ABSTRACT. Thermal imaging represents a non-invasive and contactless method for vascular assessment; however, vessel segmentation remains challenging due to low contrast, noise and motion artifacts. This work presents a comparative study of two temporal propagation strategies for vessel segmentation using thermal videos acquired with a FLIR A655 infrared camera and stored in proprietary .seq format. Both approaches are developed within a common spatio-temporal framework that integrates ROI-based preprocessing, data-driven temporal smoothing with SNR-based window optimization, arm-background separation via Gaussian Mixture Modeling, adaptive border refinement, and region growing guided by vessel statistics (µ, σ). The strategies differ in temporal propagation: a static approach based on centroid tracking, suitable for quasi-stationary conditions, and a motion-adaptive approach combining local optical flow estimation, thermal-based seed refinement, and motion-dependent search regions. Performance is evaluated against manual ground truth using region-based, boundary, and structural metrics. Under limited motion both strategies yield comparable accuracy (Dice 0.799 vs. 0.818; IoU 0.667 vs. 0.694), while the adaptive strategy markedly improves boundary fidelity in dynamic conditions, reducing mean Hausdorff distance from 16.4 px to 9.2 px (∼44%).

ABSTRACT. Active noise cancellation (ANC) may reduce listening effort in noisy environments, but its physiological effect during cognitive tasks remains unclear. This study assessed whether ANC modulates pupil-based markers of cognitive load during working memory performance. Twenty participants completed 0-back and 2-back tasks under ANC OFF and ANC ON conditions while pupil dynamics were recorded using Pupil Labs Invisible glasses. Pupil area, major axis and minor axis were analysed with generalised additive mixed models. Cognitive load increased pupil size in both ANC conditions, confirming higher physiological demand during the 2-back task. However, the load-related increase was larger and more sustained when ANC was OFF. ANC had limited effects during the 0-back task, but reduced pupil size across extended intervals during the 2-back task. These findings suggest that ANC reduces the additional physiological cost imposed by background noise when working memory demand is high. Pupillometry may therefore provide an objective method for evaluating wearable acoustic technologies and their effect on cognitive effort.

ABSTRACT. This study introduces a novel framework for estimating biological age in pediatric populations (ages 2–17) using two minutes of resting-state electroencephalography (EEG) recordings. From a cohort of 190 healthy subjects, 10 spectral and 4 non-linear EEG features were extracted to characterize neurodevelopment dynamics. For age prediction, ElasticNet and Support Vector Regression models were evaluated using a repeated nested cross-validation framework, comparing the standard 10-20 electrode configuration against a reduced 10-electrode subset optimized for clinical feasibility. All models demonstrated high predictive accuracy, with Mean Absolute Errors (MAE) ranging from 1.55 to 1.62 years. The 10-electrode ElasticNet configuration achieved the best performance (MAE = 1.55 ± 0.03 years), suggesting that reliable maturation tracking is possible even with reduced electrode setups. Feature importance analysis identified spectral-profile features (e.g., delta and alpha-frequency power) as primary drivers of the models' decision making, aligning with established neurodevelopment patterns. The best-performing model was then applied to a small cohort of five pathological subjects, where bias-adjusted Brain Age Gap values enabled the identification of subject-specific deviations from the normative developmental trajectory. Although preliminary, the present work underscores the potential of EEG-based biological age estimation as a scalable tool for monitoring neurodevelopment and detecting early signs of neurological divergence.

ABSTRACT. Epilepsy is a neurological disorder characterized by excessive neuronal activity. Accurate identification of epileptogenic zones is essen- tial for presurgical evaluation in drug-resistant focal epilepsy. In this study, we propose a framework that combines brain connectivity and graph-theoretical analysis to investigate the pre-ictal period for epilep- togenic zone identification. Thirteen patients with drug-resistant focal epilepsy were retrospectively analyzed using MRI and long-term scalp EEG recordings. Cortical sources were reconstructed on MRI-derived surfaces, and temporal Granger causal- ity was computed during pre-ictal and ictal phases. Functional connec- tivity networks were then analyzed using graph-theoretical measures to estimate centrality and hubness indices for seizure lateralization and lo- calization. Six patients underwent surgical treatment, and five achieved favorable outcomes. In the patient with an unfavorable outcome, the analysis sug- gested a bilateral epileptic network. Among the patients with favorable outcomes, both lateralization and localization were correctly identified in three cases, and lateralization alone in one. The remaining case was affected by motion artifacts. Among the three patients monitored with stereo-EEG, lateralization was correctly identified in all cases, and localization in one. In the remaining four patients who underwent neither surgery nor stereo-EEG, lateraliza- tion and localization agreed with the qualitative clinical prediction.

ABSTRACT. This study presents a functional prototype for autonomous post-market surveillance of medical devices, bridging the execution gap between theoretical AI frameworks and live hardware orchestration. Using the OpenClaw multi-agent system powered by Gemma 4 language model and Exa for real-time regulatory intelligence, we developed a proactive technical loop that interfaces a whitelisted WhatsApp control layer with digital twin telemetry. To ensure deterministic device identification and minimize AI hallucinations, we implemented a declarative metadata layer for autonomous agent discovery. The system's reliability was validated through forced-stress simulations where the orchestrator autonomously refactored device logic to trigger and report critical thermal anomalies in compliance with ISO 13485:2016. Results demonstrate 100\% discovery accuracy and successful closed-loop alerting, proving that generative AI integrated with deterministic safety gates can effectively automate the link between hardware drift and mandatory vigilance reporting. This architecture provides a scalable, human-in-the-loop framework that maintains strict adherence to the safety mandates of the EU AI Act.

ABSTRACT. Breast cancer remains one of the leading causes of cancerrelated mortality among women worldwide, underscoring the importance of early and accurate detection to improve prognosis and survival rates. Conventional imaging modalities, including X-ray mammography, ultrasound (US), computed tomography (CT), and magnetic resonance imaging (MRI), are widely used in clinical practice but each suffers from intrinsic limitations in terms of sensitivity, specificity, spatial resolution, cost, or patient comfort. In recent years, microwave tomographic imaging (MWTI) has emerged as a promising complementary modality due to its non-ionizing nature, low cost, and intrinsic sensitivity to the dielectric contrast between healthy and malignant tissues. In this work, we propose a deep-learning-based multimodal imaging framework that jointly exploits microwave and ultrasound measurements for breast cancer detection. The proposed approach integrates scattered electric and acoustic field data within a unified neural architecture to directly produce a pixel-wise classification map of breast tissues. By leveraging the complementary physical information provided by electromagnetic and acoustic wave propagation, the method aims to enhance tumor localization without relying on intermediate image reconstruction steps. The framework is validated on a synthetic dataset of numerical breast phantoms, showing encouraging performance in terms of structural similarity and classification accuracy.

ABSTRACT. Magnetocardiography (MCG) signals, which are highly susceptible to environmental noise during acquisition, which may result in acquisition failure. Meanwhile, due to the lack of publicly available universal datasets as evaluation criteria, it is highly necessary to develop a method capable of online automatic quality assessment to improve the efficiency and quality of clinical diagnosis based on MCG. This study proposes an intelligent MCG signal quality evaluation method integrating time-frequency analysis and deep learning. Raw MCG signals undergo preprocessing, including noise reduction via signal averaging and R-wave-based normalization to ensure consistent segment length. Continuous wavelet transform (CWT) is then applied to generate two-dimensional time-frequency representations. A novel InceptionTime-CBAM network, enhanced with convolutional block attention modules, is developed to classify signal quality as acceptable or unacceptable. Experimental results, based on a clinically collected dataset of 1625 samples, demonstrate that the proposed method achieves an accuracy of 94.36%, sensitivity of 97.46%, specificity of 89.17%, and an F1-score of 95.57% in five-fold cross-validation, significantly outperforming classical CNN models such as ResNet and VGGNet. The framework provides a robust and automated solution for MCG quality assessment, facilitating more reliable cardiac disease diagnosis and supporting the development of clinical tools based on MCG analysis.

ABSTRACT. Paroxysmal atrial fibrillation (PAF) may be preceded by subtle changes in rhythm dynamics that are difficult to characterize using conventional heart-rate variability (HRV) descriptors alone. This study proposes a multiscale statistical process control (SPC) framework for explainable pre-onset characterization of PAF from RR interval dynamics. The analysis used the PhysioNet PAF Prediction Challenge Database learning set, restricted to the main 30-min p-records. For each consecutive odd-even p-record pair, the odd record was labeled as PAF-distant and the even record as pre-onset. Continuation records ending in c and n-records from subjects without documented atrial fibrillation were excluded from the primary paired task. RR intervals were derived from QRS annotations and analyzed using RR-count windows, 5-min windows, and full-record aggregation. Interpretable HRV and SPC features were extracted using Shewhart, CUSUM, EWMA, moving-range, and runs-rule descriptors. Grouped validation by pair identifier and 500 pair-level bootstrap iterations were used to reduce leakage and quantify uncertainty. SPC configurations improved over the evaluated reduced and expanded time-domain/Poincaré HRV descriptor panels, with B_SPC reaching AUC = 0.776 (95% CI: 0.704–0.862) and pairwise concordance = 0.840 (95% CI: 0.733–0.941). Shewhart out-of-control rate and count showed the strongest paired pre-onset increases. These findings support SPC as an interpretable feature-engineering layer for ECG rhythm characterization, requiring external validation before clinical deployment.

ABSTRACT. IVF has become a standard fertility treatment. Deep learning has the potential to support automated embryo assessment from imaging data; however, most methods rely on single-plane or single time-point images. This study aims to develop a deep learning-based tool to support the workflow by predicting whether an individual mature oocyte will give rise to an embryo, using early multi-plane time-lapse imaging acquired in an intracytoplasmic sperm injection (ICSI) setting. Time-lapse stack images (11 focal planes) were acquired every 10 minutes and assembled into fixed-length sequences of 10 time points per oocyte. A focal-plane fusion module compressed each stack into a 3-channel frame, which was embedded by a masked-autoencoder (MAE) pre-trained Vision Transformer. Frame embeddings were modeled with a bidirectional long short-term memory (LSTM) network and classified via a sigmoid output. The developed deep learning model was evaluated on a held-out test set of 67 oocytes. It correctly identified 11 of 12 embryo-forming oocytes (sensitivity: 0.917), critical to ensure potential embryos are not overlooked during IVF procedures. For non-embryo oocytes, 34 of 55 were correctly classified (specificity: 0.618). Although specificity is lower, the model still effectively reduces the number of oocytes to track without discarding potentially useful material.

ABSTRACT. Digital twins (DTs) are increasingly recognized as key tools for improving Type 1 Diabetes (T1D) management, enabling patient-specific simulation of glucose dynamics and the development of personalized therapies. Central to the creation of an accurate DT is the twinning procedure, which is the identification of patient-specific parameters from continuous glucose monitoring, carbohydrate and insulin information.

Physics-Informed Neural Networks (PINNs) are an emerging paradigm for this task: by integrating a physiological model (described by ordinary differential equations, ODEs) into the training process, a PINN can learn a continuous approximation of the state trajectory and, in principle, can recover hidden physiological dynamics.

As a first step toward this goal, this paper aims to: (i) investigate whether a PINN can consistently identify three key patient-specific parameters: basal glucose (Gb), gastric emptying rate (kempt), insulin sensitivity (SI); and (ii) understand the role that different elements of the PINN loss function play in parameter identification.

To this end, we leveraged a simplified model of glucose-insulin dynamics and performed a Monte Carlo simulation study varying model parameters and exogenous inputs. Then, we challenged the PINN to recover the true parameters under three different configurations of its loss function.

The PINN in its complete configuration achieves a Mean-Absolute-Relative-Difference (MARD) of 0.95%, 3.85%, and 5.09% for Gb, kempt, and SI respectively, with root-mean-squared-error (RMSE) = 1.49mg/dL and R^2=99.3%. Removing the physics loss term causes kempt MARD to increase fivefold (to 24.3%) and SI MARD sixfold (to 32.7%), demonstrating that the ODE-based constraint in the training process is key for accurate parameter identification.

ABSTRACT. Automated detection of interictal epileptiform discharges (IEDs) in EEG is clinically relevant but remains methodologically challenging because of signal variability and sensitivity to preprocessing. The aim of this study was to investigate how preprocessing choices, particularly normalization strategies and band-pass filtering settings, affect the performance of deep learning models for automated detection of interictal epileptiform discharges in EEG signals. To this end, five normalization methods under both window-level and record-level schemes, combined with seven band-pass filtering configurations, were evaluated using the HTNet-EEG hybrid architecture on a publicly available 84-subject clinical EEG dataset. The best-performing pipeline, record-level robust median absolute deviation (MAD) normalization paired with a 5–100 Hz Butterworth band-pass filter, improved ROC-AUC from 0.92 to 0.98, PR-AUC from 0.31 to 0.84, and F1-score by +0.39 over the unoptimized baseline (p < 0.001, DeLong test). These findings suggest that preprocessing plays an important role in the automated identification of EEG patterns and may provide useful guidance for designing reliable and reproducible preprocessing pipelines for EEG-based IED detection systems.

ABSTRACT. The reliability of Infusion Pumps remains a critical concern in healthcare, as traditional reactive maintenance models often lead to abrupt shutdowns that can compromise patient safety. This study intro- duces an autonomous maintenance framework that utilizes a three-layer architecture to bridge the gap between physical hardware performance and high-level regulatory intelligence by integrating a digital twin simu- lation with an agentic orchestrator. By continuously monitoring device telemetry and cross-referencing performance data with safety standards such as ISO 13485 and IEC 62304, the orchestrator demonstrated the ability to execute graceful degradation through autonomous logic refactor- ing. Experimental results confirm that the system successfully maintains therapeutic delivery during mechanical drift, silent hardware failures, and power crises while ensuring absolute transparency through automated audit trails and clinician-facing decision logs. Ultimately, this research proves that transitioning from passive monitoring to active hardware resilience provides a viable pathway for meeting modern regulatory de- mands for high-risk medical AI, ensuring that devices are not only smart enough to report failures but resilient enough to survive them.

ABSTRACT. Background: Coronary CT angiography (CCTA) reconstruction methods are commonly evaluated through reader-based assessment and conventional image-quality metrics, although these approaches do not fully reflect how reconstructed images support downstream automated analysis. Objective: To propose and validate a reconstruction-aware CCTA evaluation framework that integrates coronary artery segmentation as a complementary criterion for reconstruction-method comparison. Methods: We analyzed a private paired-reconstruction dataset of 26 patients, in which each examination was reconstructed using three representative reconstruction categories: iterative reconstruction (IR, AIDR 3D), model-based iterative reconstruction (MBIR, FIRST), and deep learning reconstruction (DLR, AiCE). The framework combined: (i) whole-heart quantitative image-quality analysis based on masks obtained from an adapted whole-heart segmentation pipeline, (ii) coronary artery segmentation using a reproduced ImageCAS-based deep learning framework, evaluated both in inference-only conditions and after supervised fine-tuning on the manually annotated subset, (iii) coronary-specific quantitative metrics, and (iv) blinded qualitative radiological assessment. Results: Across the different evaluation levels, AiCE showed the most favorable performance overall, and AIDR 3D yielded the least favorable results in most comparisons. Importantly, coronary segmentation performance reflected the same ranking observed in quantitative image-quality metrics and blinded qualitative assessment. Conclusion: Coronary segmentability captures reconstruction-dependent differences in a meaningful and consistent way. These findings support the use of coronary segmentation not only as a downstream task, but also as a practical complementary criterion for reconstruction-method comparison in CCTA.

ABSTRACT. linical practice guidelines are essential for evidence-based care, but their length and complexity often limit their educational use in routine practice. In this study, we investigate whether large language models (LLMs) can automatically generate clinically meaningful educational questions from raw low back pain guidelines. We built a fully automatic pipeline combining PDF parsing, structure-aware segmentation, question generation, and multidimensional evaluation. We compared three chunk granularities and four LLM configurations, and introduced a novel Clinical Information Density (CID) as a measure of the concentration of clinically relevant information within each text segment. Results suggest that (i) smaller, semantically coherent chunks produced better questions than larger document-level inputs, (ii) information-dense chunks could improve the question generation, and (iii) both biomedical specialization and instruction tuning could improve LLMs' question generation from guideline text, recovering relevance even for less information-dense chunks.

ABSTRACT. This study investigates the temporal evolution of uterine electrophysiological activity within 48 hours following atosiban administration using electrohysterogram (EHG) signals. Three representative features were analyzed, including peak-to-peak amplitude (PPA), dominant frequency (DF), and multi-state Lempel–Ziv complexity (LZM). Generalized additive mixed models (GAMMs) were employed to characterize nonlinear temporal dynamics and to evaluate their modulation by gestational age (GA) and pregnancy outcome (preterm labor vs. term labor, PL vs. TL). The results demonstrate that frequency-domain and nonlinear features exhibit stronger GA-dependent modulation than amplitude-related measures. Regarding pregnancy outcome, the PL and TL groups exhibited distinct responses to atosiban therapy in PPA and LZM for GA ≥ 32 weeks, while for GA < 32 weeks, the difference was observed in DF. These findings indicate that the electrophysiological response to atosiban is GA stage-dependent and varies with pregnancy outcome. EHG-derived temporal dynamics may therefore serve as a non-invasive biomarker for assessing short-term tocolytic response and supporting outcome prediction in threatened preterm labor.

ABSTRACT. Precision medicine has long relied on bulk molecular profiling, an approach that can obscure the cellular heterogeneity, regulatory diversity, and tissue architecture that shape disease behavior and therapeutic response. Over the past decade, single-cell and spatial omics have reshaped this landscape by enabling the identification of rare and clinically relevant cell populations, the reconstruction of dynamic state transitions, the integration of multiple molecular layers, and the recovery of tissue context in situ. These advances have moved disease interpretation beyond descriptive profiling toward a mechanistic understanding of cell identity, regulatory programs, clonality, intercellular communication, and spatial organization. Here, we discuss how single-cell and spatial bioinformatics have transformed precision medicine by shifting analysis from bulk-average representations to cell state aware, network aware, and tissue context aware models of pathology. We argue that bioinformatics is not merely a downstream support, but the enabling layer that converts high-dimensional measurements into biologically interpretable and clinically actionable knowledge. From multimodal integration and trajectory inference to receptor aware immune profiling, regulatory network analysis, perturbational modeling, interactome reconstruction, and spatial computation, modern computational methods determine how these technologies can be translated into robust hypotheses, target prioritization, and patient relevant disease models. Finally, we highlight how emerging artificial intelligence frameworks may accelerate this transition by providing generalizable representations of cellular and tissue organization, opening the way to a more predictive and mechanistically grounded precision medicine.

ABSTRACT. Abstract: Optical brain monitoring offers a non-invasive approach for continuous assessment of cerebral physiology. Among these techniques, cerebral photoplethysmography (PPG) has gained interest due to its potential to estimate intracranial pressure. However, cerebral PPG signals may be affected by extracranial circulation and skin pigmentation, limiting their reliability. In this study, a multilayer pulsatile head phantom was developed and used to investigate the effects of extracerebral circulation and skin pigmentation on cerebral PPG signals. The phantom incorporated anatomically representative brain, skull, scalp, and interchangeable skin layers, and was integrated into a cardiovascular system enabling controlled perfusion of cerebral and extracranial layers. PPG signals were acquired at 770, 810, and 880 nm using a 3 cm source–detector separation. Results showed that PPG is sensitive to cerebral haemodynamic changes, as reductions in signal amplitude and signal-to-noise ratio (SNR) were observed during brain occlusion. However, extracranial circulation was found to dominate the pulsatile signal, with scalp occlusion reducing signal amplitude and quality. Increased skin pigmentation reduced amplitude and SNR, although pulsatile signals remained detectable across all conditions. These findings demonstrate that while PPG retains sensitivity to cerebral haemodynamics, measurements are influenced by extracerebral layers, highlighting the need for careful interpretation and potential correction strategies

ABSTRACT. Hypoxia in COVID-19 results from viral entry into host cells, leading to increased vasoconstriction, thrombosis, and lung inflammation, impairing oxygen metabolism. "Silent hypoxia" describes significantly reduced oxygen saturation without dyspnea. This phenomenon may be linked to ACE2 receptor overexpression, enhancing susceptibility to COVID-19 damage. Possible mechanisms for absent dyspnea include mild carbon dioxide elevations masking symptoms, direct viral effects on the brain and nervous system, and vascular changes. Silent hypoxia arises from a complex interplay of factors, including pneumonia, alveolar collapse, and impaired oxygen exchange, potentially causing indirect organ damage. Early detection is crucial to mitigate long-term effects and mortality. This review systematically examines the current literature on silent hypoxia in the context of COVID-19, based on a search of PubMed using the keywords "silent hypoxia" and "COVID-19" in February 2024. Findings remain inconclusive due to conflicting results on risk factors and comorbidities. Asymptomatic patients with severe hypoxia are at risk of rapid deterioration, underscoring the importance of improved diagnostics and early intervention. Current research is limited by small sample sizes and heterogeneous methodologies, necessitating larger cohorts, standardized study designs, and exploration of genetic predispositions and comorbidity interactions. Silent hypoxia highlights COVID-19’s unique pathophysiology, blending respiratory, vascular, and neurological impacts. Addressing this condition requires better tools for early detection and tailored interventions to prevent rapid progression and severe outcomes.

ABSTRACT. This study investigated whether angle signals and their derived features can effectively characterize unilateral standing long jump (SLJ) performance and assist in assessing trial validity. Twenty healthy participants were recorded performing jumps using inertial measurement units (IMUs). Using gyroscope data from the IMUs, the lower-body segment and joint angles were estimated and characterised using amplitude, temporal, and smoothness features. The extracted feature values were analysed with respect to gender and the leg used for execution, as well as their relevance to trial validity. The findings indicated that the leg used for jumping had a minimal impact on most features in valid trials, suggesting symmetric movement patterns among healthy individuals. Conversely, the features related to local extrema most frequently showed significant differences between female and male participants. In the validity assessment, the segment model with fewer parts exhibited the best overall fit, indicating that kinematic data at the segment level offers more meaningful insights into distinguishing valid from invalid trials.

ABSTRACT. Patient–ventilator asynchrony (PVA) is a clinically significant complication of non-invasive ventilation (NIV), occurring when ventilator-delivered breaths fail to align with a patient's spontaneous respiratory effort. Despite its high prevalence and association with increased patient discomfort, prolonged ICU stay, and elevated mortality risk, detection in clinical practice remains largely manual and highly operator-dependent. This study presents the development and preliminary validation of a simple, rule-based algorithm for the automated detection of the three most clinically important trigger asynchronies during NIV: double-triggering (DT), auto-triggering (AT), and ineffective effort (IE). Requiring only the airway pressure waveform and a measure of neural respiratory rate as inputs, the algorithm was validated against clinician-derived ground truth in identifying 52 asynchrony events using retrospective data from 13 NIV patients, achieving an overall sensitivity of 0.89 and specificity of 0.83. DT detection was perfect across all patients (sensitivity and specificity of 1.00), IE detection successfully identified all but one clinician-labelled events (sensitivity 0.97, specificity 0.6), and AT detection was weakest (sensitivity 0.50), attributed to the absence of a direct trigger signal. These results represent an encouraging proof of concept for a lightweight, automated PVA detection tool for NIV settings where continuous monitoring is currently lacking.

ABSTRACT. Skin hydration and oedema are governed by distinct physiological mechanisms; hydration is primarily influenced by local barrier properties and environmental or topical factors, whereas oedema is driven by systemic hemodynamic processes and lymphatic or renal handling. Consequently, changes in one parameter may occur independently of the other, making skin hydration a potential confounding factor in the detection of oedema using near-infrared spectroscopy. To investigate this interaction, Monte Carlo simulations were performed to model nine tissue conditions combining different hydration levels and oedema stages. The simulations evaluated light–tissue interactions, absorbance, and mean light penetration depth under two optical fibre configurations, perpendicular and parallel to the skin surface. Comparison of these geometries showed that fibre orientation affects both the magnitude of the optical response and its sensitivity to hydration changes. Although the parallel configuration remained capable of differentiating oedema stages across hydration levels, measurements at 1450 nm were particularly susceptible to hydration-related interference, with absorbance changes potentially mimicking or masking oedema signatures. Importantly, the absorbance ratio between 1200 nm and 1050 nm (∆A_(1200/1050)) was found to be a robust metric (R2 = 0.999, β1 = 0.188 ± 0.002, p < 0.001), demonstrating potential to distinguish oedema stages despite variations in skin hydration and fibre orientation.

ABSTRACT. Multiple sclerosis (MS) is a chronic neurodegenerative disease characterized by progressive motor impairment, in which gait dysfunction significantly contributes to disability and reduced quality of life. Although gait rehabilitation improves functional outcomes, the neural mechanisms underlying recovery remain poorly understood, particularly across different training modalities. In this context, identifying how rehabilitation modulates cortical haemodynamics is crucial to better understand recovery processes and optimize treatment strategies. Here, we investigate the cortical haemodynamic changes induced by different gait training exercises within the context of the PROGR-EX clinical study. Twenty-four MS patients underwent distinct gait interventions: robot-assisted gait training (RAGT) with progressively increasing walking speed at low intensity (low-RAGT), RAGT at constant walking speed (conventional-RAGT), or overground gait training (OGT). Functional near-infrared spectroscopy (fNIRS) signals were recorded during an upper-limb motor task before treatment (T0), at the end of treatment (T1), and at follow-up (T2). A cross-sectional comparison with healthy controls was also performed. At baseline, MS patients showed altered cortical lateralization in a frontal-dominant brain network, with increased ipsilateral and reduced contralateral oxygenation compared to healthy controls. Longitudinal analyses revealed intervention-specific effects. Low-RAGT promoted a shift toward a more lateralized pattern, with increased contralateral and reduced ipsilateral activity at T1, and sustained reduction of ipsilateral involvement at T2. OGT induced a reduction of ipsilateral activity at both T1 and T2, while conventional-RAGT showed no clear shift toward a physiological oxygenation pattern. Gait rehabilitation modulates cortical haemodynamics in an exercise-dependent manner, with low-RAGT appearing to be the most effective in promoting a shift toward physiological motor lateralization.

ABSTRACT. Traditional evaluation metrics, such as accuracy and AUC, provide only a global assessment of ML models and cannot indicate whether an individual prediction may be considered reliable. This limitation is particularly significant in critical domains, where incorrect predictions can have significant consequences. To address this issue, this work investigates pointwise reliability estimation, which aims to quantify the trustworthiness of each individual prediction. A comparative analysis is conducted using three representative reliability approaches: a density-based method (RCD), a local fit method (CCNR), and a knowledge-based rule-driven approach. The evaluation is performed through a case study based on a real-world clinical dataset, motivated by hospital resource management, and formulated as a regression task for length-of-stay prediction. Reliability scores are computed for each test instance in order to analyse the relationship between reliability levels and predictive performance. The results consistently demonstrate a stratification effect, where higher reliability intervals are associated with lower prediction errors, whereas lower reliability intervals capture instances with higher instability of the model performance. Across all approaches, reliability scores show a clear alignment with actual model performance at the instance level, and in addition, the rule-based approach incorporates domain knowledge and interpretability, which is particularly relevant in clinical contexts. These findings highlight the importance of pointwise reliability as a complementary tool to improve transparency and support more informed decision-making.

ABSTRACT. Continuous non-invasive blood pressure (cNIBP) monitor- ing still remains an open challenge. While numerous methods have been proposed to estimate blood pressure (BP) from physiological signals, lit- tle attention has been given to quantifying the trustworthiness of these estimates in real time, particularly in the absence of reference measure- ments. This work addresses the pointwise reliability problem in continuous BP estimation by proposing a model-agnostic evaluation framework for un- supervised reliability measures. Using a dataset of 40 surgical patients, reliability scores are computed from physiological feature spaces and eval- uated alongside blood pressure estimates obtained from a zero-order hold (ZOH) model in 10-second epochs. Two density-based approaches are analysed: Mean Neighbour Similarity (MNS) and the kernel-based den- sity measure (DENS). The proposed framework combines correlation analysis, threshold-based rejection curves, and distribution-aware diagnostics to characterize the behaviour of the reliability scores in relation to estimate error. Results show that both measures provide consistent ranking of prediction error and enable effective filtering of unreliable estimates, with error reductions of up to 40% under selective rejection. However, reliability scores exhibit limited calibration with absolute error, restricting their interpretation as confidence estimates. Overall, the findings point to the practical value of reliability measures for gating and quality control, while emphasizing the need for improved calibration and adaptive strategies for clinical deployment.

ABSTRACT. Reliability estimation is essential for the safe deployment of medical AI systems, particularly in high-stakes clinical environments where incorrect predictions may have serious consequences. However, many existing reliability assessment methods, such as distance-based and neighborhood-based approaches, rely on access to global data or sample-level comparisons, which limits their applicability in privacy-sensitive settings. In this paper, we present a density-based reliability estimation method based on Gaussian Mixture Models, which promotes data obfuscation by replacing sample-level comparisons with compact parametric density representations. The core idea is to model class-conditional data distributions in the representation space and estimate prediction reliability through likelihood-based scoring, thus avoiding explicit neighborhood search and direct distance computation. This makes the method well suited to settings where data sharing is restricted. The approach also extends naturally to federated learning, where each client learns local density models from private data without exposing raw samples. By aggregating model parameters rather than transferring instance-level information, a global reliability estimator can be constructed across distributed medical data sources. In this way, the proposed method provides a practical and privacy-preserving solution for reliability estimation in medical AI, while also offering a unified perspective across centralized and federated settings.

ABSTRACT. Continuous cuffless blood pressure monitoring has the poten- tial to improve patient safety by providing uninterrupted, non-invasive haemodynamic information. However, current data-driven estimators of- ten provide only point predictions and lack mechanisms to assess the reliability of individual estimates during operation. This is particularly important in clinical settings, where signal artefacts, physiological vari- ability, and temporal drift may cause estimation errors to vary substan- tially over time. This work proposes a local conformal prediction frame- work for uncertainty-aware cuffless blood pressure estimation. A simple regression model is used to estimate mean arterial blood pres- sure from window-based ECG, PPG, and respiratory features. To quan- tify uncertainty, we introduce a physiologically motivated nonconformity score that combines local feature-space consistency with prediction con- sistency among neighbouring calibration samples. A local conformal cal- ibration strategy is then used to construct adaptive uncertainty intervals and derive reliability labels for individual predictions. The framework was evaluated on 51 ICU recordings from the PhysioNet/Computing in Cardiology Challenge 2010 MIMIC II dataset. Results show that the proposed uncertainty estimates are positively cor- related with true prediction error and can identify unreliable predic- tion windows with high specificity. Although empirical coverage does not perfectly match nominal conformal guarantees under the non-stationary time-series setting, it remains monotonic with the confidence level, sup- porting the use of the method as a relative uncertainty and reliability indicator. These findings suggest that physiologically informed local con- formal prediction can support safer cuffless blood pressure monitoring by flagging estimates that may require reference measurement or recalibra- tion.

ABSTRACT. The classification of human emotions is being explored in diverse brain-computer interaction applications, including psychotherapies, gamming, and automotive systems, among others. However, some practical limitations in the adoption of these techniques are related to the lack of trustworthiness in the predictions provided by the associated machine learning (ML) models. In the present study, different methods for pointwise reliability assessment in the prediction of human emotions, evaluated through electroencephalography (EEG) signals, are explored. Pointwise reliability provides an estimate of how much an individual prediction for a new instance can be trusted. For this purpose, a dataset containing EEG recording from individuals exposed to video clips intended to induce positive, negative, and neutral emotions was used. The signals were preprocessed, features were extracted, and feature engineering techniques were applied to improve data quality and reduce the number of features. Five ML models (shallow learning) were trained, and eight reliability metrics were implemented. The results demonstrated moderated classification performance, with an accuracy of about 51 % (for three classes), while the reliability metrics showed strong correlation with accuracy for some of the implemented approaches, including a newly proposed metric (Pearsons’s correlation coefficient of 0.88). Although these methodologies require further exploration to improve performance, the study demonstrates the suitability of incorporating reliability metrics into the prediction of human emotions, representing a novel approach.