in diabetes mellitus type 1 patients, applying different class imbalance correction techniques. The results show the ability of class weighting and a combined undersampling and oversampling approach to significantly improve the recognition of DR cases (the class of interest in the medical application) classification.

ABSTRACT. Determining the type of kidney stone is crucial to prescribe appropriate treatments and prevent recurrence. Currently, there are different approaches to identify the type of kidney stone; however, obtaining results through the reference ex vivo identification procedure can take several weeks, while in vivo visual recognition requires highly trained specialists. For this reason, machine learning models have been developed to provide urologists with an automated classification of kidney stones during ureteroscopy. Nevertheless, a common issue with these models is the lack of training data.

This methodology presents a deep learning method based on few-shot learning, aimed at producing sufficiently discriminative features for identifying kidney stone types in endoscopic images, even with a very limited number of samples. This approach was specifically designed for scenarios where endoscopic images are scarce or uncommon classes are present, enabling classification even with limited information. Additionally, the model was enhanced through a transfer learning approach and the use of few-shot learning-based methods. The results demonstrate that Prototypical Networks, using up to 25% of the training data, can achieve performance that is equal to or better than traditional deep learning models trained with 100% of the data.

ABSTRACT. Deep learning models enhance breast cancer detection in mammograms but struggle with domain shifts, where test data differ from training data. Domain adaptation (DA) helps address this issue but often relies on unstable adversarial techniques. Breast lesion classification in mammographic images also faces challenges like data scarcity and overfitting. Mixup mitigates these by generating synthetic samples, increasing variability, and improving robustness. Meanwhile, contrastive learning enhances feature alignment, boosting generalization and classification accuracy across domains. This paper proposes a DA model that integrates mixup and contrastive learning to improve feature alignment and generalization, leading to more accurate breast lesion classification. Our approach outperforms standard DA methods, achieving 82.5% accuracy, 0.774 F1 score, and 0.7868 AUC on INbreast (target dataset), surpassing DANN (63.6%) and Deep CORAL (67.7%). It also generalizes well, reaching 70.4% accuracy on CMMD and 63.64% on CDD-CESM, demonstrating its effectiveness in addressing domain shifts.

ABSTRACT. Hospital readmission is a complex health outcome, which burdens the patient, their family network, and the health system. The black box machine-learning techniques present challenges in validation, regulation, and understanding of health outcomes, such as hospital readmissions. Complex models based on deep learning generally achieve greater accuracy, although they generate tension between accuracy and interpretability. XGBoost outperforms traditional statistical methods and can potentially improve the prediction of negative health outcomes. In a tertiary university hospital, we carried out a retrospective cohort study with patients under 18. Demographic, clinical, and nutritional data were extracted from electronic the hospital system. We used extreme gradient boosting (XGBoost) to build a predictive model for potentially avoidable 30-day readmissions. We use methods to calculate the importance scores of each variable for the generated model. Our study showed that it is possible to develop an interpretable prediction model for potentially preventable pediatric readmissions using the XGBoost algorithm.

ABSTRACT. This paper presents a novel approach to non-invasively assess different levels of perfusion using remote photoplethysmography (rPPG) from an RGB camera. In a proof-of-concept study involving 20 healthy participants, three degrees of stenosis were induced in the upper extremity by applying external pressure. The effects on the extracted rPPG signal from the affected extremity were analyzed by correlating it with a reference rPPG signal from a well-perfused area. The results demonstrate that three distinct levels of perfusion can be distinguished. The results may contribute to the development of a non-invasive, cost-effective method for detecting perfusion disorders, such as peripheral artery disease (PAD).

ABSTRACT. Circadian rhythm is essential for regulating physiological functions, and is compromised in people with dementia. Objective measurements are needed to track circadian trend and stability, usually from actigraphy data. This paper explores the feasibility of applying the stability model on activity counts obtained from raw accelerometer, and discusses preliminary results limited to two participants included in a randomized controlled trial that investigates the effect of virtual darkness as a treatment.

ABSTRACT. Polymyalgia rheumatica (PMR) is an inflammatory rheumatic disease primarily seen in older patients, most often diagnosed in primary care, and treated with oral corticosteroids. It can be challenging to distinguish PMR from conditions with similar presenting symptoms, with an incorrect diagnosis of PMR potentially leading to delayed diagnosis for the true condition, and unnecessary steroid treatment. A digital pain manikin is a measurement tool for collecting self-report pain data, presenting an overview of the human body where people can mark the location of their pain to produce a pain drawing. Advanced summary and analysis methods of PMR pain manikin data may be able to improve the accuracy of PMR diagnosis in primary care. Participants (n=195) were referred to the study by GPs who suspected PMR, with a final diagnosis made by rheumatologist specialising in PMR. PMR diagnosis was confirmed for approximately 30% of participants. Of the participant data already analysed (n=91), participants were majority white (n=87), female (n=60), with an average age of 72. I will develop clinically relevant summary measures for PMR data, and build a machine learning classification model to distinguish PMR from other painful conditions with similar presenting symptoms.

ABSTRACT. Fragmented sleep is a characteristic of dementia. For people with dementia at the end of life, sleep has not been investigated using digital measures. In this preliminary analysis we compare data from a smartwatch (heart rate variability) and an environmental sensor (sleep stages). Results show that this data has potential to provide clinically meaningful observations. Further research will include more participants and biosignals.

ABSTRACT. Diabetes Mellitus has emerged as a major global health challenge, impacting over 10% of adults worldwide. This condition is marked by a complex pathophysiology that leads to notable morbidity and socioeconomic consequences, requiring precise predictive models and efficient management tactics. In this paper, we investigate the advantages of employing causal modeling in both predicting and managing diabetes. Utilizing the PIMA diabetes dataset, we construct a causal model to assess the effects of potential interventions guided by diabetes management protocols, conducting a what-if analysis to pinpoint potential preventive measures for diabetes onset. The intervention outcomes show that the causal graph used precisely mirrors real-world results, notably reflecting a decrease in glucose levels when a healthy lifestyle is maintained. Moreover, the what-if analysis results are promising, suggesting that altering certain conditions might have prevented the development of diabetes in some cases. These findings underline the value of causal inference methods in identifying risk factors for diabetes and steering effective prevention strategies. Expanding current data limitations with more extensive, varied datasets and expert insights will boost predictive precision and clinical relevance.

ABSTRACT. Clinical data generated in healthcare systems is increasingly recognized as a key resource for biomedical research, healthcare optimization, and population health monitoring. However, its full potential remains underexploited due to fragmentation, heterogeneity, and lack of interoperability between data sources. The ELADAIS project addresses this challenge by designing, developing, and deploying a scalable, modular, and interoperable technological platform for the extraction, transformation, storage, and advanced analysis of high-impact clinical data. Grounded in the OMOP Common Data Model (OMOP-CDM), ELADAIS integrates a microservice-based architecture, analytical environments, workflow orchestration, and federated data capabilities. The platform will be deployed at two major hospitals in Madrid, Spain, with the expectation of standardizing access to over 1 million patient records and more than 19 million clinical events. This paper presents the architectural principles and expectations of ELADAIS, highlighting its potential to accelerate reproducible and collaborative clinical research.

ABSTRACT. Anxiety and depression are psychological disorders characterized by persistent and impairing symptoms. They affect milions of people worldwide and have a significant impact on individuals' well-being and daily functioning. Although highly effective treatments exist, delayed diagnoses and limited access to mental health care contribute to a significant number of undiagnosed individuals. Therefore, it is important to explore predictive modeling to anticipate and address potential issues before the symptoms increase. In that context, this study proposes a machine learning approach to predict anxiety and depression scores based on the Generalized Anxiety Disorder (GAD-7) and Patient Health Questionnaire (PHQ-9). In a regression scenario the proposed multi-layer perceptron (MLP) achieved the lowest MAE values of 5.3924 for anxiety and 5.06 for depression, as well as the lowest MAPE values of 0.1101 for anxiety and 0.1043 for depression. For a classification scenario the best-performing models were the random forest (RF) and LightGBM with an F1-score of 0.8997 and 0.8918 for anxiety, respectively, and 0.7593 and 0.7480 for depression. These results highlights the potential of neural network-based models to outperform traditional ensemble and kernel-based approaches to predict mental disorder scores. Additionally, the classification results also suggests that tree and kernel-based models can effectively maintaining balanced predictive performance.

ABSTRACT. Effective management of pediatric diabetes remains a clinical challenge, particularly due to the onset of lipodystrophy resulting from repeated insulin injections and inadequate rotation of injection sites. These complications adversely affect subcuta- neous tissue integrity and insulin pharmacokinetics, ultimately compromising glycemic control. In this study, we introduce di- Pen, a novel smart case designed to integrate with commercial insulin pens and equipped with a dual-sensor system: an optical module for non-invasive lipodystrophy detection and an inertial measurement unit for monitoring injection-site rotation. To evaluate the feasibility of the proposed approach, a clinical study was conducted involving a pediatric cohort. The system employs a personalized machine learning pipeline, leveraging a leave-one-acquisition-out validation strategy to replicate real- world deployment scenarios, wherein newly acquired data from the same subject are evaluated without prior exposure during training. The system demonstrated promising performance in both detection and classification tasks, suggesting that di-Pen may represent a viable tool for enhancing insulin therapy through personalized, data-driven injection guidance and tissue health monitoring.

ABSTRACT. The Advancing Neurofeedback in Tinnitus (ANT) project aims to develop improved neurofeedback protocols and BCI technology by systematically designing engaging feedback stimuli and optimizing neural targets. General design principles for audiovisual feedback stimuli are established, system and software engineering for general purpose real-time M/EEG is developed, and ultimately integrated for the clinical use case tinnitus. This interdisciplinary effort combines expertise in clin- ical neuroscience, design, user experience research, psychology, and biomedical signal processing to create a novel neurofeed- back approach with potential for both clinical and home-based applications.

ABSTRACT. This work addresses the gaps in research and management of rare diseases in the Community of Portuguese-Speaking Countries. There is a lack of studies assessing rare disease-related digital and informational readiness, especially in African countries. International cooperation can lead to significant technological advances by measuring community members' digital maturity in managing health data, information systems, and medical technology. The main objectives include creating a collaborative network for rare diseases in the community, improving the registration and monitoring of patients, and promoting continuing education actions. From a scientific, technological, and innovation point of view, the project seeks to map the scenario using digital health tools, create strategies to establish and strengthen cooperation networks, define formative second opinion processes, and promote equitable access to digital health technologies to support decision-making. These actions will help to identify and fill gaps in governance processes, documentation, and practices, leading to realistic and equitable recommendations for better disease management in challenging contexts. Therefore, the cross-border collaboration is expected to promote a unified approach through the transnational sharing of scientific and clinical evidence, expanding the cooperation between team members and partner institutions.

ABSTRACT. Given a set of historical, unlabeled patient records, how can we model the relationships between various concepts related to health conditions and treatments? Modeling health data as knowledge graphs can aid in better understanding and mining recurrent and abnormal patterns within massive amounts of records. However, for generic data modeling, the concepts and relationships of EHRs are often defined manually, which can be laborious and prone to human errors. The lack of structure in textual reports makes modeling challenging, as there is typically no standard in information, terminology, or other elements. This work proposes CliniKG for utilizing Large Language Models (LLMs) to structure patient records, enabling the modeling of health data concepts. First, we employ LLMs with zero-, or few-shot learning to define the graph's relationships of pairwise concepts. Based on the resulting modeling, we extract meaningful measures from nodes and define their semantics automatically. The proposed visualizations highlight key data findings, such as recurrent or rare relationships. The experimental evaluation shows CliniKG in action using a real dataset from a public hospital in Brazil. We performed a qualitative analysis with 40 domain experts to evaluate medium-size LLMs and prompt configurations. The study reveals interesting patterns automatically identified within the data. CliniKG exhibits linear performance relative to the number of nodes, and the visual tools can assist specialists in monitoring patients' conditions.

ABSTRACT. Drug-drug interactions are a major cause of mortality during hospitalization, causing toxicities and unexpected side effects. This work utilizes a knowledge graph derived from biomedical literature and open databases, to predict different classes of drug drug interactions. To this end, a path analysisbased machine learning approach is compared with various graph embedding techniques in a lung cancer use case. The experiments aim at analysing different type of interactions from a public database, to define five general classes. Focusing on under-represented classes of interactions, in order to boost their predictive performance via different configurations, the path analysis-based approach achieves better performance results, allowing for their subsequent interpretation using the most important features of each path.

ABSTRACT. Deep learning models have achieved great success in medical imaging tasks. However, recent work on fairness in healthcare has shown these models can be biased, potentially leading to discriminatory treatment of patients based on demographic attributes such as race, gender, and age. Data bias, often resulting from imbalanced and non-representative datasets, can negatively impact model fairness. While aggregating data from multiple sources can help mitigate data bias, privacy concerns make the data-sharing process challenging. In this scenario, Federated Learning (FL) has emerged as a solution for the collaborative training of models without data sharing. This paper presents FairMed-FL, a methodology to assess fairness in FL for medical imaging tasks. By utilizing two public chest X-ray datasets partitioned by sex and age, we compared federated models trained with clients from single and multiple datasets against centralized models trained on each client’s data. The results indicate that FL reduces performance discrepancies between demographic groups, enhances the performance of the worst-performing groups, and improves overall metrics compared to centralized approaches. These findings highlight its potential for promoting fairness in medical imaging.

ABSTRACT. Privacy and regulatory barriers often hinder centralized machine learning solutions, particularly in sectors like healthcare where data cannot be freely shared. Federated learning has emerged as a powerful paradigm to address these concerns; however, existing frameworks primarily support gradient-based models, leaving a gap for more interpretable, tree-based approaches. This paper introduces a federated learning framework for Random Forest classifiers that preserves data privacy and provides robust performance in distributed settings. By leveraging PySyft for secure, privacy-aware computation, our method enables multiple institutions to collaboratively train Random Forest models on locally stored data without exposing sensitive information. The framework supports weighted model averaging to account for varying data distributions, incremental learning to progressively refine models, and local evaluation to assess performance across heterogeneous datasets. Experiments on two real-world healthcare benchmarks demonstrate that the federated approach maintains competitive predictive accuracy—within a maximum 9\% margin of centralized methods—while satisfying stringent privacy requirements. These findings underscore the viability of tree-based federated learning for scenarios where data cannot be centralized due to regulatory, competitive, or technical constraints. The proposed solution addresses a notable gap in existing federated learning libraries, offering an adaptable tool for secure distributed machine learning tasks that demand both transparency and reliable performance. Tool available at https://github.com/ieeta-pt/fed_rf.

ABSTRACT. Alzheimer’s Disease (AD) and Frontotemporal Dementia (FTD) are two of the most common neurodegenerative disorders, most commonly leading to progressive cognitive decline. Electroencephalography (EEG) has been proven as a promising tool for early and non-invasive detection of these disorders, leveraging advanced computational methods for classification. In this study, we propose a novel Convolutional Neural Network (CNN) architecture for the automatic classification of AD and FTD using EEG-based biomarkers. Our approach utilizes Power Spectral Density derived Relative Band Power and Spectral Coherence Index as input features, extracted from preprocessed EEG signals of 88 participants following Independent Component Analysis and Artifact Subspace Reconstruction for noise reduction. The extracted features capture both spectral power variations and functional connectivity disruptions, key findings of neurodegenerative disorders. The proposed CNN model is trained and validated using a Leave-One-Subject-Out cross-validation approach to ensure robust generalization. Experimental results demonstrate that our model achieves high classification accuracy, outperforming state-of-the-art machine learning methods and achieving competitive performance with state-of-the-art deep learning approaches. These findings highlight the effectiveness of EEG-based deep learning models for differential dementia diagnosis, providing a potential avenue for early detection and clinical decision support.

ABSTRACT. This work introduces a novel methodology for Electroencephalography (EEG) data analysis in the context of neurological diseases, emphasizing feature extraction through covariance matrices and their integration with Case-Based Reasoning (CBR). Departing from traditional techniques such as Fast Fourier Transform (FFT) and statistical analysis, we investigate the synergy between covariance matrices and CBR, highlighting their potential to improve the efficacy of EEG data analysis over conventional methods like Random Forest (RF). Covariance matrices analyze the relationships between channels, indirectly capturing interactions between brain regions, while CBR uses similarities in these relationship patterns across cases to make decisions, both techniques focusing on understanding the data through its interrelationships. Additionally, we incorporate Tangent Space Mapping (TSM) to make the covariance matrices more suitable for traditional classifiers by projecting them into a space that preserves their geometric properties. Empirical results on public EEG datasets show that CBR, using covariance matrices with TSM, achieves the best accuracy of 0.72 for Alzheimer’s Disease (AD) and up to 0.83 for Parkinson’s Disease (PD).

ABSTRACT. We use a beta-variational autoencoder regularized by a denoising autoencoder to learn latent representations of sleep electroencephalography (EEG) time-frequency spectrograms in a self-supervised manner. Our approach enables the discovery of sleep EEG spectral features unbiased by extrinsic sleep stage labels. We provide sample results that demonstrate the effectiveness of this approach in identifying EEG features that exhibit a variety of patterns of specialization in time and frequency without supervision, including features resembling sleep spindles and theta bursts that occur in sleep stages N2 and N3.

ABSTRACT. Cardiovascular disease (CVD) is a leading cause of global mortality, accounting for an estimated 17.9 million deaths annually. CVD is broadly defined as a group of medical conditions influenced by modifiable or non-modifiable risk factors that affect the heart's ability to function properly. Machine learning (ML) has emerged as a powerful tool for analyzing complex medical data, aiding in early detection and accurate diagnosis of CVD and improving patient outcomes. Recent studies proposed various deep learning (DL) architectures for detecting CVD, yet there is a lack of robust benchmarks for comparing their performance on large-scale databases. In this work, we benchmark six state-of-the-art DL architectures for multi-label heart disease classification using 12-lead electrocardiogram (ECG) data from the large-scale publicly available Medical Information Mart for Intensive Care (MIMIC) database. Specifically, we evaluate a 1-dimensional convolutional neural network (CNN) with residual blocks (1D-CNN-ResNet); bidirectional long-short-term-memory neural network with convolutional layers (CNN-Bi-LSTM); spectrogram-based CNN (SpG-CNN); convolution-attention-transformer network (CAT-Net); hierarchical attention network (HAN), and structured state space sequence (S4) model; on a multi-label heart disease classification task with seven diagnostic targets. Model accuracy is assessed using the Hamming distance and its complexity is measured by number of model parameters. By contrasting models' accuracies versus their complexity, we establish a reliable benchmark.

ABSTRACT. It is well-known that severe pain and powerful pain medications cause short- and long-term damage to the developing nervous system of newborns. Caregivers routinely use physiological vital signs [Heart Rate (HR), Respiration Rate (RR), Oxygen Saturation (SR)] to monitor post-surgical pain in the Neonatal Intensive Care Unit (NICU). Here we present a novel approach that combines continuous, non-invasive monitoring of these vital signs and Computer Vision/Deep Learning to make automatic neonate pain detection with an accuracy of 74% AUC, 67.59% mAP. Further, we report for the first time our Early Pain Detection (EPD) approach that explores prediction of the time to onset of post-surgical pain in neonates. Our EPD can alert NICU workers to postoperative neonatal pain about 5 to 10 minutes prior to pain onset. In addition to alleviating the need for intermittent pain assessments by busy NICU nurses via long-term observation, our EPD approach creates a time window prior to pain onset for the use of less harmful pain mitigation strategies. Through effective pain mitigation prior to spinal sensitization, EPD could minimize or eliminate severe post-surgical pain and the consequential need for powerful analgesics in post-surgical neonates.

ABSTRACT. Sickle cell disease (SCD) is a severe hereditary disorder that affects multiple systems and compromises blood flow, potentially leading to serious complications such as tissue damage and vision loss. Despite treatment developments, early detection is essential, even more so in regions with limited resources. In this context, clinical manifestations observed in ocular fundus images, such as vascular tortuosity, provide the opportunity to detect SCD non-invasively by applying artificial intelligence algorithms and techniques. This study analyzes the potential of deep learning methods to detect SCD using ocular fundus images and to identify relevant retinal patterns in this disorder, which leads to improved comprehension and clinical management of the disease. To this end, we used ocular fundus images from SCD patients and healthy controls to train and evaluate several models of neural networks models, including CNN and a hybrid CNN-Transformer Vision model. In addition, activation maps were built to identify the most relevant retinal characteristics for the classification problem. ResNet-50 and EfficientNet-b0 models showed better performance in the F1 score metric, getting 88% and 83% values, respectively. The activation maps analyze highlighted vascular tortuosity as an important feature of disorder detection. Notwithstanding certain limitations, such as the size of the data set or the variability between the images, the results obtained are promising. Solving these problems could improve the effectiveness of the models for the detection and characterization of this disorder.

ABSTRACT. Large Vessel Occlusion is one of the most critical neurological emergencies in stroke care, requiring rapid and accurate diagnosis to optimize clinical outcomes. Automated detection tools have demonstrated the potential to significantly reduce treatment time, thereby improving patient prognosis. In this paper, we propose a novel two-stage deep learning approach for detecting large vessel occlusion and assessing its volume directly from computed tomography angiography. Our method consists of two sequential stages. The first stage employs a two-dimensional convolutional neural network-based detector, built upon GravityNet, specifically adapted for single lesion detection with a novel pixel-based configuration. The second stage applies a three-dimensional false-positive reduction technique to refine predictions within the brain volume. Our method achieves 80% sensitivity at two false positives per case, demonstrating its robustness and effectiveness in detecting large vessel occlusions on computed tomography angiography.

ABSTRACT. Accurate electric dosimetry is essential for predicting treatment outcomes in Irreversible Electroporation (IRE), a promising non-thermal tumor ablation technique. However, real-time 3D electric field computation remains computationally expensive, limiting its practical use in clinical settings. In this study, we propose a hybrid approach combining deep learning (DL)-based initialization with an iterative numerical solver to accelerate dose map calculation. Specifically, we employ a convolutional neural network (CNN) to predict basis functions, which are subsequently refined using a Biconjugate Gradient Stabilized (Bi-CGSTAB) solver to maintain physical accuracy. We evaluate our method on data from 10 patients undergoing IRE liver ablation under real-time clinical conditions. Our results demonstrate a 10× speedup compared to conventional solvers while maintaining comparable accuracy. This hybrid method offers a promising pathway toward fast and reliable electric dosimetry for on-line IRE procedures.