ABSTRACT. Automating steps in early skin cancer detection is critical for early-stage intervention and treatment. While numerous solutions are being developed for skin cancer detection and classification, those are usually not evaluated and validated in real clinical settings, primarily due to challenges associated with image processing, standardization, and limited annotations resulting in class imbalance. To address this, we propose a mobile app driven by Wasserstein GAN with Gradient Penalty (WGAN-GP) with an improved optimization function to augment the synthetic lesions. We enhance feature representation via Local Binary Patterns (LBP) and perform classification via convolutional neural network ensembling with confidence calibration. The proposed framework improves accuracy of 97% and robustness, and a prototype mobile app demonstrates feasibility for real-world deployment

ABSTRACT. Knee osteoarthritis (KOA) is a leading cause of disability worldwide. Radiographic assessment using the Kellgren–Lawrence (KL) grading system is widely used but primarily reflects osseous changes rather than cartilage integrity. Arthroscopy remains the gold standard for cartilage evaluation but is invasive. This study develops a deep learning model for predicting cartilage damage using AP knee radiographs. A retrospective cohort of 833 patients was analyzed. ROIs were labeled using the ICRS grading system and dichotomized into mild (I–II) and severe (III–IV). A ResNet-50 shared-backbone model achieved balanced accuracy of 0.822 and ROC-AUC of 0.893. This approach demonstrates potential as a non-invasive clinical decision support tool.

ABSTRACT. Medication errors resulting from overlooked contraindications, allergies, or interactions remain a critical concern in clinical care. We propose MediScene, a decision support prototype designed to alert prescribers of potential risks at the point of prescription. The system integrates structured drug interaction data from DrugBank with unstructured safety content extracted from pharmaceutical leaflets. Patientspecific information such as known allergies, chronic diseases, and lab results is cross-checked using semantic similarity techniques to detect risks related to four categories: drug–drug, drug–allergy, drug–disease/lab, and drug–food interactions. Embedding models are applied to represent leaflet sentences and patient conditions in a shared vector space. Preliminary testing shows that MediScene can identify risks that would typically be missed by keyword-based or rule-based systems. This paper presents our architecture, methodology, and early outcomes of this hybrid AI-powered framework for safer prescribing.

ABSTRACT. This study presents a novel real-time approach for blood pressure estimation from photoplethysmography (PPG) signals processed using Artificial Neural Networks (ANN). The framework employs the MIMIC-IV dataset for ANN training, allowing continuous, non-invasive blood pressure assessment. Signal acquisition is per-formed in real time, during which the PPG waveform is subjected to key prepro-cessing operations, including DC component removal, band-pass filtering, and normalization, in order to improve signal quality. Morphological characteristics, including cardiac period, diastolic time, and systolic upstroke time, are derived from the preprocessed PPG waveform and utilized as ANN input features. The pro-posed model demonstrates high accuracy in estimating both systolic and diastolic blood pressure and shows better performance than conventional methods. The find-ings suggest that the ANN-based approach can deliver accurate real-time blood pressure estimation and is suitable for personal health monitoring in practical healthcare settings today. Future work will focus on further improving the model and exploring its integration into mobile health platforms for wider application.

ABSTRACT. Multi-label concept recognition from radiology images is challenging due to the large label space and severe long-tailed class imbalance. This study examines the effects of some specific loss functions and their multiple configurations for multi-label concept recognition on ROCOv2, a long-tailed radiology dataset consisting of 79,789 images annotated with 1,947 UMLS (Unified Medical Language System) concepts. To this end, we conduct a three-phase empirical study. The first phase evaluates two base loss functions, Binary Cross-Entropy (BCE) and Asymmetric Loss (ASL), using the EfficientNet-B0 backbone. The second phase investigates the effect of different ASL hyperparameter settings under increased model capacity with EfficientNet-B4. Finally, the third phase analyzes hybrid loss formulations that combine BCE and ASL under the same network settings. In the experiments, promising concept recognition rates were achieved with ASL. ASL improves the overall concept recognition performance score from 0.5809 (BCE baseline) to 0.5882 on the EfficientNet-B0 backbone and to 0.5916 on the EfficientNet-B4 backbone. Although this increase in the overall score may seem modest, the local increases in specific classes are even stronger. In addition, hybrid and other variants were evaluated and it was observed that they do not provide further gains over the best ASL-only configuration.

ABSTRACT. The lifecycle management of biomedical equipment represents a strategic challenge for healthcare organizations, particularly in technologically complex environments. Replacement Priority Indices (RPIs) are commonly used to support decision-making for equipment replacement; however, traditional approaches based on static models and predefined weightings are limited in capturing non-linear relationships, complex interactions, and temporal variability. This study presents the development of a dynamic RPI based on Machine Learning techniques to enhance decision-making accuracy and reliability in clinical engineering. The analysis was conducted using real-world data from the Fondazione IRCCS Istituto Neurologico Carlo Besta (Milan), including demographic, operational, economic, and clinical-functional variables. Two datasets (2024 and 2025) were constructed to ensure a clear separation between training and validation phases. Following preprocessing and feature engineering, multiple classification algorithms were evaluated, addressing class imbalance. XGBoost achieved the best performance (Macro F1-Score = 0.83), outperforming both baseline models and a traditional RPI. Feature importance analysis highlighted the relevance of dynamic indicators such as downtime and failure rates. Prospective validation on updated data, along with comparison against the institutional Investment Plan, confirmed the model’s consistency and operational value, demonstrating its ability to anticipate technological criticalities and support investment planning. The proposed system enables a transition from static planning to dynamic and proactive management of the technological asset portfolio.

ABSTRACT. Alzheimer’s disease classification often depends on heterogeneous patient-level variables, including demographic indicators, lifestyle factors, comorbidities, cognitive scores, and functional assessments. Predictive models can perform well when direct cognitive and functional markers are included, but this does not necessarily imply early diagnosis or clinically transparent reasoning. This paper presents a revised Gemini Flash Thinking methodology for interpretable feature assessment in a synthetic Alzheimer’s disease dataset containing 2,149 patient records. The main contribution of this study is a structured Gemini Flash Thinking protocol for categorizing Alzheimer’s disease dataset variables as direct diagnostic markers, indirect risk factors, possible early indicators, or unsupported associations. The method uses a fixed prompt protocol, explicit output categories, tabular reasoning summaries, and author review to improve reproducibility. The dataset is examined under full-feature and reduced-feature conditions to distinguish current symptomatic evidence from weaker non-clinical risk information. The resulting Gemini-based reasoning framework identifies MMSE, ADL, Functional Assessment, Memory Complaints, and Behavioral Problems as direct clinical markers carrying the dominant diagnostic signal, while lifestyle and demographic features are treated as indirect and hypothesis-generating rather than sufficient for point-in-time diagnosis. The study therefore supports Gemini Flash Thinking as an interpretability and clinical-plausibility layer for feature reasoning, not as a validated early-diagnosis system. Validation on real-world longitudinal data is required before claims about earlier diagnosis can be made.

ABSTRACT. Artificial intelligence (AI) has demonstrated significant potential to improve healthcare outcomes, yet its adoption in clinical practice is threatened by concerns about transparency and trust. Explainable AI (XAI) can address these challenges, ensuring more reliable AI implementation in healthcare. This study introduces an XAI-supported decision system for knee osteoarthritis (OA) classification, using a probabilistic Support Vector Machine (SVM) to estimate the likelihood of OA. The model uses logistic regression (LogReg) and a sigmoid function to transform SVM hyperplane distances into probabilistic interpretations, thereby enhancing explainability. The dataset consists of 131 knee scans (52 OA-diagnosed, 79 healthy). From these images, densitometric and geometric features of cartilage were extracted, yielding eight key features for classification. The model was evaluated using stratified 5-fold cross-validation, achieving strong performance with an F1-weighted score above 80%. Notably, the combined SVM + LogReg system improves classification performance to 83% while providing a probabilistic score, making AI-driven diagnostics more interpretable for clinicians. This approach links AI models to their applicability, particularly in clinical diagnosis, thereby enhancing their usability in healthcare environments.

ABSTRACT. Passive infrared (PIR) motion sensors are frequently used in Ambient Assisted Living (AAL) because they are low-cost, unobtrusive, and less privacy-invasive than camera- or microphone-based monitoring. However, PIR sensors provide sparse binary activations rather than direct activity labels or continuous occupancy information. This study proposes a privacy-preserving home-monitoring pipeline that converts PIR events into room-level occupancy states using a Finite State Machine (FSM) and then applies a transformer encoder to predict multi-room occupancy at a 5-minute lead time. The median sampling interval was 0.483 min, corresponding to approximately 10 samples for the 5-minute prediction horizon. The model achieved a training binary accuracy of 89.4% and a validation binary accuracy of 87.3%, with final training and validation losses of 0.273 and 0.322, respectively. Habitual discor- dance was computed as the mean absolute deviation between predicted and observed room-occupancy vectors. The held-out 24-hour test window showed how short-lead discordance can be visualized as an interpretable candidate flag for routine deviation. These results support the feasibility of transformer-based, short-latency discrepancy monitoring in PIR-based AAL settings.

ABSTRACT. Surgical Flow Disruptions in the operating room can increase cognitive workload and the risk of technical errors. While electroencephalography (EEG) provides an objective framework for monitoring mental effort, traditional statistical methods often rely on group-level analysis. This approach potentially ignores the significant degree of inter-subject variability among individuals. This study evaluated the effectiveness of various EEG features in monitoring the impact of SFDs using a Linear Mixed-Effects Model (LMEM) framework to distinguish between group trends and individual-specific responses. EEG data were recorded from nine surgical trainees during a simulated microvascular anastomosis, during which participants were subjected to six predefined SFDs, including auditory alerts and visual distractions. EEG features were analyzed using LMEM with random intercepts and slopes, and a subjectivity index was introduced to quantify the degree of individualization across 36 extracted EEG features. An initial analysis of variance on the fixed effects revealed that, generally, theta and beta power bands were significantly impacted by disruptions across the group. However, the subjectivity index identified several features, such as Gamma Asymmetry, that exhibited high individualization despite the absence of a significant group trend. Features characterized by both significant fixed effects and high subjectivity indices, such as Mean Power Beta, were identified as the most reliable metrics for tracking cognitive load. These results demonstrate that traditional group-averaging methods can mask critical individual physiological responses. By utilizing LMEM and the associated subjectivity index, this research establishes a statistical foundation for developing personalized intraoperative monitoring systems tailored to a surgeon's unique cognitive profile.

ABSTRACT. Chronic lymphocytic leukemia (CLL) comprises two major biological subtypes defined by the somatic hypermutation status of the immunoglobulin heavy chain variable region (IGHV): mutated (M CLL) and unmutated (U CLL). Although these subtypes differ markedly in clinical behavior, B cell receptor biology, epigenetics, and clonal evolution, statistical distinctions at the level of raw DNA sequence structure remain insufficiently explored. In this study, we analyze V–D–J region DNA sequences from 1,032 CLL patients using a suite of entropy based methods to determine whether information theoretic metrics can differentiate IGHV subtypes. Classical measures—including nucleotide distributions, Kullback–Leibler divergence, and Shannon entropy—revealed no significant differences between M CLL and U CLL, indicating highly similar base level composition. A robust differential entropy estimator (EnRE) likewise showed no subtype specific divergence. In contrast, two newly developed metrics—Randomized Entropy Divergence (EnRD) and the Entropy based Homogeneity Index (EnHI)—identified statistically significant differences. M CLL sequences exhibited greater intra sequence relational complexity, while U CLL sequences demonstrated markedly higher group level homogeneity, consistent with known biased IGHV usage and stereotyped BCR subsets. These findings show that deeper structural dependencies, rather than simple nucleotide frequencies, distinguish IGHV subtypes. Entropy derived metrics therefore provide a refined quantitative framework for characterizing CLL clonal architecture and complement existing prognostic markers. Integrating IGHV status with EnRD and EnRE substantially improves survival prediction, underscoring the clinical relevance of entropy based approaches in CLL stratification.

ABSTRACT. A negative correlation between EEG functional connectivity and small-worldness in the alpha band has been previously reported. The aim of the present study was to examine whether this correlation is preserved in the resting-state following cognitive tasks, as mental fatigue has been shown to affect both functional connectivity and small-worldness. Functional connectivity was quantified using phase locking value (PLV) and weighted phase lag index (wPLI), and Pearson correlation coefficients were calculated between these measures across a range of network densities. The results demonstrated a statistically significant decrease in functional connectivity in the post-task resting-state compared to the pre-task state for both PLV and wPLI, along with a significant increase in small-worldness for wPLI. A strong negative correlation between functional connectivity and small-worldness was consistently observed across both resting-state conditions. Although the correlations were stronger in the post-task state, no statistically significant differences in correlation strength were found between the states. These findings suggest that although functional connectivity and small-worldness change following cognitive tasks associated with increased mental fatigue, their inverse relationship remains stable. This supports the view that the balance between functional connectivity and small-world organization may represent a fundamental characteristic of brain network organization that is important for maintaining brain's efficiency also during mental fatigue. However, as post-task resting-states showed stronger correlations, future research could investigate whether and how this balance is altered during active cognitive engagement.

ABSTRACT. Entropy-based measures of heart rate variability (HRV) capture nonlinear aspects of cardiac regulation that time- and frequency-domain methods may miss, but most studies assess them separately, making it difficult to determine which measures provide unique information and which are redundant. This study compares seven entropy metrics within a single framework for distinguishing Normal Sinus Rhythm (NSR) from Congestive Heart Failure (CHF): four traditional measures (Shannon Entropy, Approximate Entropy, Sample Entropy, and Spectral Entropy) and three recently proposed ones (Robust Entropy Estimator, Randomized Entropy Divergence, and Entropy Index of Group Homogeneity). RR interval sequences from 54 NSR subjects and 29 CHF subjects, obtained from PhysioNet, were segmented into 500-beat windows. Six of the seven metrics differed significantly between groups, whereas Approximate Entropy did not. Univariate models identified Randomized Entropy Divergence as the strongest individual predictor, with an AUC of 0.854, while multivariate models achieved an AUC of 0.892. Principal Component Analysis and Factor Analysis showed that the six significant metrics span three independent dimensions, which helps explain why some combinations perform better than others. Generalized Linear Models and Linear Discriminant Analysis produced highly consistent results across 57 combinations (r = 0.993). Selecting one measure from each latent factor, namely Spectral Entropy, Randomized Entropy Divergence, and Entropy Index of Group Homogeneity, provided the best balance between classification performance and coefficient stability.

ABSTRACT. Pre-stimulus brain state is a major source of variability in transcranial magnetic stimulation (TMS) responses, motivating predictive models that remain physiologically interpretable. We propose a fuzzy-logic framework for predicting binary TMS-evoked EEG response levels from pre-stimulus EEG features, offering a transparent alternative to black-box machine-learning approaches. The post-stimulus target was defined as the area under the absolute EEG waveform in the stimulated channel over 0–1000 ms, and trials were labeled as low or high using the median target value. Pre-stimulus predictors included alpha, beta, gamma, delta, theta, and Higuchi fractal dimension (HFD). Four model configurations were compared: Manual Only, Auto Only, Auto + Fallback, and Hybrid. Under a stratified trial-level 80/20 held-out split, the Manual Only model achieved the best performance (macro-F1 = 0.782, accuracy = 0.782), followed by the Hybrid model (macro-F1 = 0.770, accuracy = 0.772). The reduced four-feature automatic setting matched the six-feature setting while slightly improving coverage. These findings support fuzzy logic as a transparent framework for modeling TMS responsiveness and suggest that expert-defined rules already capture much of the relevant structure.

ABSTRACT. Electroencephalography (EEG) signals exhibit complex, nonlinear dynamics that can be characterized using nonlinear measures such as Higuchi’s Fractal Dimension (HFD), which quantifies the signal’s complexity based on its self-similarity. While HFD has been widely applied in clinical research, the influence of reference montages on its estimation remains insufficiently explored. This study investigates the effect of EEG reference montage on HFD estimates and assesses their ability to differentiate depressive and control groups. The EEG signals were originally recorded with the linked mastoids reference and subsequently re-referenced using three additional montages: vertex reference (Cz), reference electrode standardization technique (REST), and common average reference (CAR). Results indicate that the reference choice substantially influences HFD estimates and the detectability of group differences. At the global level, only the mastoid reference yielded statistically significant group separation. At the channel level, mastoid referencing produced the highest number of significant channels and the largest effect sizes, with significant channels located in central, centroparietal, and parietal regions, as well as in frontal regions. In contrast, REST and average yielded significant channels that were more spatially confined, while the Cz reference exhibited the least pronounced group differences, limited to frontal channels. These findings indicate that re-referencing is a transformation that can influence HFD estimates, potentially affecting the detectability of group differences. Future studies with larger cohorts are needed to further investigate the effects of referencing and to enable more generalizable conclusions.

ABSTRACT. Covert visuospatial attention enables the selection of a spatial location without overt eye movements and is classically associated with a lateralized modulation of posterior alpha-band activity. However, how other frequency bands contribute to the cortical organization of this process remains less clear. In this study, we compared leftward and rightward covert orienting using source-level EEG, with the aim of investigating frequency-specific modulations of both oscillatory power and functional connectivity during the anticipatory interval. Twenty-five healthy participants performed a cueing task requiring covert attention to the left or right hemifield. EEG source activity was reconstructed with eLORETA and analyzed in 20 regions of interest occipital, parietal, and frontal. Power and directed connectivity, estimated via spectral Granger causality, were assessed in theta, alpha, low-beta, high-beta, and gamma ranges using a time-resolved approach, and connectivity network organization was summarized through hubness and authority indices. Power results showed the typical lateralized modulation in posterior regions in the alpha band, consistent over time, with weaker but qualitatively similar effects in theta and low beta, whereas high-beta and gamma power exhibited less clear modulation patterns. Connectivity and centrality analyses revealed lateralized reorganization of the cortical networks, with interactions increasing ipsilaterally in all bands, but with distinct timing and spatial patterns, as high-beta and gamma connectivity effects emerged later and engaged more ventral regions than lower-frequency bands. These results support the view that covert visuospatial attention relies on multi-band oscillatory mechanisms, with lower and higher frequency bands possibly contributing differently.

ABSTRACT. The Auditory Brainstem Response (ABR) is an electrophysiological signal routinely used in clinical settings to determine hearing threshold and assess auditory pathway integrity. Due to background electroencephalographic (EEG) noise, its assessment requires averaging a large number of sweeps acquired after repeated stimulus presentations. To reduce the number of sweeps, several automated detection indices have been proposed and implemented in commercial devices. However, their relative performance remains uncertain, especially in presence of noise and artifacts, and they generally do not account for non-stationarity in EEG noise. The present work provides a simulation-based comparison of six indices for automated ABR detection: five from the literature (Hotelling’s T² test in time and frequency domains, q-sample uniform score test, Fsp, and Fmp) and a novel modification of the Hotelling’s T² statistic designed to account for EEG non-stationarity. Performance was evaluated using true positive rate (TPR) and false positive rate (FPR), under different signal-to-noise ratios (SNR) and increasing levels of noise non-stationarity. Results show that methods based on the Hotelling’s T² statistic achieved the best performance, with TPRs above 95% at medium SNR, while Fsp and Fmp consistently showed lower detection rates (below 40% at medium SNR) and increased detection times (about three times those of other methods at high SNR). The proposed method mitigated the effects of non-stationary noise, maintaining high TPR (above 90%) across conditions, whereas the others exhibited substantial degradation (down to below 75%). In conclusion, Hotelling’s T²-based methods outperform alternative approaches, and the proposed modification shows promising robustness to non-stationary noise.

ABSTRACT. Photoplethysmography (PPG) is widely used for non-invasive cardiovascular monitoring; however, its performance is known to vary with skin pigmentation, raising concerns regarding reliability across diverse populations. While these effects have been previously investigated, most studies rely on in vivo measurements or global metrics, limiting detailed pulse-level analysis under controlled conditions. This study presents a controlled in vitro investigation of PPG signal quality using a multilayer vascular finger phantom with interchangeable skin tone layers representing different pigmentation levels according to the Fitzpatrick skin colour scale. PPG signals were acquired at three wavelengths (530 nm, 665 nm, and 940 nm – representing the most common wavelengths used in wearable devices) and analysed using a pulse-based framework incorporating six signal quality indices (SQIs), including morphology, signal strength and frequency characteristics. Results demonstrated significant main effects of both skin pigmentation and wavelength across all SQIs (p < 0.001), with infrared (940 nm) consistently providing the highest signal quality, and green (530 nm) showing the greatest signal degradation. A strong interaction effect demonstrated that signal degradation is not uniform but is amplified at shorter wavelengths under darker skin conditions. While most pairwise differences were statistically significant, the entropy-based metric showed limited differences between red (665 nm) and infrared (940 nm) wavelengths. These findings improve understanding and provide a detailed pulse-level characterisation of skin pigmentation effects on light tissue interactions in PPG signals for wearable optical technologies.

ABSTRACT. Wearable optical sensors based on photoplethysmography (PPG) are widely used for non-invasive cardiovascular monitoring; however, their validation is limited by the lack of physiologically realistic and reproducible testing platforms. This study presents the development of a multilayer wrist phantom designed to replicate the anatomical, optical, and mechanical properties of human wrist tissue. The phantom consists of an adipose-mimicking silicone matrix embedding three vascular layers representing the capillaries microcirculation, arterioles and radial artery, positioned at controlled depths together with interchangeable skin layers representing different pigmentation levels. Optical properties of the adipose and skin layers were tailored using absorbing and scattering agents and validated against literature data, showing good agreement across the 500–1000 nm wavelength range. Mechanical characterisation confirmed that the phantom exhibits tissue-like hardness within the range of human soft tissue. The phantom was integrated into a closed-loop pulsatile cardiovascular system to enable controlled flow conditions and perfused using a blood-mimicking fluid. Laser Doppler Flowmetry (LDF) verified continuous pulsatile flow within the microvascular layer, while multiwavelength PPG measurements demonstrated realistic signal generation across all skin tones, with expected variations in amplitude and signal‑to‑noise ratio due to pigmentation-dependent absorption. The proposed platform provides a robust and reproducible tool for the evaluation and optimisation of wearable optical sensors under physiologically relevant conditions.

ABSTRACT. Characterizing task-related brain network reorganization requires analytical frameworks that capture both local interactions and global information flow. Functional connectivity derived from magnetoencephalography (MEG) is often analyzed using static, edge-based measures, which may overlook higher-order and sequential dependencies in neural signal propagation. Here, we investigate a Markov-based framework for task-related brain network analysis by modeling MEG functional connectivity as a transition system and studying the resulting random walk dynamics. Using Human Connectome Project MEG data from healthy young adults, we compared working-memory and story-math tasks. From the corresponding transition matrices, we derived multiscale descriptors, including weighted random-walk motifs, the spectral gap, and a two-step round-trip (return) measure. Statistical analyses across frequency bands revealed robust task-related differences, with consistent effects in the alpha band across all measures and additional beta-band effects localized to right-hemisphere sensorimotor regions. These results demonstrate the potential of Markov-based measures to capture taskdependent brain network organization beyond pairwise connectivity.

ABSTRACT. Atrial fibrillation (AF) is the most common arrhythmia, affecting approximately 50 million people worldwide, with prevalence expected to rise. It is commonly treated with catheter ablation to achieve pulmonary vein isolation (PVI). Traditional thermal methods, radiofrequency ablation (RFA) and cryoablation are effective but may damage surrounding tissue, whereas the newer pulsed field ablation (PFA) appears similarly effective and safer due to its nonthermal mechanism. RFA-based PVI often causes parasympathetic denervation, reflected by increased heart rate and reduced heart rate variability (HRV). HRV is primarily governed by the balance between the sympathetic and parasympathetic branches of the autonomic nervous system (ANS). The long-term effects of PFA on cardiac autonomic function and HRV remain unclear. Here, we present preliminary HRV comparisons from an ongoing clinical study. After ablation, mean and minimum heart rates were higher in the RFA group than in the PFA group for up to 6 months. HRV was significantly lower after RFA at 3 months, with a similar trend at 6 months, although differences were attenuated—likely reflecting partial recovery of autonomic function in the RFA group. These findings suggest better preservation of autonomic regulation with PFA, likely due to less parasympathetic denervation compared with RFA.

ABSTRACT. Pigmented lesions of the face and neck pose a fine-grained differential-diagnosis problem in which several classes overlap visually on dermoscopy and are ultimately separated by histology, which limits what any image-only system can do. The MEDICON 2026 Scientific Challenge frames the task as a seven-class problem on 777 training images and a hidden test set of 335. We describe a 20-model ensemble of four families, three built on the DermLIP dermatology vision-language backbone (two fine tuned on the challenge data, one on external dermatology images) and a ConvNeXt Base classifier at 768 px input. A supervised contrastive term used during DermLIP fine tuning makes its image embedding double as a retrieval index, so any prediction can be inspected against the nearest training images from the predicted and the annotated class. We treat this built-in diagnostic view as a central feature of the system, not a post-hoc add-on. The ensemble reaches an out-of-fold combined score of 66.9, with errors clustering in the clinically expected LM/LMM and PAK/SL groups.

ABSTRACT. Facial pigmented lesions represent a challenging diagnostic setting because benign, premalignant, and malignant entities may show overlapping dermoscopic features, particularly on chronically sun-damaged skin. This study developed and evaluated an image-based deep learning model for the multiclass classification of facial pigmented lesions. A two-step transfer learning strategy was used: a ResNet-152 model was first pretrained on a merged ISIC-derived dataset including 92,044 dermoscopic images across nine diagnostic categories, and subsequently fine-tuned on a facial pigmented lesions dataset comprising 1,112 images from seven diagnostic classes. Class imbalance was addressed using weighted sampling and class-weighted focal loss. On the internal test set, the model achieved a macro-AUC of 0.860, averaged recall of 0.697, macro-F1 score of 0.622, accuracy of 0.675. External testing showed preserved discriminative ability, with a macro-AUC of 0.85. These findings suggest that deep learning models based on contrastive learning training technique may support the differential diagnosis of facial pigmented lesions, although further multicenter validation and calibration refinement are required before clinical implementation.

ABSTRACT. Pigmented lesion classification remains challenging because clinical images are often collected under non-standardized conditions and different lesion categories may show highly similar visual characteristics. This study proposed a two-stage probability-level stacking ensemble framework for seven-class pigmented lesion classification using transfer learning. The dataset consisted of 777 clinical lesion images with histological diagnosis labels, including atypical nevus, lentigo maligna, lentigo maligna melanoma, pigmented actinic keratosis, seborrheic keratosis, seborrheic-lichenoid keratosis, and solar lentigo. To address class imbalance, clinically constrained data augmentation was applied to increase the training data to 1,610 images, with 230 images per class. Three ImageNet-pretrained CNN backbones, including EfficientNet-B0, DenseNet121, and ConvNeXt-Tiny, were trained independently using stratified cross-validation. EfficientNet-B0 was trained using five folds, whereas DenseNet121 and ConvNeXt-Tiny were each trained using three folds. The seven-class probability outputs from the three backbones were concatenated into a 21-dimensional meta-feature vector and used to train a multinomial logistic regression meta-classifier. In internal validation, DenseNet121 achieved the highest mean validation accuracy among the individual backbones, reaching 0.5135. The final stacking ensemble was evaluated on the official challenge test set and achieved a Macro AUC of 0.80, averaged recall of 0.47, Macro F1-score of 0.45, Brier score of 0.67, accuracy of 0.54, and global score of 59.39. These results suggest that probability-level stacking of multiple pretrained CNN backbones provides a feasible baseline strategy for multiclass pigmented lesion classification under limited and non-standardized clinical image conditions.

ABSTRACT. Automated classification of facial pigmented lesions from dermoscopic images is a clinically important yet challenging problem due to the high visual similarity between malignant and benign lesions, severe class imbalance, and scarcity of validated training data. This paper presents our approach to the MEDICON 2026 Scientific Challenge, which targets seven-class differential diagnosis of facial pigmented lesions. Our key contributions are twofold. First, we curated an extended, dermatologist-validated dataset by merging the official MEDICON 2026 training images with publicly available external sources (ISIC Archive and HAM10000), followed by GPT-4 Vision-assisted subclass disambiguation and manual clinical validation by a board-certified dermatologist who reviewed and removed any suspicious or incorrectly labelled samples - yielding approximately 400-500 validated images per class. Second, we propose a classification pipeline based on Swin Transformer V2 (pre-trained on ImageNet-22K) with a custom two-layer SwiGLU classification head, trained with Focal Loss, MixUp/CutMix augmentation, Exponential Moving Average (EMA) of weights, and 8-view Test-Time Augmentation (TTA) at inference. On our internal validation set, the model achieved a competition score of 87.95, with Macro-AUC of 0.972, Averaged Recall of 0.811, and Macro-F1 of 0.810. The official hidden test score was 61.65, revealing a distribution shift between the external-data-enriched training set and the clinical test set - a finding with important implications for future challenge design.

ABSTRACT. Automated classification of skin lesions located on the face and neck presents a significant clinical challenge due to chronic ultraviolet exposure and overlapping morphological characteristics across different diagnostic categories. In this work, we present an image-based deep learning framework developed for the MEDICON 2026 Scientific Challenge. Relying exclusively on raw dermoscopic pixel data as input, our core architecture leverages an ensemble of state-of-the-art convolutional backbones to promote diverse feature representation. These networks are enhanced with integrated Convolutional Block Attention Modules (CBAM) to maximize the extraction of localized discriminative features from subtle malignant transitions while suppressing irrelevant background noise. Evaluated on the organizer-side private hidden test set (N=335), our method achieved a Global Score of 54.92, with a Macro-AUC of 0.78 and an Accuracy of 0.50. These results demonstrate that a neural network with specialized attention blocks can achieve robust diagnostic discrimination across seven facial lesion classes without relying on supplementary clinical metadata. Ultimately, this ensemble-based method provides a scalable solution to assist primary care physicians and nurse practitioners in screening for skin cancer.

ABSTRACT. We present a multi-modal deep learning pipeline for the MEDICON 2026 Challenge, which requires 7-class classification of pigmented skin lesions from dermoscopic images and clinical metadata. Our models combine three pretrained vision backbones, EfficientNetV2-M, Vision Transformer (ViT-B/16), and Swin Transformer-B, with a metadata-guided gating mechanism that adjusts image features. Also, an XGBoost-based metadata classifier complements the deep models by being fully trained on metadata. All components are trained using 5-fold stratified cross-validation with minority class oversampling, Focal Loss with label smoothing, MixUp and CutMix augmentation, and a custom hair-simulation transform. In inference, 8-view Test-Time Augmentation (TTA) is applied, and predictions from all four models are combined via weighted ensemble averaging with temperature scaling. After all these steps, an LLM Agent is used to optimize possible integrations, such as different parameter settings, changing class weights, and retraining XGBoost with different parameters. The proposed system is evaluated using the official challenge score $S_{total}$, which integrates Macro-AUC, Averaged Recall, Macro F1-Score, Brier Score, and Accuracy. The best performance obtained with the model was: $S_{total}$ = 62.81, Macro-AUC = 0.88, Averaged Recall = 0.46, Macro F1-Score = 0.45, Brier Score = 0.59, and Accuracy = 0.60.

ABSTRACT. Classification of images of pigmented skin lesions images from face and neck region is clinically crucial. Benign, malignant, and intermediate lesions may exhibit very similar visual characteristics. This study addresses the seven-class facial pigmented lesion classification problem within the scope of the MEDICON 2026 Scientific Challenge. Two-stage deep learning pipeline was proposed for identifying skin lesions. In the first stage, an ensemble structure consisting of image-only EfficientNetV2-S models trained with three different random seeds was created. External images selected from ISIC-DICM-17k and DERM12345 datasets were harmonized to target classes and included in mixed-source training process. Following mixed-source training, model was fine-tuned using only MEDICON data to better adapt it to MEDICON image distribution. In the second stage, metadata information such as patient age, sex, lesion site, and lesion maximum diameter used. These metadata were combined with image representation using feature-level fusion architecture. Final prediction was obtained by combining the probability outputs of the image-only ensemble model and the models that take metadata into account. Proposed model achieved official evaluation results of 65.99 Global Score, 0.87 Macro AUC, 0.54 Macro F1-score, 0.56 Averaged Recall, 0.60 Brier Score and 0.57 Accuracy.

ABSTRACT. Facial pigmented lesions present substantial diagnostic complexity because benign, premalignant, and malignant entities may share overlapping dermoscopic patterns. This study proposes a transfer-learning framework for the seven-class classification of facial pigmented lesions using a pre-trained foundation model. Methods: To address target-domain data scarcity, the original MEDICON dataset was expanded by harvesting external public datasets, including HAM10000, PAD-UFES-20, ISIC 2019, DERM12345, BCN20000, and ISIC_ARCHIVE. These datasets were mapped to the target 7-class taxonomy and directly merged to create a unified training pool of 11,020 images. The proposed vision-only pipeline extracts 6144-dimensional embeddings using the Google Derm Foundation model [1], followed by a lightweight Multi-Layer Perceptron classifier. Results: Preliminary validation showed moderate discriminative performance, with Macro-AUC of 0.6838, a Brier Score of 0.1079, and an Accuracy of 0.3886. However, class-balanced classification remains limited, reflected by an Average Recall of 0.2904 and a Macro-F1 of 0.3012. Conclusions: Foundation model embeddings can support facial pigmented lesion classification by effectively extracting visual features without computationally expensive pretraining.

ABSTRACT. This paper describes the submitted solution for the MEDICON 2026 Scientific Challenge on multiclass differential diagnosis of facial and neck pigmented lesions. The task is to assign dermoscopy images, optionally accompanied by clinical and acquisition metadata, to seven lesion classes. The submitted candidates used frozen image encoders rather than end-to-end fine-tuning: DINOv2-base, Google Derm Foundation, and PanDerm representations were combined with metadata features and shallow probabilistic classifiers. Metadata were aligned by sample identifier, imputed, standardized or one-hot encoded. Model selection used stratified five-fold out-of-fold validation on the 777 public training samples. The highest internal validation estimate was obtained by blending a DINOv2 + Derm Foundation + metadata branch with PanDerm Large probabilities, yielding Macro-AUC 0.835, averaged recall 0.490, Macro-F1 0.520, Brier score 0.581, accuracy 0.569, and official-style weighted score 63.205.

ABSTRACT. Accurate automated classification of skin lesions is a clinically significant task, particularly under severe class imbalance and limited training data. This paper presents our approach to the MEDICON 2026 Scientific Challenge on dermoscopy image classification, which requires distinguishing seven skin lesion categories from only 777 training images distributed across classes with a maximum imbalance ratio of approximately 15:1. We propose an ensemble of three ConvNeXtV2 models: a Tiny and a Base variant pre trained on ImageNet-1k via Fully Convolutional Masked Autoencoder (FCMAE), together with a third Tiny model initialised from a publicly available ISIC-2019 skin lesion checkpoint through a two-stage domain specific fine-tuning protocol. All models are trained with five-fold stratified cross-validation, class-weighted cross-entropy loss, and AdamW optimisation. The final ensemble averages out-of-fold probabilities from all fifteen fold-checkpoints and applies per-class multipliers optimised by differential evolution combined with Nelder-Mead local search to maximise the official composite score. On the official held-out test set our system achieves a global score of Sₜₒₜₐₗ = 65.46 (AUC = 0.87, Macro Recall = 0.52, Macro F1 = 0.52, Brier = 0.56, Accuracy = 0.60). An empirical analysis of the out-of-fold to test generalisation gap reveals that per-class threshold optimisation appears to overfit strongly at this data scale, whereas the base ensemble transfers with near-zero gap.

ABSTRACT. Obstructive Sleep Apnea Syndrome (OSAS) is a chronic sleep-related breathing disorder defined by recurrent upper airway obstructions. OSAS can lead to a significant reduction in quality of life, and it can be associated with the onset of severe cardiovascular diseases, metabolic disorders, and increased mortality. Current standard diagnostic exams, such as polysomnography, are accurate but often uncomfortable and intrusive to patients and inaccessible for long-term monitoring. To address these limitations, this work presents an innovative non-contact monitoring solution integrating Frequency-Modulated Continuous-Wave (FMCW) Radar technology with a cloud-based data management platform for the automated analysis of vital signs. By leveraging high-precision radar sensing, the solution is able to continuously track the respiratory patterns and heart rate without the need for wearable sensors, ensuring maximum patient comfort. The integrated platform employs advanced signal processing algorithms to detect apnea events and quantify syndrome severity. Preliminary results demonstrate that the system offers a reliable, low-cost, and scalable alternative for early screening and longitudinal home-based monitoring, potentially reducing the clinical burden on healthcare facilities and improving patient outcomes through timely intervention.

ABSTRACT. Objective To evaluate the signal-to-noise ratio (SNR) and image uniformity of 5.0T MRI using a standardized water phantom, and to enhance the reproducibility of quality control procedures. Methods Existing methodologies for SNR and uniformity assessment were systematically reviewed, and a standardized testing protocol based on the SMR 170 phantom was developed. Under identical pulse sequence conditions, various image reconstruction parameters were compared and analyzed to determine optimal settings for image quality control in 5.0T MRI. Results The SNR and image uniformity of the 5.0T MRI system at our institution were evaluated. The results demonstrated that different reconstruction parameters significantly influenced the quantitative outcomes of image quality metrics. Conclusion This study establishes a standardized approach for SNR and uniformity evaluation, along with optimized reconstruction parameters, thereby supporting consistent, reliable, and scientifically sound quality control of 5.0T MRI equipment.

ABSTRACT. Pulsed field ablation (PFA) has recently emerged as a promising non-thermal technique for cardiac arrhythmia treatment. However, the high-voltage pulsed electric fields used in PFA procedures introduce significant challenges for intracardiac electrogram (iEGM) acquisition, including amplifier saturation, severe electromagnetic interference, and signal distortion. Pulsed field ablation (PFA) has emerged as a promising non-thermal technique for the treatment of cardiac arrhythmias. During PFA procedures, real-time intracardiac sensing is essential for evaluating catheter-tissue contact, lesion formation, and electrical activity. However, the high voltage pulsed electric fields introduce significant challenges to conventional sensing systems, including amplifier saturation, electromagnetic interference, and signal distortion. In this paper, a high voltage tolerant multi-modal intracardiac sensing system integrating electrogram (iEGM) and impedance measurement is proposed. The system combines a multi-stage protection circuit, low-noise analog front-end (AFE), and synchronous multi-channel analog-to-digital converters (ADCs) to enable robust signal acquisition under high-voltage conditions. In addition to electrical activity monitoring, an impedance measurement module is incorporated to assess tissue characteristics and catheter-tissue interaction. A field programmable gate array (FPGA) based control unit is employed to coordinate synchronous data acquisition and system operation. The proposed system supports scalable multi-channel configurations and provides complementary electrical and impedance information for comprehensive intracardiac monitoring. This design offers a practical platform for future PFA catheter systems and advanced electrophysiological studies.

ABSTRACT. Transcatheter Aortic Valve Implantation (TAVI) requires extreme precision, yet current navigation mainly relies on 2D fluoroscopy, increasing cognitive load and radiation exposure. Building on our previous work in holographic visualization, this study presents a Proof-of-Concept (PoC) for an integrated Mixed Reality (MR) platform. We propose a twofold approach: (i) a pre-operative planning module that employs hierarchical scaling algorithms to ensure 1:1 metric accuracy for patient-specific anatomical analysis, and (ii) a markerless intraoperative tracking pipeline. By leveraging Sobel edge detection, RGB thresholding, and DBSCAN clustering, we demonstrate the feasibility of aligning 3D holograms with the surgical field using intrinsic anatomical landmarks instead of physical markers. Preliminary results show sub-millimetric accuracy in measurements and robust holographic stability, providing a reliable foundation for enhancing surgical awareness and safety in structural heart interventions.

ABSTRACT. X-ray microtomography (micro-CT) is a non-destructive 3D imaging modality, increasingly used to assess the morphometry of soft tissues, including formalin-fixed paraffin-embedded biopsies. As micro-CT imaging is implemented into histological workflows, sample mounting and holder design becomes relevant, especially for small and mechanically delicate core biopsies. This study evaluated how sample holder type and wall thickness influence X-ray attenuation and reconstructed image quality in micro-CT imaging of tissue biopsies. Two custom 3D-printed polylactic acid (PLA) holders with increasing wall thickness were compared with a routinely used syringe-based mounting. Holder-related attenuation was assessed using transmission measurements of empty holders, while image quality was evaluated through qualitative assessment of reconstructed volumes and tissue-background contrast-to-noise ratio (CNR) in representative prostate and intestine biopsies. Increasing the holder wall thickness resulted in a mechanically more stable setup but resulted in higher attenuation and was associated with increased image noise and reduced CNR. The trends were consistent between sample types. These results demonstrate that the sample holder design influences the imaging conditions and image quality in micro-CT imaging of tissue biopsies and should be considered a part of the acquisition setup. Balancing robustness with minimal holder attenuation is essential when designing or selecting mounting solutions for soft-tissue micro-CT workflows

ABSTRACT. This study explores the transformation of clinical engineering from traditional maintenance to data-driven strategic governance, focusing on the integration of patient safety, operational ethics, and Environmental, Social, and Governance (ESG) goals. Utilizing the i-MMS (Intelligent Medical Management System), we conducted a comprehensive lifecycle assessment of medical assets. In the domain of Technical Assessment, risk evaluations of aging ICU monitors (10–17 years) identified critical End-of-Life (EOL) risks, preventing potential care interruptions. By analyzing maintenance data for 1,036 infusion pumps, the implementation of a "Preventive Battery Replacement" strategy reduced failure rates by 88%. Furthermore, the application of a Predictive Maintenance (PdM) model for CT tubes—monitoring the "Fatigue Index" via i-MMS—allowed for preemptive replacements during non-clinical hours, successfully averting 5-day downtimes and approximately US$30,000 in operational losses per incident. Regarding ESG Sustainability, the transition to digital workflows eliminated tens of thousands of paper records, reducing cumulative carbon emissions by 3,365.37 kgCO2 over 26 months, while centralized dispatching reduced maintenance logistics by 95%. From an Ethical and Social perspective, optimizing point-of-care testing (POCT) systems increased Wi-Fi data upload rates from 31.1% to 89.4%, ensuring 100% data integrity within Electronic Medical Records (EMR). Additionally, digitalization saved 15,101.1 administrative man-hours, significantly mitigating staff burnout. In conclusion, this study demonstrates that modern clinical engineering serves as a vital evaluator of medical technology, leveraging data-driven precision to uphold patient safety, enhance hospital resilience, and achieve measurable sustainability milestones.

ABSTRACT. Background: Traditional clinical engineering departments must maintain rigorous policies for managing medical device recalls and hazard alerts. These notifications originate from diverse sources, including manufacturers, regulatory authorities, and independent services such as ECRI Alerts [1] and RASMAS [2]. However, according to the U.S. Food and Drug Administration (FDA) Center for Devices and Radiological Health (CDRH), the number of recalls issued increased by 97% between 2003 and 2012 [3]. This upward trajectory has persisted into recent years, representing a continuous escalation in global alert volumes up to the present day. Within large-scale healthcare systems, daily manual screening of this expanding volume poses an operational difficulty and increases the risk of omissions [4]. Conventional manual cross-referencing of textual reports against extensive inventories causes delays, potentially postponing corrective actions and compromising patient safety [5]. This study presents a closed-loop, AI-assisted framework to optimise workflow efficiency, traceability, and standardisation. Method: An AI-assisted Safety Alert Workflow Engine was developed utilising a centralised asset database. An automated data pipeline was created to extract safety information directly from regulatory websites, such as the FDA and the Therapeutic Goods Administration (TGA), while Optical Character Recognition (OCR) technology was integrated to parse unstructured recall documents from manufacturers [6]. A Large Language Model (LLM) was deployed to summarise these raw data streams, creating structured database entries containing specific fields: manufacturer, brand, model, serial number, lot, recall class, and source. The engine then cross-referenced these structured attributes against the inventory to identify affected assets. A clinical risk matrix prioritised alerts, automatically generated action plans, and dispatched notifications to engineers via a centralised platform, integrated with device location identifying and tracking [7]. Result: The system enabled automated cross-checking against the master medical device asset database and enables rapid identification of affected device locations, significantly reducing search time across hospital sites for timely corrective action. The transition to digital workflow has reduced the safety alert-to-field action window from several days to a few hours for high-risk cases. In addition, a dual-layer verification mechanism, integrating automated screening with human validation, has enhanced overall system reliability. Together with real-time auto-logging of field actions, the framework further strengthened auditability and data traceability. Conclusion & Significance: With the increasing diversity of medical device models and the growing volume of safety alerts, this study demonstrated the value of automating and digitalising safety alert handling workflows, enabling a proactive, risk-based approach that enhances operational efficiency and supports timely implementation of corrective actions.

ABSTRACT. Background: Managing large-scale healthcare asset across multiple hospital clusters presents challenges including data fragmentation, transcription errors arising from paper-based records and inconsistent adherence to original equipment manufacturer (OEM) testing requirements [1]. Fragmented and localised record-keeping within individual hospitals limits system-wide oversee and foresee of device degradation, maintenance costs and vendor performance. In response to these challenges, this study proposes a data-driven clinical engineering framework that integrates international management standards, specifically ISO 13485 and ISO 55001, to optimise asset management across the entire lifecycle of medical devices [2]. Methods: The framework was developed based on international management standards, with ISO 13485 for medical device quality and ISO 55001 for asset management as primary references, complemented by the 4M1E methodology (Man, Machine, Material, Method, and Environment). A digitalised Biomedical Engineering Services (BES) e‑Form platform was deployed to support numbers of medical devices across more than 40 public hospitals. Conventional maintenance checklists were transformed into model-specific electronic forms embedded with validation logic derived from original equipment manufacturer specifications, thereby standardising data entry and maintenance workflows [3]. The system architecture also incorporates two analytical modules: 1) Master Tester Equipment Calibration System, which manages calibration schedules and records to ensure the use of valid and calibrated test instruments; and 2) Scoring Engine integrating historical maintenance data for each equipment category, including fault frequency, parts costs, labor hours and downtime records, to provide predictive maintenance and performance insights for lifecycle value assessment. Further optimisation was achieved through a hybrid maintenance workflow, combining centralised and decentralised approaches for high-risk medical devices, including a Centralised Infusion Pump Maintenance Workshop for high-volume, controlled-condition servicing and a Decentralised Renal Technology Special Duty Unit for responsive on-site support, thereby maximising the efficient use of labor and technical expertise [4]. Results: Following implementation of the developed framework and BES e-Form platform, over 65,000 maintenance cases were managed by the platform annually, with preventive maintenance compliance consistently exceeding 95%. Moreover, the adoption of centralised and decentralised workflows enhanced turnaround time and resource utilization, resulting in up to a 30% increase in technician productivity while maintaining consistent maintenance quality across hospitals. Furthermore, operating costs were reduced by approximately 15% per asset category, attributable to the integration of predictive maintenance strategies, optimized spare parts allocation, standardised procedures and improved workforce deployment. Conclusion & Significance: The integration of ISO-based governance frameworks approach with a centralised digital data infrastructure facilitates the transition of large public healthcare systems from reactive maintenance practices to a proactive, lifecycle-oriented management model, thereby enhancing asset management across the entire medical device lifecycle.

ABSTRACT. Early diagnosis and monitoring of neurodegenerative diseases can benefit from quantitative digital biomarkers of neuromuscular function. In this context, handwriting analysis provides a rich, non-invasive source of motor and cognitive information, capturing fine features of sensorimotor control. Here, we present the development of a vision-based tactile sensing interface for quantitative assessments of handwriting and drawing. The novel device integrates a soft mechanochromic tactile sensor that changes structural colour under mechanical deformation. By detecting pixel colour (hue) variations with an internal camera, the system can track contact position and applied pressure with high spatial resolution, fast response, and low hardware complexity. We present ongoing characterizations of the first prototype version of this new device.

ABSTRACT. Mental health diagnostics often rely on symptom based assessments, limiting the discovery of predictive biomarkers. This paper introduces a scalable framework that integrates normative modeling and self-supervised representation learning to decode individual heterogeneity in brain structure and function across 100,000+ multi-modal scans. Using Gaussian process regression for normative trajectories and variational autoencoders with contrastive learning for latent feature extraction, we generate subject-level deviation scores that capture clinically meaningful variation. Key results show that (i) deviation patterns identify four neurobiological subtypes within schizophrenia with distinct symptom profiles, (ii) multi-modal representations achieve 0.84 AUC in distinguishing patients from controls, and (iii) longitudinal models predict treatment response with 71% accuracy at 12 weeks. These findings demonstrate that scalable brain-based models can provide personalized, interpretable insights for improved and fairer mental health care.

ABSTRACT. Simulating soft object interactions is essential for creating immersive and realistic virtual reality environments, particularly in applications such as medical training, where the perception of tissue softness is critical. However, current haptic technologies still face limitations in reproducing the mechanical compliance of fingertip interactions. Soft actuators offer a promising solution, thanks to their intrinsic compliance, lightweight design, and ability to conform to the skin, enabling more natural cutaneous stimulation. Taking advantage of these properties, we recently developed compact, wearable tactile displays based on pneumatic actuation. To evaluate the performance of this technology in a realistic scenario, we developed a virtual abdominal palpation system integrating hand tracking with wearable pneumatic fingertip displays. We present preliminary psychophysical tests on healthy participants that show high accuracy in discriminating between regions of a simulated abdomen with different levels of softness.

ABSTRACT. Accurate continuous prediction of lower-limb gait trajectories is essential for the effective control of rehabilitation robots that provide personalized assistance to post-stroke patients. However, current state-of-the-art machine learning methods rely on offline batch learning trained with pre-collected data, which is constrained by the limited availability of post-stroke data and the substantial distribution differences in sEMG signals and kinematics across patients. These differences violate the assumption that source and target domains share the same distribution, undermining the direct deployment of offline-trained models. While offline transfer learning can mitigate this gap, it still requires patient data to be collected in advance, which is costly and burdensome—particularly for severely impaired individuals. To address these challenges, this paper proposes an online transfer learning (OTL) framework for end-to-end continuous prediction of hip joint angles in post-stroke patients across multiple walking modes using sEMG signals. The framework couples a temporal convolutional network (TCN) with a recursive least squares (RLS)–based autoregressive residual compensator. The source-domain network is pre-trained offline on data from healthy individuals walking at self-selected speed, and the target-domain network is updated in real time from streaming patient data via a system-wide update strategy with two data buffers, eliminating any need for pre-collected patient data. The end-to-end design removes manual feature engineering, simplifying the signal-processing pipeline and enabling real-time deployment. The method was validated on a public dataset (15 healthy adults, 26 post-stroke survivors, 16-channel sEMG) across five walking modes and benchmarked against LSTM, MLP, offline TCN, and CNN-BiLSTM baselines. The proposed framework achieved RMSE ≤ 2.49°, NRMSE ≤ 0.0679, NMAE ≤ 0.051, and R² ≥ 0.91 across all walking modes, significantly outperforming every baseline (P < 0.001). Its total system response time of approximately 40–50 ms—well below the 200 ms prediction interval—confirms real-time feasibility, and channel-level ablation showed the learned feature attributions are consistent with established biomechanical findings. In conclusion, the proposed OTL framework delivers accurate, robust, and generalizable real-time hip-angle prediction without pre-collected patient data, reducing experimental burden and demonstrating strong potential for cost-effective, personalized lower-limb rehabilitation robots.

ABSTRACT. Hepatocellular carcinoma (HCC) is still one of the most important causes of cancer-related death in the world, which is largely attributable to the delay at diagnosis and the limited sensitivity of the current biomarkers for early diagnosis. High-throughput genomic sequencing, artificial intelligence (AI) and precision medicine have created potential to change the way people are diagnosed with HCC from reaction to prediction and prevention. This chapter introduces a detailed machine learning based analysis of the gene pair signatures for the early detection of HCC, highlighting the growing importance of the relative gene expression of the machine-learned gene pair signatures as informative and powerful biomarkers. In contrast to single-gene signatures, gene-pair signatures capitalize on the concept of relative expression ordering, thus improving platform, population, and disease context reproducibility, and are able to detect the dynamic molecular interactions in hepatocarcinogenesis. Next generation computational approaches are discussed – deep learning, graph neural networks, transformer models, self-supervised learning and foundation models – in development of the predictive gene-pair signatures, discovery, optimisation and validation. It also explores some of the challenges of integrating multi-omics data, federated learning, digital twin cohorts, liquid biopsy, and explainable AI into creating an adaptive diagnostic ecosystem that learns over time and can support clinical decision making. The focus is on the challenges around translation such as regulatory approval, clinical validation, algorithmic fairness, data privacy and the ethical considerations of AI-powered genomic diagnostics. Future outlook includes organoid validation and validation of features, feature selection in the context of quantum technologies, translating criteria for precision oncology to specific populations and establishing cost-effective point-of-care screening systems to enhance global access to precision oncology. It presents an outlook on the problems of deep learning in the era of post-genomic HCC diagnostics by combining the developments in computational biology, machine learning, and systems genomics. These plans feature gene-pair signatures as surrogates for living patients, continuously growing with biological discoveries and clinical experiences, and ultimately responsible for earlier detection, individualized surveillance, and enhanced therapeutic results, all ideally with a backdrop of a long-term vision of precision prevention to diminish the global burden of HCC.

ABSTRACT. Bite force is an important functional parameter used to assess the biomechanical performance of the stomatognathic system. Its measurement is relevant in dentistry, prosthodontics, orthodontics, physiotherapy and biomedical engineering, particularly for evaluating occlusal function, masticatory efficiency, parafunctional activity and temporomandibular disorders. This work presents the development and preliminary verification of a prototype intraoral device for the non-invasive measurement of bite-related compression forces. The proposed solution is based on a mouthpiece-like measuring insert equipped with pressure-sensitive elements and designed to enable independent assessment of loading on the left and right sides of the dental arch. The design process included the development of the bite element geometry, material selection and preparation of the prototype for controlled compression testing. Particular attention was paid to minimizing the influence of the device on natural mandibular positioning while maintaining mechanical stability and measurement sensitivity. An important feature of the concept is its potential for personalization using CAD/CAM workflows and additive manufacturing, allowing adaptation to individual dental arch morphology. Preliminary tests confirmed the feasibility of distinguishing unilateral loading and registering side-specific pressure changes under laboratory conditions. The prototype may serve as a basis for further calibration, optimization and development of a low-cost, patient-specific tool for objective occlusal load assessment.

ABSTRACT. Abstract. Reliable access to medical oxygen remains one of the most urgent unmet healthcare needs in many low- and middle-income countries. In Africa, hospitals and rural health facilities continue to experience shortages caused by weak infrastructure, unstable electricity, inadequate maintenance systems, and fragmented supply chains. This paper examines sustainable medical oxygen ecosystems and green hospital technologies as scalable solutions for last-mile healthcare delivery. A review-based and implementation-oriented approach was adopted using published evidence, field experiences from oxygen infrastructure projects, biomedical engineering interventions, and existing literature on sustainable healthcare technologies Findings indicate that decentralized oxygen systems, Oxygen-as-a-Service models, renewable-powered concentrators, pressure swing adsorption plants, digital monitoring platforms, and strengthened biomedical engineering systems can significantly improve reliability and affordability. Lifecycle-oriented approaches outperform donation-based models that often fail after initial deployment. Africa has an opportunity to leapfrog legacy systems by integrating resilient oxygen infrastructure with climate-smart hospital design. Sustainable oxygen ecosystems should therefore be prioritized in health financing, universal health coverage plans, and emergency preparedness strategies.

ABSTRACT. The integration of artificial intelligence (AI) and machine learning (ML) into medical imaging has generated significant optimism for closing the diagnostic gap in low and middle income countries (LMICs). Yet a critical and largely unaddressed prerequisite persists in the literature: the functional integrity and calibration quality of the radiology equipment on which AI algorithms depend. A miscalibrated X-ray unit or poorly maintained CT scanner introduces image artefacts and dosimetric errors that no algorithm can correct downstream. This paper argues that without rigorous quality assurance (QA) and quality control (QC) frameworks for radiology equipment, AI assisted diagnostics cannot achieve reliable clinical outcomes in resource constrained settings. Drawing on a scoping review of AI adoption in LMIC medical imaging and field evidence from the Christian Health Association of Kenya Medical Engineering Services (CHAK MES), a certified centre of excellence in radiation safety and radiology equipment maintenance in Sub-Saharan Africa, we examine the intersection of equipment integrity, AI diagnostic performance, and regulatory compliance. Five structural barriers are identified and a staged implementation framework is proposed that positions QA/QC as the foundational layer upon which context-appropriate AI diagnostic tools must be built. The findings carry direct implications for biomedical and clinical engineers, health policymakers, and the global community working toward equitable diagnostic access.

ABSTRACT. With the progressive aging of the population, the prevalence of chronic diseases, such as type 2 diabetes mellitus (DM2) and cardio- vascolar disease, including chronic heart failure (CHF), increases signifi- cantly, posing a growing challenge for healthcare systems. Recent advances in imaging analysis, particularly the application of non- linear trimodal regression analysis (NTRA) to radiodensitometric distri- butions in CT cross-sections, have enabled a detailed characterization of muscle, fat, and connective tissue quality; using the AGES-Reykjavik dataset, we analyzed 3,156 older adults over a five-year period and ex- tracted 11 parameters derived from NTRA, where N, μ, σ, α represent amplitude, location, width, and skewness, respectively. The data were analyzed using Linear Mixed Effects Models (LMMs) which allow mod- eling both intra- and inter-individual changes over time. The model was used to compare subjects who remained healthy with those who devel- oped the disease during follow-up, evaluating differences both at baseline and in their longitudinal trajectories over time. LMM analysis showed significant differences (p-value ≤ 0.05) for both diseases. Compared to the healthy group, DM2 cases showed higher (Nfat, Nmuscle, μmuscle and Nconnective) at T0, with Nconnective exhibiting also a faster increase over time. For CHF, baseline predictors included μfat, σmuscle, μmuscle, αmuscle, Nconnective and μconnective , while longitudinal divergence was driven by αfat, Nmuscle and σconnective. These results demonstrate that NTRA parameters can effectively dis- criminate pathological progression from physiological aging, offering a non-invasive tool for early detection of DM2 and CHF via routine CT scans.

ABSTRACT. Cold atmospheric plasma (CAP) has attracted increasing interest for biomedical applications including sterilization, wound healing, tissue modulation, and experimental oncology. However, many existing CAP platforms rely on continuous external gas supply, increasing system complexity, operational cost, and limiting portability. This paper presents the design and preliminary characterization of a compact programmable plasma system based on a sealed reduced-pressure rare-gas applicator energized by resonant high-voltage excitation generated through a dedicated flyback transformer stage. The applicator simultaneously functions as gas reservoir, dielectric barrier, discharge chamber, and treatment interface, thereby eliminating the need for continuous gas flow during operation. The prototype operates from a 12 V DC supply and uses a pulsed high-voltage stage capable of kilovolt-class excitation. Electrical measurements demonstrated no-load electrode voltages up to 17 kVpp and loaded operation in the kilovolt range under multiple coupling conditions. The chamber interface was intentionally designed to exhibit a negative mean surface potential, which may be relevant for future bioelectrical interaction studies. Continuous thermal testing over 30 min showed chamber temperature increase from 22°C to only 30°C, while the switching-stage heatsink stabilized near 60°C. In addition to the fixed-frequency hardware control mode, the architecture includes a BLE-capable controller module for future programmable pulse delivery and user-defined treatment protocols. These findings support the feasibility of sealed rare-gas plasma systems as a promising platform for portable and controllable biomedical plasma technologies.

ABSTRACT. Rapid advances in technology and software tools are reshaping the expectations placed on biomedical engineers, prompting a critical re evaluation of existing curricula. While traditional foundational subjects remain essential, emerging domains must be thoughtfully integrated to prepare a new generation of professionals for evolving roles in research, industry, and health care. Future biomedical engineers will require broader and more adaptive competencies, extending beyond core engineering to include computer science, data analytics, regulatory, legal and ethical frameworks, digital health and soft skills. A key challenge is ensuring that education remains closely aligned with real world job requirements and reflects current trends in clinical practice, industry Innovation, and global health priorities. Equally important is clarifying and strengthening the relationship between biomedical engineering and the rapidly expanding field of digital health, including AI-driven solutions and data-intensive health technologies.

ABSTRACT. Rapid advances in technology and software tools are reshaping the expectations placed on biomedical engineers, prompting a critical re-evaluation of existing curricula. While traditional foundational subjects remain essential, emerging domains must be thoughtfully integrated to prepare a new generation of professionals for evolving roles in research, industry, and healthcare. Future biomedical engineers will require broader and more adaptive competencies, extending beyond core engineering to include computer science, data analytics, regulatory, legal and ethical frameworks, digital health, and soft skills. A key challenge is ensuring that education remains closely aligned with real-world job requirements and reflects current trends in clinical practice, industry innovation, and global health priorities. Equally important is clarifying and strengthening the relationship between biomedical engineering and the rapidly expanding field of digital health, including AI-driven solutions and data-intensive health technologies. This paper discusses strategic guidelines and practical recommendations for future curriculum development, including the identification of critical topics and the design of flexible, competency-based teaching approaches. Emphasis is placed on experiential learning through hands-on exercises, interdisciplinary seminars, problem-based and project-based learning, and real-world case studies embedded within relevant regulatory, technical, and professional contexts to ensure meaningful and sustainable educational impact.

ABSTRACT. Rapid advancements in digital health—including artificial intelligence, interoperable platforms, and data-driven healthcare—are fundamentally reshaping the competencies required of biomedical engineers. This shift necessitates a transition from traditional curricula toward integrative, experiential learning models that align with real-world healthcare ecosystems. This work presents an interdisciplinary educational framework designed to bridge the gaps between academia, industry, and healthcare providers. Inspired by a long-standing Summer School on emerging technologies for healthcare and independent living, the model integrates engineering, health sciences, and social sciences to address challenges related to healthy ageing within the Latin American context. Implemented over ten consecutive years during inter-semester periods at Pontificia Universidad Javeriana—and supported by international societies such as IFMBE and CORAL—the program aims to strengthen soft skills often underrepresented in traditional curricula. It fosters meaningful learning environments where students apply technical knowledge to real-world case studies through critical and reflective thinking. The program targets senior undergraduate students in biomedical engineering, computer science, mechatronics, and medicine. Participants leverage prior knowledge in signal and image processing, data science, machine learning, and project formulation. The pedagogical approach emphasizes experiential learning through case studies focused on chronic disease management, fall prevention, functional decline, and the digital inclusion of older adults. Students employ structured innovation methodologies to design user-centered, scalable digital health solutions, supported by clinical simulation environments, engineering laboratories, and industry collaboration. The framework incorporates interoperability, cybersecurity, cloud infrastructure, and AI-driven decision-making, while integrating essential ethical and regulatory considerations. Designed for versatility, this model can be embedded into standard curricula as elective or capstone courses in engineering, innovation, and entrepreneurship. It provides a practical pathway for aligning biomedical engineering education with the evolving demands of the global healthcare industry.

ABSTRACT. Biomedical Engineering (BME) has become a vital component of the modern healthcare architecture, driving technological innovation and necessitating a highly specialized, multidisciplinary workforce. Due to the field's inherent complexity, educational programs across the European Higher Education Area (EHEA) display significant curricular heterogeneity, which complicates graduate competency assessment and hinders student mobility. This paper examines the historical progression of BME education and provides an exhaustive update to the pan-European BME program database. By mapping 559 active Bachelor (BSc), Master (MSc), and Doctoral (PhD) programs in 2026, this study offers a comparative analysis against the 2020 benchmark data. The primary objective is to establish a transparent, accessible repository that empowers students to navigate the EHEA and provides a baseline necessary to modernize Curricula Reformation and Harmonization (CRH-BME) criteria.

ABSTRACT. Medical devices are essential enablers of modern healthcare, with the broader sector accounting for approximately 4.4% of global greenhouse gas emissions. While recent research has begun to map the environmental impact of the medical device lifecycle, from manufacturing to end-of-life, the industry now faces a dual challenge: the race toward Net Zero and the rapid integration of Artificial Intelligence and Software-as-a-Medical-Device. As emissions shift from physical supply chains toward high-energy computational infrastructures, existing regulatory frameworks, such as the EU Medical Device Regulation (MDR 2017/745), reveal a critical blindspot regarding the full-lifecycle impact of digital health. Drawing on the methodology established in our recent scoping review, this presentation maps the intersection of circular economy principles, medical device regulatory requirements, and the emerging environmental footprint of digital health. While this is something current biomedical engineers are just coming to terms with, the next generation of biomedical engineers will need to be better prepared for these aspects. This talk provides a concrete blueprint for BME education, advocating for a transition from traditional engineering competencies to a hybrid model that embeds frugal innovation and digital sustainability literacy. By fostering a design framework that balances hardware longevity under strict MDR compliance with resource-efficient software deployment, we can equip future engineers to advocate for accountability across both physical and digital healthcare lifecycles.

ABSTRACT. Motion sickness represents a critical occupational hazard in maritime and transport sectors, creating a pressing need for objective, preemptive screening tools to assess an individual's susceptibility. Traditional predictive models often rely solely on subjective questionnaires or isolated physiological signals. This study investigates the predictive power of combining Center of Pressure (COP) and Head kinematics to identify Motion Sickness Proneness (MSProne). Using a subset of the BioVRSea database (N=18), we extracted 54 kinematic features—including time-domain, frequency-domain, and non-linear complexity metrics (Sample Entropy, Higuchi Fractal Dimension), from both COP and Head trajectories. A targeted feature fusion approach followed by a Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression with 10-repetition 3-Fold Cross-Validation yielded a mean classification accuracy of 67.2% and a mean Area Under the Curve (AUC) of 0.72. The most robust predictive biomarkers were non-linear: Head Antero-Posterior Higuchi Fractal Dimension and COP Ellipse Area. These findings demonstrate that non-linear sway complexity and the targeted fusion of Head and COP kinematics are robust, objective biomarkers for predicting innate Motion Sickness Proneness using a VR-based screening protocol.

ABSTRACT. Hypergraph-based methods have emerged as a powerful framework for the analysis of brain structural connectivity, extending classical graph-theoretic approaches by encoding higher-order multi-region interactions. The Boundary Scale (BS) model provides an iterative mathematical framework that progressively amplifies these interactions starting from a binary incidence matrix. While previous studies have applied the BS model to Region-to-Region adjacency matrices, this approach requires an artificial conversion into incidence matrices, resulting in representations of variable dimensions and high computational cost. In this study, we investigate the use of Tract-to-Region (T-R) matrices — which constitute a structurally native incidence matrix input for the BS model — applied to a longitudinal dataset of 32 healthy subjects undergoing twelve weeks of Quadrato Motor Training. T-R matrices were binarized at four overlap thresholds (0.15, 0.20, 0.25, 0.30) and analysed through four iterative BS scales. A paired t-test on average node degree between baseline and the final timepoint revealed significant increases at thresholds 0.20 and 0.25 from the first BS iteration onward (p ≤ 0.030), with the strongest effect at threshold 0.25, iteration 2 (t = 2.603, p = 0.014). These results demonstrate that T-R incidence matrices are sensitive to training-induced structural changes and represent a computationally efficient and methodologically sound alternative for hypergraph-based connectome analysis.

ABSTRACT. Background: The incidence of pleural effusion (PE), particularly secondary to malignancy and heart failure, is expected to increase due to population ageing. During therapeutic thoracentesis (TT), an excessive pleural pressure (Ppl) fall caused by the lack of lung re-expansion may lead to complications such as pulmonary edema and pneumothorax.

Objectives: The aim of this preliminary analysis of an ongoing prospective study was to investigate effects of cough, CPAP, and a pause in PE withdrawal on re-expansion in patients undergoing TT.

Methods: Before TT, patients were randomly assigned to one of three groups: cough on request, CPAP (5 cmH₂O) and no intervention. PE was drained in portions of 200 mL, each followed by the 60-second period of Ppl measurement. After drainage of each 600 mL, longer 240-second periods with coughing, CPAP or nothing were introduced. Ppl increase (dPpl) during the measurement periods was treated as an index of re-expansion.

Results: Strong Pearson’s correlation (r=−0.9) between dPpl and the Ppl value at the periods’ beginning suggested the following differential equation: dPpl/dt=−alpha*(Ppl−Ppl0), where alpha indicates the re-expansion rate and Ppl0 is some steady-state Ppl value. This equation explained the observed exponential Ppl increase. Neither coughing nor CPAP accelerated re-expansion significantly. However, these findings must be validated in the full study cohort once enrolment is complete.

Conclusion: Re-opening needs time and thus slower PE drainage seems to be the simplest method to prevent excessive Ppl fall.

This study was supported by the National Science Center, Poland (grant N 2019/35/B/NZ5/02531) and IBBE PAS.

ABSTRACT. Background: The evaluation of heart sounds produced by the beating heart and the resulting blood flow through it provides a useful method for diagnosing cardiovascular disease (CVD). A phonocardiogram (PCG) signal captures a range of heart sounds, including normal and pathological ones like S3, S4, and murmurs, gathered through stethoscope-assisted cardiac auscultation. Countering subjectivity and the high rate of diagnostic errors, computer-aided detection (CAD) systems play a crucial role in identifying abnormal heart sound conditions with significant importance. In the case of analysis with a computer application statistical model or dynamic model-based could be used. Method: The primary objective of this study is to provide heart condition differentiation using dynamic model simulation based on three coupled ordinary differential equations (ODE). The correct positioning and algorithmic alignment of each heart cycle have been implemented on experimental data, significantly influencing the model's efficacy. In this method, 895 cardiac cycles were segmented, comprising 149 normal beats with only S1 and S2 and 746 abnormal beats with either S3 or S4. The model was fine-tuned by fitting it with the Genetic algorithm using input derived from actual signals, enabling the identification of normal and abnormal sounds associated with S3 and S4. Result: In the heart condition diagnosis using dynamic models, the time-domain correlations between the simulated models and experimentally observed data consistently exceed 0.9 across all examined cases, indicating strong agreement. Conclusion: The study introduces a strong methodology utilizing dynamic modeling to fit experimental data with the Genetic algorithm by extracting key parameters from real signals to perceive the physiological aspects of heart conditions quantitatively.

ABSTRACT. Judo-specific training may promote inter-limb asymmetries, however, evidence comparing dominant (DOM) and non-dominant (NDOM) lower-limb force-time behavior during the countermovement jump (CMJ) in elite judokas remains limited. This study compared DOM and NDOM limbs using both discrete force-time variables and continuous waveform analysis. Thirty-seven elite Brazilian judo athletes performed bilateral CMJs on dual force plates. Peak force, phase-specific impulse, and mean force were calculated for the yielding, braking, and concentric phases. Force-time waveforms were also analyzed using Statistical Parametric Mapping (SPM). Cross-correlation and coefficient of variation (CV) were used as complementary waveform analyses, and an exploratory asymmetry classification was performed using the asymmetry index (AI). No significant differences were found between DOM and NDOM limbs for any discrete variable (p > 0.05), with trivial-to-small effect sizes. Likewise, SPM revealed no significant clusters across the movement cycle. Cross-correlation showed very high waveform similarity between limbs (r = 0.970 ± 0.032). Most athletes were classified as symmetrical, although a greater proportion of asymmetrical athletes was identified for yielding impulse and yielding mean force. These findings indicate a high degree of lower-limb symmetry at the group level in elite judo athletes, while suggesting that the yielding phase may be more sensitive to individual asymmetry patterns.

ABSTRACT. Abstract. This pilot proof-of-concept study evaluates a multilevel linguistic assessment paradigm combined with task-based functional magnetic resonance imaging (fMRI) to characterize brain activation patterns associated with second-language processing in university students. Participants were assigned to beginner (n = 5), intermediate (n = 5), and advanced (n = 4) proficiency groups based on standardized linguistic assessments administered by foreign language faculty. The experimental design employed a block-design task-based fMRI paradigm with language-processing stimuli presented in the native language (L1) and the second language (L2). Neuroimaging data were analyzed using the General Linear Model (GLM). As an exploratory pilot study, activation maps were examined at a voxel-wise threshold of p < 0.001 (uncorrected), with cluster sizes ranging from 13 to 695 voxels. Results revealed distinct activation patterns across proficiency levels. Beginner participants exhibited widespread bilateral activation involving frontal and occipital regions, reflecting increased cognitive demand. Intermediate participants showed transitional patterns with reduced spatial dispersion and greater engagement of language-related regions. Advanced participants demonstrated focal, left-lateralized activation in classical language areas, including the inferior frontal gyrus and temporal cortex. These findings suggest a shift from distributed to specialized neural recruitment as language proficiency increases.

ABSTRACT. Early-life exposure to lead remains a public health problem because the developing brain has limited biological margin for error. Even concentrations that do not produce overt poisoning may interfere with synaptic maturation, vascular regulation, and the integrity of the blood-brain barrier. This conference paper integrates a neurovascular interpretation with structural magnetic resonance imaging findings in Mexican preadolescents with prenatal and early childhood lead measurements. Blood lead was characterized across seven developmental windows, from the second and third trimesters of pregnancy to birth and early childhood. Three-dimensional T1-weighted structural magnetic resonance imaging was used to quantify regional brain volumes, with atlas-based morphometric processing and region-wise analysis adjusted for total brain volume. Median lead concentrations were highest around birth and gradually declined during childhood. Nominal associations were observed in frontal, temporal, parietal, occipital, ventricular, and subcortical regions, including the hippocampus, inferior temporal gyrus, frontal pole, posterior orbital gyrus, precuneus, posterior cingulate gyrus, calcarine cortex, ventral diencephalon, and basal forebrain. The third trimester showed the largest number of associated regions, suggesting that late gestation may be a particularly sensitive window for brain morphometry. Taken together, these findings support the idea that lead-related structural changes should be interpreted not only as isolated regional volume differences, but also as possible signatures of disrupted neurovascular development.

ABSTRACT. The growing trend in the design of artificial intelligence models used for Magnetic Resonance Imaging (MRI) has been the emphasis on validation schemes that ensure reproducibility via controlled experiments, without regard for the representativeness of datasets – the degree to which datasets can cover variations in real-world patient populations, imaging modalities and environments. This paper presents a perspective-based analytical synthesis of this issue from a biomedical engineering point of view, examining the ways in which choices made in various phases of AI model creation can lead to or reduce the generalization problem. A classification scheme with four types of domain shift that occurs during MRI imaging is introduced, based on previous classifications but taking labeling and temporal shifts into consideration separately. The trade-off model highlights the inherent incentive discrepancy that results in the systematic underinvestment in representativeness within the field, whereby optimizing reproducibility offers easily observable feedback during development whereas poor representativeness poses delayed costs realized only when the technology is implemented clinically. Empirical studies have documented substantial performance degradation on external datasets – with reductions as large as 0.21 AUC points observed in specific MRI contexts (e.g., prostate imaging) – illustrating the practical consequences of this trade-off. Pipeline-based harmonization and validation techniques, as well as stratified performance reporting, are some proposed solutions to the issue, along with implications in foundation models and regulatory guidelines by the FDA and EU AI Act.

ABSTRACT. TraumaKc spinal cord injury (SCI) leads to severe neurological disability with vast medical and socioeconomic repercussions, affecKng an esKmated 2.5 million people worldwide. Despite various experimental approaches, most current therapeuKc strategies fail to significantly restore lost motor or sensory funcKons. This study invesKgates the neuroprotecKve and regeneraKve potenKal of plasma- synthesized polymers containing amines (pP-NH), which have previously demonstrated the ability to recover voluntary movement in posterior joints in rat models. We conducted a longitudinal evaluaKon using 3.0 Tesla magneKc resonance imaging (MRI) in adult Wistar rats subjected to a complete spinal cord transecKon at the T9 level and subsequent pP-NH implantaKon. Advanced protocols, including Diffusion Tensor Imaging (DTI) and 1H MagneKc Resonance Spectroscopy (MRS), were uKlized eight weeks post-injury to quanKfy pathophysiological metabolites and microstructural integrity. Findings revealed FracKonal Anisotropy (FA) values ranging from 0.3 to 0.7, indicaKng highly organized fibre orientaKon and potenKal axonal reconnecKon at the lesion site. Furthermore, metabolic analysis idenKfied an NAA/Cr raKo of 0.917 and a Cho/NAA raKo of 0.907, reflecKng a specific profile of neuronal integrity and acKve cellular remodelling or reacKve gliosis. These results confirm the feasibility of in vivo metabolic monitoring and suggest that pP-NH implants posiKvely modulate the injured microenvironment to promote funcKonal recovery.

ABSTRACT. Manufacturers of medical devices (MDs) and in vitro diagnostic devices (IVDs) must comply with increasingly complex and evolving requirements regarding the labeling information to be included with MD/IVD. At the same time, devices are becoming smaller, more sophisticated, and increasingly connected, adding challenges for users to access, understand, and correctly interpret essential labelling information. MedTechLabel is a three-year (2026–2029) Innovative Health Initiative (IHI) project set out to develop a universal, digital labelling framework specification for all categories and classes of MDs and IVDs. To reshape labelling practices, improve user experience, reduce administrative burden, and enhance operational efficiency and regulatory compliance for industry, MedTechLabel addresses two key challenges: (1) scalability and efficiency through greater digitalisation of labelling information; and (2) more accessible information for diverse users, with curated displays that support different roles and needs. To achieve this, the digital labelling framework specification seeks consensus among all involved stakeholders: manufacturers, notified bodies (NB), regulators, health authorities, healthcare professionals, patients, and others. This poster presents how MedTechLabel aims to fulfil this mission and invites input from MEDICON participants on how best to address their needs in the MD/IVD ecosystem.

ABSTRACT. Reading difficulties in children are increasingly prevalent and one of the digital technologies that is commonly used in reading assessment studies is eye-tracking. In this paper, eye-tracking metrics were used to introduce a rule-based system intended for personalized ranking of text presentation strategies for children. The system was evaluated on 23 Serbian-speaking third-grade children, selected based on the assessment of reading and related language skills. Initially, each child specified the preferred text presentation through an interactive application, including font type and size, spacing, and color combinations. An eye-tracking experiment was then conducted in which participants read two stories across separate sessions, with text segments presented using different modification strategies: basic text formatting, average preferred settings, individual preferences, and modifications based on experience from practice. The modifications were evaluated using eight eye-tracking metrics related to eye fixations, including span, duration, dispersion, and gaze self-intersection. Each feature produced a text modification ranking with higher feature values corresponding to more reading struggles. For each child, the ranks were then averaged across all features and both stories to obtain a personalized ranking. Results show variability in individual preferences, with no universally optimal configuration. However, standard text formatting consistently ranked lowest, highlighting the possible benefit of adaptive presentation. The proposed results support further development of personalized reading assistance for children, providing a non-invasive framework for assessing a broad spectrum of personalization approaches.

ABSTRACT. Anatomically-realistic digital twins of neural implants are useful in the scope of neurotechnology research and development. We present an open-source, modular pipeline that bridges neuroanatomical atlases and physics-based simulation, enabling finite element (FE) analysis of brain-implant mechanics without requiring commercial software or manual geometry preparation. Starting from the Waxholm Space rat brain atlas, the pipeline automatically produces a watertight (i.e., free of boundary gaps), simulation-ready tetrahedral mesh compatible with the FEniCS FE solver, alongside a fully parametric implant geometry (electrode shaft, cement layer, anchoring cross) ready for Boolean assembly. The modular architecture supports straightforward substitution of subject-specific segmentations or alternative implant designs via a single configuration file. The assembled multi-domain mesh is intended for viscoelastic simulations of tissue deformation under physiological cyclic loads, such as those induced by breathing and locomotion, providing a practical basis for computational studies of neural implant mechanics and device design. All code is publicly available in GitHub.

ABSTRACT. Game features can improve engagement in cognitive assessment, but they may also change response strategies and threaten score interpretability. This paper combines general design and validation considerations for gamified cognitive assessment with an initial formative evaluation of a three-part prototype based on Mini-Cog, Trail Making Test, and Stroop principles. The paper argues that such systems should be interpreted in relation to intended use, construct definition, task equivalence, reliability, usability, subgroup fairness, and documentation. Simple progress indicators and neutral feedback are treated as lower-risk engagement supports, whereas recognition prompts, scoring incentives, time pressure, and device-specific input demands are treated as modifications requiring targeted validation. The implemented prototype was evaluated in an initial formative usability study with ten participants. The System Usability Scale was administered separately for the Mini-Cog-derived, Trail Making Test-derived, and Stroop-derived modules, producing mean SUS scores of 95.0, 95.0, and 89.3, respectively. Qualitative feedback indicated that the game-like design was perceived as clear, enjoyable, and motivating, while also identifying task-specific refinements related to instruction clarity, error recovery, feedback, visual accessibility, and colour perception. The findings support the usability and engagement potential of the prototype, but they should be interpreted as formative evidence rather than evidence of diagnostic validity or clinical effectiveness. Future work should refine and version-lock the prototype, compare it with conventional or low-gamification reference measures, and evaluate it with intended clinical or screening-support populations

ABSTRACT. Intraoperative neurophysiological monitoring (IONM) signals are used to assess neural irritation during surgery, yet their underlying structure remains difficult to characterize due to substantial variability and class overlap. This study investigates the latent structure of EMG-based IONM signals and proposes a geometry-aware framework that enables both discrete classification and continuous ordinal estimation of irritation severity in an interpretable and clinically transparent manner. To capture transitional dynamics, signals are segmented using sliding windows with labels assigned by the terminal timestamp (late-anchor labeling), inducing boundary-aware label blending. High-dimensional features are then projected into a two-dimensional latent space using Linear Discriminant Analysis (LDA), preserving class separability while organizing samples along a continuous ordinal manifold. A smooth manifold is fitted through class centroids, and samples are projected onto it to obtain a continuous severity coordinate, enabling unified discrete and continuous inference. Analysis reveals substantial overlap between the baseline and weak irritation classes, suggesting these labels reflect a shared low-severity physiological regime rather than distinct states. Classification using classconditional Gaussian modeling with maximum likelihood estimation achieves consistent performance, with errors predominantly occurring between adjacent severity levels, supporting the consistency of the learned ordinal structure.

ABSTRACT. In safety-critical environments, human performance depends on the interplay be-tween cognitive demands and physiological regulation. Heart rate variability (HRV) is widely used to assess autonomic activity, yet its relationship with perfor-mance remains inconsistent, partly due to linear models that neglect underlying physiological states. This study proposes a distribution-based framework that dis-cretizes autonomic activation into subject-specific regimes and assesses HRV-performance relationships within each regime. Fifty-six participants performed a dual-task paradigm (Tracking and System Mon-itoring) on the AeroStim platform at four ranges of workload levels. ECG signals were continuously recorded; HRV metrics (SDNN, RMSSD, LF/HF) were extracted and normalized within subjects. SDNN defined three activation ranges (low, inter-mediate, and high), and nonparametric analyses assessed performance and HRV-performance associations within each regime. Performance decreased with increasing task difficulty. Critically, HRV-performance relationships varied across ranges: SDNN was primarily associated with performance in low and high activation, RMSSD predicted performance in in-termediate activation, and LF/HF showed complementary associations in low acti-vation. Intermediate activation was linked to higher accuracy and fewer omissions than low activation, supporting a non-linear arousal–performance relationship. These findings demonstrate that HRV-performance associations are regime-dependent, supporting adaptive closed-loop systems to maintain optimal op-erator performance in real time.

ABSTRACT. This study focuses on a feature extraction algorithm based on the analysis of dynamic spectral components of magnetoencephalography (MEG) and electroencephalography (EEG) data. During its activity, the brain produces electromagnetic fields that can be sensed outside the head by means of appropriate sensors (e.g., SQUIDs in the case of Magnetoencephalography and electrodes in the case of Electroencephalography). These recordings are commonly acquired and processed to obtain estimates of brain activity with high spatial and temporal resolution. When an individual performs a task, brain activity exhibits task-dependent variations in the measured signals. Consequently, in recent years, the scientific community has increasingly focused on the study of brain signals to detect and classify features associated with specific tasks. Within this framework, we propose a novel paradigm whose core consists of a synthetic feature extraction approach based on the time-varying contribution of brain spectral components. These features are subsequently used as inputs to a machine learning classifier for task discrimination. The proposed strategy is designed to provide a novel and lightweight solution for human--machine interaction pipelines.

ABSTRACT. Visual comfort in indoor spaces, particularly those intended for study and work, represents an important aspect in the design of lighting and shading systems and their potential automation. Glare occurs as an undesirable effect of illumination, especially daylight. To mitigate this phenomenon in the design of both new and existing projects, it is essential to have reliable equipment for glare measurement. This paper aims to analyze methods for glare assessment and to develop a costeffective measuring device. Glare is evaluated primarily using the robust probabilistic metric Daylight Glare Probability (DGP), for which extensive support exists within the RADIANCE software package, particularly through the evalglare function. The device is based on a Raspberry Pi microcontroller and a wide-angle Waveshare camera with an OV5647 sensor. The outcome of this work is a functional prototype for measuring daylight glare in indoor environments. Its functionality was experimentally verified under real indoor conditions in both residential and university environments, and potential limitations of the implementation were identified along with proposals for further development of the project.

ABSTRACT. As healthcare environments transition toward hyper-connected digital ecosystems, the role of the Clinical Engineering (CE) professional is undergoing a radical shift. No longer confined to traditional preventative maintenance, today’s CE workforce must navigate the intersection of medical device interoperability, cybersecurity, and data-driven life cycle management. This presentation addresses the critical need for a modernized educational framework within Clinical Engineering departments. We will explore a dual-stream approach to workforce development:Technical Upskilling: Integrating IT fundamentals, network security protocols, and AI-driven diagnostics into standard CE training. Strategic Competencies: Developing "soft" skills such as cross-departmental communication, vendor negotiation, and clinical workflow integration.Participants will gain insights into establishing internal mentorship programs, leveraging manufacturer-agnostic training resources, and creating "career ladders" that incentivize continuous learning. By moving from a reactive "fix-it" mindset to a proactive technology management strategy, CE departments can reduce device downtime, enhance patient safety, and secure their seat at the hospital’s strategic leadership table.

ABSTRACT. This study evaluates the economic justification for introducing rotavirus vaccination into the national immunization program in Bosnia and Herzegovina (B&H). A pharmacoeconomic framework was applied, including Cost-Utility Analysis (CUA), Cost-Benefit Analysis (CBA), and Budget Impact Analysis (BIA), to compare a universal vaccination strategy with the current no-vaccination approach. The findings demonstrate that vaccination represents a dominant strategy, simultaneously improving health outcomes measured in Quality-Adjusted Life Years (QALYs) and reducing overall healthcare expenditures. The calculated Benefit-Cost Ratio (BCR) of 3.14 indicates substantial economic return relative to investment, supporting the integration of rotavirus immunization into the publicly funded healthcare system.

ABSTRACT. Sodium-glucose cotransporter 2 inhibitors (SGLT2i) are the most recently approved class of drugs (since 2012) for type 2 diabetes mellitus (T2DM), but their economic merits have yet to be fully confirmed. Using Agency for Identification Documents, Registers and Data Exchange of Bosnia and Herzegovina (IDDEEA) data, demographic and population data from the registry, and the available budget in Bosnia and Herzegovina, we analyzed whether to switch from sulfonylurea (SU) derivatives to sodium-glucose cotransporters-2 (SGLT2) inhibitors is economically justified. We constructed a decision-tree model spanning a 15-year period, using a 3% discount rate, to compare SGLT2 inhibitors to sulfonylureas (SU). Analyses included Cost-effectiveness, Cost-utility, Cost-minimization, Cost-benefit and Budget impact analysis and were informed by clinical probabilities, utility weights, and the costs of treatment and hospitalization. The results showed that SGLT2 therapy, while offering a marginal improvement in Quality-adjusted life years (0.0402 QALYs), came at a significantly higher lifetime cost (27,725 BAM vs. 6,000 BAM for SU). The calculated Incremental cost-effectiveness ratio (ICER) of approximately 540,000 BAM per QALY is well above the willingness-to-pay threshold in Bosnia and Herzegovina of 30,000 BAM/QALY. The analysis predicted a significant annual increase in spending, ranging from 72.3 million BAM to 216.8 million BAM, depending on the percentage of patients who would switch therapy. For the general patient population, the switch from SU to SGLT2 is not economically justified. However, due to the benefits of SGLT2 therapy, the switch is certainly justified in high-risk subgroups.

ABSTRACT. The rapid advancement of artificial intelligence (AI), particularly the widespread adoption of generative models such as ChatGPT, is reshaping the field of biomedical engineering. This transformation extends beyond technological innovation and is redefining professional roles, required competencies, and educational practices. This paper examines the transition from a predominantly technical profile of biomedical engineers toward a socio-technical role. Drawing on recent literature and industry perspectives, it argues that traditional technical expertise alone is no longer sufficient and must be complemented by communication skills, critical thinking, and interdisciplinary collaboration. Particular attention is given to emerging educational challenges associated with AI integration, including the development of data literacy, ethical awareness, and responsible AI use. The paper highlights the non-neutral nature of data, emphasizing structural biases—such as gender bias in clinical datasets—that may lead to unequal healthcare outcomes and raise critical questions of accountability in AI-driven decision-making. These challenges are further compounded by the risk of deskilling resulting from overreliance on AI systems. The paper aligns with recent studies emphasizing the need for pedagogical strategies that prepare students for responsible AI use, including academic integrity and transparency.

ABSTRACT. Bridging the gap between engineering education and real-world clinical research remains a persistent challenge in biomedical engineering. Here, we examine the practical competencies that academic curricula should cultivate to prepare graduates for translational research, drawing on direct experience across several methodological domains. A central difficulty for early-career engineers is the complexity of clinical data, which are frequently affected by missing values, artifacts, limited sample sizes and population heterogeneity. Effective training should therefore encompass not only analytical methods for handling such data, but also preventive strategies embedded in rigorous study design. Required competencies span robust statistical inference, data processing, programming, visualization, and the critical use of machine learning and deep learning, whose clinical application is constrained by issues of overfitting, interpretability and ethics. Open-source computational environments such as Python and R further support reproducible analytical workflows. Beyond data analysis, physiological numerical modeling enables virtual experimentation and hypothesis generation when direct in vivo measurement is ethically or technically infeasible. For physical device evaluation, hybrid numerical-physical simulators provide a reproducible and ethically preferable alternative to early-stage animal or human testing. Together, these domains define a curriculum oriented not toward narrow specialization, but toward the formation of biomedical engineers capable of integrating methods, disciplines and clinical perspectives in translational research.

ABSTRACT. This paper presents an interpretability-focused study on the effect of sensor placement on human activity recognition (HAR) using wrist-worn and hip-worn wearable sensors. By performing classification of various activities of daily living (walking, running and sitting) and simulated falling activity on a dataset containing accelerometer, gyroscope and barometric altimeter data, this study focuses on the comparison only wrist-worn part of the dataset versus the hip-worn part in order to establish whether there are any statistically significant differences in model performance. The approach consists of using Random Forest classifier for classification, and Leave-One-Subject-Out cross-validation. Model behaviour was further analysed using SHapley Additive exPlanations (SHAP) for global and local feature attribution and t-distributed Stochastic Neighbor Embedding (t-SNE) for feature space analysis. Wilcoxon signed-rank test found no statistically significant difference between placements, a result attributed to low statistical power given the small number of activity classes rather than an absence of a meaningful effect. Hip-worn data proved to be more suitable for the performance of all three models, with significant macro F1 gaps for accelerometer and gyroscope models respectively. Barometric altitude exhibited the lowest drop-off in model performance and the lowest t-SNE centroid distances generally, proving its robustness regarding sensor placement.

ABSTRACT. Biomedical engineering education varies significantly across countries and institutions. Each system believes it is preparing graduates for real-world demands. However, the rapid transformation of technologies creates new challenges in biomedical engineering. The constantly changing landscape requires graduates who can quickly adapt to new tools, learn unfamiliar systems, and yet remain careful, as every decision lands on a human being. Adaptability is not a soft skill. It is a technical necessity. This paper draws on my personal journey as an Erasmus Mundus scholar within the EMMBIOME program — a two-year master's conducted across the University of Patras (Greece), the University of Kragujevac (Serbia), and the University of Medicine and Pharmacy "Grigore T. Popa" (Romania). The aim is to identify how international, interdisciplinary education addresses competency gaps and prepares graduates for the rapidly transforming biomedical engineering landscape. I will reflect on how learning and working across borders reshaped my understanding of biomedical engineering — not only as a discipline, but as a responsibility. Rather than presenting a single best model, this paper will explore how combining different educational environments can better prepare future biomedical engineers. Beyond formal coursework, it will present the importance of student-led initiatives in biomedical engineering, and the crucial lessons that emerge when building them in unfamiliar countries, cultures, and systems. This paper offers concrete recommendations for curriculum development, drawn from experience inside the system — arguing that the goal of education is not only technical preparation, but building in graduates the steadiness, care, and capacity to keep learning that a constantly changing field demands.

ABSTRACT. The rapid advancement of artificial intelligence (AI) is transforming the role of biomedical engineers, requiring a shift from traditional technical expertise toward interdisciplinary, AI-augmented capabilities. Positioned at the intersection of human physiology and advanced technology, biomedical engineers must address persistent challenges, including workforce skill gaps, increasing regulatory demands such as compliance with ISO 13485 and HIPAA, as well as technical limitations related to data bias, limited datasets, and lack of model interpretability. Beyond algorithm development, modern biomedical engineers are expected to integrate sensing technologies, signal processing, and AI-driven decision support into deployable healthcare systems while addressing constraints such as biocompatibility, power consumption, and device miniaturization. In this context, core competencies extend to analytical thinking, interdisciplinary collaboration, creativity, and ethical responsibility to ensure patient-centered innovation. Ultimately, the future workforce will be defined by professionals who can bridge biological complexity and technological advancement through continuous learning and system-level integration, enabling reliable, explainable, and clinically meaningful AI applications in healthcare.

ABSTRACT. Early detection of Diabetic Peripheral Neuropathy (DPN) is hindered by the subjectivity of psychophysical screening tests such as the 10 gram-force Semmes-Weinstein monofilament (SWM-10g), whose accuracy depends on the patient’s perceptual judgment and motor response. Building upon a previously validated robotic platform that standardized SWM-10g application across plantar regions in healthy volunteers, we propose a design framework for an objective sensory assessment system that replaces the behavioral response with a neurophysiological marker. The proposed architecture integrates the existing robotic stage with synchronized electroencephalography (EEG) and deferred Artificial Intelligence (AI) analysis to correlate controlled mechanical stimuli with Somatosensory Evoked Potentials (SSEPs). Dedicated hardware (32-bit microcontroller and isolated front-end) supports sub-millisecond synchronization between plantar pressure and cortical activity. A hybrid CNN-RNN pipeline is outlined to identify the cortical P40 component and a data-driven “neural inflection point” that operationally defines an Objective Perception Threshold (OPT). The framework is presented as a hypothesis-generating roadmap; clinical validation against Nerve Conduction Studies in diabetic cohorts is outlined as future work toward a neurophysiological biomarker for diabetic foot risk prevention.

ABSTRACT. The interventionwill introduce participants to the core digital foundations behind the OneAquaHealth project. Building on the project’s mission to strengthen urban freshwater ecosystem health and human wellbeing, the session will explore how the deployment of the One Digital Health (ODH) framework - the so-called Steering Wheel - is meant to highlight the "digital balance" lying under the OAH project. This becomes possible via the definition of FAIR‑aligned digital tools, within the greater scenario of the ODH Intervention, with the objective of enabling transparent, connected, and actionable insights for researchers, policymakers, and communities working toward resilient aquatic ecosystems.

ABSTRACT. Shared decision-making in breast cancer locoregional treatment remains challenging due to the need to balance oncological safety with aesthetic outcomes, quality of life, and patient preferences. While artificial intelligence (AI) has shown considerable promise in personalized medicine, most available applications in oncology lack prospective clinical validation and real-world implementation. This presentation will describe the development and clinical validation of the CINDERELLA approach, a Horizon Europe-funded initiative designed to support shared decision-making in women undergoing breast cancer locoregional treatment. The CINDERELLA platform combines AI-based prediction of aesthetic outcomes with patient-centred educational content and digital support tools to facilitate informed treatment decisions. The presentation will summarize the rationale, methodology, and implementation of a large prospective multicentre study evaluating the clinical impact of the intervention, including patient engagement, expectation–outcome matching, patient-reported outcomes, and acceptability across diverse healthcare settings. Particular emphasis will be placed on the challenges of integrating explainable AI into routine clinical workflows, ethical and regulatory considerations, and the role of digital innovation in empowering patients. The CINDERELLA experience offers an example of how AI can move beyond proof-of-concept toward clinically validated and scalable patient-centred cancer care.

ABSTRACT. App usage data can provide valuable insights into the information needs of breast cancer patients. By analysing which topics are accessed most frequently, how much time is spent on different sections, and which content is revisited, it is possible to identify the aspects of the disease, treatment, and surgical options that generate the greatest interest or concern. However, registered usage times frequently do not reflect true engagement. For example, a user who leaves a reconstruction-information page open for 30 minutes while doing something else may appear more interested in that topic than a user who spends 10 minutes actively reading and interacting with the content. Thus, observed usage times may arise from two latent processes: active and inactive periods. Finite mixture models offer a robust framework for addressing classification challenges when the underlying process is not fully observable. We propose a two-component finite mixture model to classify session times into real (active sessions) or inflated (unattended sessions) categories, exploring parametric specifications including log-normal and Gamma distributions. Covariates capturing socio-demographic factors, clinical characteristics, and user–application interactions were incorporated using mixtures of generalized linear models. Model estimation was conducted using EM (Maximum Likelihood framework) and NUTS (Bayesian framework) algorithms. Data analysis results indicate that the proposed models effectively distinguish real and inflated usage times, with covariate inclusion improving model fit and classification flexibility. Frequentist and Bayesian approaches yield comparable results for well-specified models. The study highlights the heterogeneity of user behaviour and the importance of accounting for covariates and latent outliers, offering a robust methodology for analysing digital application usage.

ABSTRACT. Despite growing interest in artificial intelligence (AI) in healthcare, most AI-enabled solutions remain at the early validation stage and do not achieve sustained integration into routine clinical practice. A key barrier is the limited understanding of the organizational and contextual factors that influence successful implementation. The ongoing CINDERELLA project responds to this gap by assessing an innovative AI-based APProach intended to enhance shared decision-making and manage patient expectations in breast cancer locoregional treatment through a multidimensional assessment framework encompassing economic, financial, organizational, and implementation dimensions. The organizational impact of the CINDERELLA APProach and its acceptability among healthcare professionals (HCPs) were assessed through focus groups conducted with HCPs involved in the breast cancer care pathway across five clinical centers participating in the CINDERELLA trial. An initial focus group at each site examined the existing organizational model, while follow-up focus groups at 12 and 24 months explored organizational changes associated with implementation and gathered perceptions regarding the acceptability of the intervention. All discussions were audio-recorded, transcribed verbatim, and analyzed using a hybrid inductive–deductive thematic approach. Overall, the CINDERELLA APProach was perceived as acceptable and potentially valuable for enhancing patient engagement, supporting shared decision-making, and strengthening interprofessional collaboration in breast cancer care. Although it was generally integrated into existing clinical workflows without major disruption, its implementation required varying degrees of organizational adaptation across participating centers, shaped by context-specific conditions. These contextual variations are critical not only for interpreting trial outcomes but also for informing future scale-up and adoption across diverse healthcare settings. The talk will also provide an opportunity to discuss current challenges in the assessment and implementation of AI-based health technologies, and to identify priorities for advancing their adoption into clinical practice, drawing on lessons learnt from the CINDERELLA project.

ABSTRACT. Responsible innovation in healthcare demands professionals equipped to navigate highly complex technological, regulatory, and environmental landscapes. While advancements like AI-driven diagnostics, digital health ecosystems, and Net Zero targets are redefining medical device lifecycles, traditional Biomedical Engineering (BME) curricula often treat ethics as a compartmentalised compliance checkbox, leaving emerging competencies like sustainability literacy and regulatory strategy on the periphery. This interactive 90-minute workshop shifts the educational paradigm from passive listening to collaborative action, inviting educators, researchers, and practitioners to co-create a modernised BME competency framework.

Moving systematically through four structured phases, participants will work in mixed-career groups to execute a hands-on curriculum reform exercise. The session begins with a rapid gap analysis to isolate the primary deficiencies in current engineering education regarding AI ethics, EU Medical Device Regulation (MDR 2017/745) compliance, and frugal engineering. Groups will then utilize a specialized 2D framework to map these challenges across two distinct axes, namely Technical Complexity (Basic to Advanced) and Ethical Responsibility (Awareness to Active Implementation), to pinpoint where critical learning gaps exist.

In the final phases, teams will select a high-priority competency from their matrix and build a concrete implementation strategy to integrate it into a standard 3-year Bachelor’s program, defending their blueprint in a series of rapid-fire, 1-minute peer pitches. The workshop concludes with a live digital poll to benchmark the systemic barriers to curriculum reform across the participants' global home institutions. Sponsored by the EAMBES and IFMBE Ethics and Education Committees, this session transitions dialogue into a tangible roadmap, equipping the MEDICON 2026 community with the educational tools necessary to train the next generation of socially accountable and future-ready biomedical engineers.

ABSTRACT. This study aims to evaluate the reliability and quality performance of in-service Magnetic Resonance Imaging (MRI) systems in China and to analyze their fault modes through multi-center data analysis. Fault data from 333 MRI devices across 53 hospitals nationwide were collected and analyzed using indicators such as Mean Time Between Failures (MTBF), annual failure rate, and Failure Modes, Effects, and Criticality Analysis (FMECA). The results show that radiofrequency coils, software defects, and water-cooling systems are common high-incidence fault areas across the industry. In addition, equipment reliability decreases significantly as service years increase. MRI systems that have been in use for more than seven years generally enter an aging phase of degradation, with failure rates increasing year by year. At this stage, enhanced maintenance and servicing are required for MRI equipment.

ABSTRACT. Medical equipment aging places growing pressure on healthcare systems that must balance safety, performance, and constrained budgets. Most hospitals still schedule maintenance and replacement by calendar age, an approach that ignores the real deterioration and variability of device functional life. This talk shows how AI can support a sustainable and data-driven view of the medical equipment lifecycle. Deep learning and natural language processing can turn maintenance records, adverse event databases, and operational logs into reliability indicators that reflect actual condition and usage. These models estimate performance degradation, predict failure risk, and produce age-informed metrics that help clinical engineers prioritise interventions and plan capital replacement. By linking such indicators to lifecycle governance, hospitals can move beyond fixed thresholds toward decisions grounded in evidence. Semantic ontologies and evidence-based maintenance are also addressed in structuring heterogeneous data across device fleets, which improves comparability and supports transparent reasoning. Attention is given to the regulatory context and to the need for explainable outputs that managers can trust. AI can extend the judgement of clinical engineering teams. Used well, it reduces waste, lengthens the safe service of reliable equipment, and frees resources in constrained settings.

ABSTRACT. Analyses of maintenance and management data collected from ~342,000 pieces of equipment in 30+ years show certain equipment types exhibit clear impact of aging (CIA)—i.e., increasing number of repairs with age, while others don’t (no impact of aging – NIA), despite consistent scheduled maintenance for all types. In other words, the demand of corrective maintenance resources increases with age for CIA types but remain mostly constant for NIA types, suggesting maintenance planning need to consider aging behavior. Likewise, equipment replacement needs to be planned by type for CIA types, while replacement for NIA types can be made on a contingency basis. A simple, quantitative model has been created and tested for characterizing medical equipment aging behavior, thus eliminating the need for analyzing large amounts of historical data. This aging model enables better maintenance and replacement planning, as well as providing better life expectancy estimates than those offered by the American Hospital Association and ECRI.

ABSTRACT. Medical equipment replacement decisions in resource-constrained healthcare systems are typically governed by calendar-age thresholds rather than by empirical assessment of functional state and operational risk. This paper proposes a data-driven framework that transforms routinely collected maintenance data into a structured Replacement Prioritization Score (RPS) for aging medical devices. The framework integrates four operational indicators — Maintenance Frequency Rate, Maintenance Cost Index, Repair-to-Replace Cost Ratio, and Device Operational Availability — into a composite Degradation Urgency Index (DUI), which is further weighted by clinical acuity and normalized device age to produce a device-level prioritization score. An illustrative case study applied to five representative device types in a low- and middle-income country (LMIC) hospital context demonstrates how the framework distinguishes operationally degraded devices from well-functioning aging equipment, enabling more accurate capital investment decisions. The framework is designed to be implementable using data routinely available in Computerized Maintenance Management Systems (CMMS) and budget records, without requiring additional instrumentation or specialized data infrastructure.