ABSTRACT. Collective variable (CV) identification in biomolecular simulations represents a dimensionality reduction challenge in which high-dimensional molecular configurations must be projected onto low-dimensional representations capable of capturing the key dynamics of the system. Machine learning approaches have emerged as powerful alternatives to traditional physics-based methods; however, their adoption among computer scientists and computational practitioners has remained limited due to literature that often emphasizes chemical intuition over algorithmic considerations. This review provides a computational perspective on ML-based CV discovery by analysing the trade-offs between automation, interpretability, and performance across major algorithmic classes. The findings indicate that automation has largely been resolved across modern methods, leaving interpretability versus performance as the primary remaining trade-off. Linear methods, such as LDA variants and lasso regression surrogate models, offer high interpretability but may sacrifice performance, whereas neural networks achieve substantially greater sampling efficiency, with improvements of up to 200-fold, at the expense of mechanistic understanding. Hybrid approaches, including surrogate models and Bayesian frameworks, demonstrate strong potential to bridge this gap by combining near-optimal performance with improved interpretability. The review further presents a decision-making framework for algorithm selection based on research objectives, discusses transferability limitations, and identifies scalability and the preservation of interpretability as critical challenges for the next generation of methods.

ABSTRACT. Background: Pneumonia is a leading cause of ICU mortality, yet existing severity scores have limited accuracy for short-term mortality prediction. Objective: To develop a machine learning model for dynamic 24-hour mortality prediction in ICU patients with pneumonia. Methods: We analyzed 29,847 clinical snapshots from 17,648 pneumonia patients in the eICU database (200+ US hospitals). Snapshots were triggered at critical events (ICU admission, ventilation start, vasopressor start, lactate elevation). We proposed the Soft Voting ensemble which combined LightGBM and XGBoost as our primary solution, compared it with other models via the hold-out validation scheme, and assessed performance against an APACHE-only baseline. Results: The Soft Voting ensemble (LightGBM + XGBoost) achieved the highest PR-AUC of 0.372 (95% CI: 0.265–0.484) and ROC-AUC of 0.855 (95% CI:0.817–0.888), with Platt-calibrated calibration (Brier 0.0422, ECE 0.0107). The Soft Voting ensemble achieved a high PR-AUC improvement over APACHE-only logistic regression. SHAP analysis identified age, APACHE score, and maximum lactate as top predictive features. Conclusions: The Soft Voting ensemble substantially outperformed APACHE-based prediction, suitable for early warning applications. Nonetheless, external validation is required before clinical deployment.

ABSTRACT. Regulatory requirements for safe levels of CO₂ rebreathing in continuous positive airway pressure (CPAP) masks used to treat obstructive sleep apnoea are specified in the ISO 17510:2015 Annex F protocol. Satisfaction of these requirements is currently mandated using physical bench testing on custom-built rigs. We used a previously validated cardiopulmonary simulator to develop physiologically informed digital twins that could facilitate in-silico assessment of CO₂ rebreathing. The digital twins were shown to be capable of closely matching the outputs of an industrial bench testing rig. Several enhancements to the digital twins were also introduced and evaluated: (1) replacement of the sinusoidal breathing profile with a subject-specific, asymmetric breathing waveform; (2) personalisation of tidal volume (VT), anatomical dead space (VDanat), and oxygen consumption (V̇O₂) using anthropometric data; (3) incorporation of a layered and conal mixing gas dynamics model in the CPAP mask; and (4) introduction of a CPAP jet-flow module capturing inlet mixing effects. Simulations were performed at baseline and at CPAP pressures of 3, 4, 5, and 10 cmH₂O under two pathways: constant tidal volume at different CPAP pressures, replicating ISO 17510:2015 bench test conditions, and constant inspiratory effort at different CPAP pressures, replicating spontaneous breathing patterns typically observed in patients. Under constant inspiratory effort, end-tidal CO₂ pressure (PexpCO₂) increased monotonically with CPAP, with the largest relative increase of 18.8% remaining within the ISO-defined 20% safety threshold. The conal mixing model revealed physiologically plausible mask washout behaviour consistent with computational fluid dynamics literature, including jet-driven CO₂ clearance at higher pressures. Collectively, these results demonstrate that digital twins can replicate bench test outcomes and also potentially provide more patient-specific and more physiologically realistic testing than standardised physical bench tests, advancing the case for their incorporation into the regulatory assessment of CPAP and other non-invasive respiratory support systems.

ABSTRACT. Automated hepatorenal index (HRI) computation from abdominal ultrasound requires organ segmentation that remains reliable across all frames of a routine examination, including liver-only views where no renal cortex is visible. Standard deep learning segmentation models fail silently on these out-of-distribution inputs, producing anatomically implausible masks that corrupt HRI estimates. In this study we developed a novel single-network architecture that replicates the clinician's cognitive process: first confirming the view type, then delineating organ boundaries. A shared ResNet34 encoder drives both a classification branch and a U-Net-like segmentation decoder augmented with explicit feature injection points at every upsampling stage. Two complementary injection mechanisms were evaluated (Feature-wise Linear Modulation (FiLM) for channel-wise conditioning and spatial attention gating) alongside a six-level progressive ablation study. On a multi-platform dataset of 801 patients, the combined model achieved mean organ DSC=0.91 and IoU=0.84 on hepatorenal views, with 94.3% balanced view-classification accuracy. Critically, on acoustically challenging frames where the segmentation-only baseline yields implausible masks (DSC=0.714), the proposed model recovers accurate predictions (DSC=0.965). Classification-guided feature injection is an effective and practical strategy to improve segmentation robustness across the distributional variability of routine abdominal ultrasound, providing a reliable foundation for automated multi-image HRI computation in MASLD screening.

ABSTRACT. Vector-borne diseases pose a serious threat to global health, but traditional entomological surveillance is constrained by the slow, costly bottleneck of manual taxonomic identification. To overcome these barriers, we propose a rigorous pipeline for the development and validation of highly optimized Edge Computer Vision models designed for automated, in-the-field mosquito identification using ultra-low-cost, general-purpose hardware (Raspberry Pi 5). The methodology evaluated four distinct convolutional architectures through comprehensive ablation and data scalability studies, directly comparing standard transfer learning against deep fine-tuning strategies to effectively address severe class imbalance. Subsequently, the models underwent Full Integer Quantization, mapping floating-point parameters to an 8-bit discrete domain (INT8) to bypass the computational limitations of edge devices.

Experimental validation revealed two striking operational achievements. First, prioritizing the epidemiological necessity to minimize false negatives, the fine-tuned MobileNetV2 architecture emerged as the optimal solution by achieving an exceptional recall of 95.65% on the target vector class. Second, the INT8 software optimization drastically compressed inference latency on the edge device's standard ARM CPU from 99.37 ms to a record 2.11 ms. This extraordinary 47-fold acceleration effectively matches the throughput of expensive dedicated AI hardware accelerators, laying a technologically robust and economically sustainable foundation for a highly scalable global IoT surveillance network.

ABSTRACT. Introduction: Respiratory diseases account for a substantial proportion of global deaths, disproportionately affecting populations in low- and middle-income countries (LMICs). In many low-resource settings (LRS), access to reliable diagnostic equipment such as spirometers remains limited due to cost, infrastructure requirements, and contextual constraints. Conventional medical devices are typically designed for high-income contexts and do not adequately address these challenges. This study presents the design and preliminary validation of a frugally engineered, low-cost spirometry device intended to support respiratory assessment in LRS using 3D printing, low-cost electronics, and data-driven processing. Materials and Methods: The spirometer’s flow tube and casing was designed using CAD on fusion 360 and manufactured with 3D printing. The airflow modelling was based on Bernoulli’s principles, relating volumetric flow rate and differential pressure in a constricted tube. Differential pressure across the tube was measured using a low-cost pressure sensor and interfaced with an Arduino-based microcontroller. Signal acquisition and processing were carried out using custom MATLAB routines, incorporating data processing techniques such as low pass filtering to compute standard spirometry variables. Bench validation was conducted using a handpump with a 2 L volume, while early-stage usability and performance were assessed through expiratory tests conducted by 30 participants, with results compared against a commercially available CE-marked spirometer. Results: The fabricated spirometry tube generated consistent and measurable differential pressure signals, which were successfully converted into airflow waveforms. Across 30 handpump trials, the device demonstrated a mean percentage error of 1.53%. During the usability evaluation, participants produced spirometry results within expected normal ranges and when compared to the CE-marked reference device, it achieved high reliability and consistency with this device. For FEV1 it achieved r= 0.97 and an ICC value of 0.97, indicating very high consistency between devices. Discussion: The proposed device demonstrated accuracy comparable to commercially available spirometers while remaining significantly lower in cost, with an estimated small-scale manufacturing cost of approximately £60. These results indicate strong potential for the device to contribute to improved respiratory diagnostics in LRS. Future work will focus on expanding functionality through the integration of machine learning techniques and further clinical validation.

ABSTRACT. High-impact diseases, such as cervical cancer and Leishmaniasis, result in high morbidity and mortality rates, especially in low-resource settings due to a shortage of specialists and the centralization of diagnostic services. In Perú, rural healthcare providers face these challenges, known as structural gaps, which significantly delay conventional screening. This study proposes the AIM-LRS scheme as a three-hierarchical workflow to improve diagnostics of these diseases through the use of artificial intelligence and telemedicine. Methodology: The three layers proposed are defined as, Layer 1 describes the telemedicine workflow and the integration of processing via Convolutional Neural Networks. Layer 2 includes process monitoring using the Lean Healthcare methodology, and Layer 3 features a Store-and-Forward model for deferred data transmission. Results: Evaluation of the Value Stream Map revealed that Total Lead Time decreases from 25 days and 2.5 hours to just 7 hours and 10 minutes. Furthermore, the preliminary proposed architecture model for Layer 1-step 3, demonstrated an accuracy of 0.991 in detecting host cells and 0.993 in detecting nuclei. This validates the system’s ability to generate early alerts and optimize workflows in rural areas.

ABSTRACT. Reliable augmented reality (AR) guidance in robot-assisted partial nephrectomy (RAPN) is still limited by the difficulty of accu- rately aligning preoperative models with the intraoperative scene in the presence of tissue deformation, manipulation, partial occlusions, and viewpoint changes. To support research on this problem, we present a patient-specific deformable kidney phantom dataset together with a proof-of-concept workflow for frame-wise reference alignment annotation in simulated RAPN. The dataset comprises 21 stereo recordings acquired from 7 participants across 3 difficulty levels and 4 representative RAPN task phases, for a total of 158,520 synchronized stereo pairs. Stereo en- doscopic videos were acquired on the da Vinci Research Kit (dVRK) during simulated tasks performed on a CT-derived deformable phantom. Phantom segmentation, stereo 3D reconstruction, and offline annotation- oriented rigid registration were used to associate the visible reconstructed phantom surface with the corresponding phantom model. A subset of 563 frames was manually annotated with validated frame-wise rigid reference alignments. Preliminary results indicate that the proposed workflow can recover plausible rigid alignments while also highlighting the limitations of a purely rigid formulation. Overall, the proposed framework provides a realistic reference dataset and a practical first step toward future de- formable annotation and registration studies for AR-guided RAPN.

ABSTRACT. Genetic factors contribute to inter-individual variability in athletic performance, yet their practical relevance in heterogeneous athlete populations remains underresearched. This study investigated the distribution of angiotensin-converting enzyme (ACE) I/D and alpha-actinin-3 (ACTN3) R577X polymorphisms in 85 athletes from Bosnia and Herzegovina and explored their associations with sex and sport type. Genotyping was performed using real-time PCR, and genotype and allele frequencies were analysed using descriptive and inferential statistics. The ACE ID genotype (43.5%) and ACTN3 RX genotype (47.1%) were the most prevalent, and both polymorphisms were in Hardy-Weinberg equilibrium. No statistically significant associations were observed between genotype distribution and sex or sport type, although descriptive trends suggested a somewhat higher prevalence of power-oriented profiles in athletes from power disciplines. These findings indicate limited predictive value of individual polymorphisms in this cohort and support a polygenic and multifactorial model of athletic performance. The results provide initial insight into the genetic structure of athletes from Bosnia and Herzegovina and highlight the need for integrative approaches in personalised sports performance profiling.

ABSTRACT. Knee osteoarthritis (KOA) is increasingly recognized as a multi‑tissue disease involving articular cartilage, subchondral bone, and surrounding musculature. However, most imaging studies still focus on individual tissues, limiting insight into joint‑level degeneration. In this study, we propose a multi‑tissue imaging framework to investigate coordinated structural alterations across cartilage, bone, and muscle in KOA. Quantitative morphological and densitometric features were extracted from cartilage segmented on MRI and projected onto co‑registered CT scans, as well as from subchondral bone and knee muscles segmented directly on CT. Canonical Correlation Analysis (CCA) was used to characterize multivariate associations between cartilage and bone, and between cartilage and muscle, with statistical significance assessed via permutation testing. Strong and significant multivariate coupling was observed for both tissue pairs. The first canonical correlation reached r=0.80 for cartilage–bone and r=0.95 for cartilage–muscle, indicating particularly strong shared variation with muscle tissue. Loadings analysis revealed that density‑related features dominated cross‑tissue associations, with subchondral bone density variability and intramuscular adipose tissue (IMAT) properties emerging as key contributors. Classification analyses further demonstrated the added value of a multi‑tissue approach: combining cartilage features with CCA‑selected bone or muscle features improved discrimination between degenerative and control knees, achieving F1 scores up to 0.82 ± 0.09, compared to 0.72 ± 0.17 using cartilage alone. Overall, these findings support a system‑level view of KOA and highlight the potential of multi‑tissue quantitative imaging biomarkers for improved disease characterization.

ABSTRACT. Digital twins are increasingly proposed as patient-specific, dynamically updated computational representations that integrate multimodal data to support prediction, monitoring, and personalized care. In cognitive decline and dementia, their potential is considerable because disease progression is slow, heterogeneous, and shaped by biological, cognitive, clinical, and social factors. However, digital twins will remain clinically incomplete unless they capture sex- and gender-related differences in disease burden, symptom presentation, testing performance, care pathways, treatment exposure, and support needs. Women carry a disproportionate burden of Alzheimer's disease and related dementias and often show better early verbal-memory performance, which can mask disease and delay diagnosis. The objective of this paper is to define a gender-aware digital-twin perspective for cognitive loss and to translate the literature into design requirements for an integrative IT framework. Electronic health records (EHRs), natural language processing (NLP), digital cognitive assessment, gamified testing, and longitudinal monitoring apps together provide an important technological basis for such a framework. The paper argues that the digital twin should be understood not as a single tool or as a conventional electronic health-record platform, but as a modular, continuously updated patient model that links interoperable assessment, prediction, follow-up, and support components. Such systems should be co-developed with patients, caregivers, relatives, and clinicians to better address women's diagnostic and long-term care needs across the care continuum

ABSTRACT. Accurate, practical measurement of milk intake during breastfeeding remains a major unmet need, as existing methods are limited by cost, inconvenience, or lack of suitability for routine use. Motivated by this gap in maternal and infant care, we developed a flexible sensing technology for live tracking of milk expression from the breast. To support development of this approach, we modified a silicone-based breast phantom to produce controlled internal volume changes representative of lactation-related deformation and suitable for repeatable benchtop testing. Using this platform, we evaluated multiple attachment methods for Velostat-based strain sensors, including adhesives, fabrics, and tapes, to determine their ability to reliably track phantom deformation. The phantom produced repeatable displacement fields with peak surface displacements on the order of millimeters, and when the internal volume was increased by 50 mL, DIC-based surface reconstruction estimated a volume change of 45.49 mL, corresponding to a 9% error. Among the attachment methods tested, medical tape-based approaches, particularly kinesiology tape and transparent film medical dressing, showed the most reliable sensor performance, while glue- and bra-based attachments were less effective and more difficult to implement consistently. These results support the use of the phantom as a benchtop testbed for lactation sensing and indicate that surface-mounted Velostat-based sensors are a promising approach for future wearable device development.

ABSTRACT. Robot-assisted gait training (RAGT) is widely used in paediatric neurorehabilitation for children with cerebral palsy (CP), yet the physiological mechanisms underlying individual rehabilitation response remain insufficiently characterized. This study investigates the relationship between multimodal physiological adaptations and clinical outcomes through a comparative case analysis of two children with CP characterized by different levels of motor impairment (Case A: GMFCS II; Case B: GMFCS V). A multimodal framework integrating heart rate variability (HRV), facial infrared thermography (IRT), and exoskeleton-derived symmetry indices was employed. Physiological features were expressed as changes between baseline and post-intervention (ΔT2–T0) and analysed alongside clinical outcomes, including GMFM, MAS, WeeFIM, and PedsQL. Both cases exhibited clinical improvements, although with different domain-specific patterns. Case A showed moderate gains in motor function and autonomy, while Case B demonstrated larger improvements in quality of life and functional independence, associated with distinct physiological adaptation profiles. Case A exhibited a pronounced modulation of autonomic activity, whereas Case B showed a more consistent increase in thermal high-frequency components and improvements in biomechanical symmetry. These findings suggest that rehabilitation response is characterized by subject-specific physiological patterns and cannot be described by a single biomarker. Multimodal monitoring may support the development of personalized rehabilitation strategies.

ABSTRACT. Continuous Glucose Monitoring sensors (CGMs) are portable, minimally invasive devices that are transforming the management of Type 1 Diabetes (T1D). Pressure-induced sensor attenuations (PISAs) are artifacts that can lead to false low glucose readings, compromising the reliability of the sensors. This study investigates the application of Dynamic Time Warping (DTW) algorithm, a pattern recognition technique to assess the similarity between time-series data, for the retrospective identification of these faults. DTW algorithm is employed for detecting a sequence with a specific shape (associated to a PISA episode) in a larger signal and has the ability of handling variations in the duration and timing of the events through a non-linear alignment of the time-series considered. The algorithm is tested using a dataset generated by a state-of-the-art T1D patient simulator. Results show that DTW achieves a recall of 0.87 with an average of 2 false alarms during the monitoring period (10 days). These findings highlight the potential of DTW to enhance CGM reliability, providing a robust method for detecting sensor faults and contributing to the development of more accurate diabetes monitoring technologies.

ABSTRACT. Medication administration remains a critical and error-prone process in healthcare systems, with significant implications for patient safety and clinical outcomes. This study presents PillsGuard, an Internet of Things-based platform designed to enhance the safety and accuracy of medication administration through automated verification and real-time alert mechanisms. The proposed system integrates barcode identification, a microcontroller-based architecture, and multimodal feedback interfaces to validate the correspondence between patient, medication, dosage, and administration schedule. The platform includes modules for data acquisition, processing, visualization, and automated medication dispensing using servo-controlled mechanisms. Additionally, the system provides visual and auditory alerts to support healthcare personnel during decision-making. The solution was developed following structured design methodologies and evaluated in a simulated clinical environment. Results demonstrate the feasibility of the system and its potential to reduce medication errors and improve workflow efficiency. PillsGuard represents a scalable and cost-effective approach for strengthening patient safety by integrating real-time validation and automation into medication management processes, with future applicability in real clinical settings.

ABSTRACT. This study evaluated whether inspiratory power–controlled ventilation (IPCV) produces distinct mechanical loading patterns compared with volume-controlled (VCV) and pressure-controlled ventilation (PCV), with respect to Ventilatory-Induced lung injury (VILI). A hybrid respiratory simulator connected with a novel IPCV ventilator was used to reproduce lung conditions: normal adult mechanics and Acute Respiratory Distress Syndrome (ARDS). Ventilatory modes: VCV, PCV, and IPCV were tested at respiratory rates (RR) of 10, 15, and 20 breaths/min. Airway pressures, inspiratory and expiratory flows, ventilator-delivered work of breathing, and inspiratory mechanical power were analysed. Simulations proved that ventilation modes significantly affected all analysed variables. The biggest differences between IPCV, VCV and PCV were observed in maximal inspiratory mechanical power (MPmax), peak inspiratory flow (PIF), and work of breathing (WOB). For example, in IPCV mode at 15 breaths/min, MPmax, PIF, and WOB were 81.26, 29.36, and 7.32% lower, respectively, than in PCV (P<0.001). MPmax was also higher (12.74%; P<0.001) in VCV, compared to IPCV. However, PIF and WOB in VCV were lower by 24.49 and 5.86%, respectively, than in IPCV (P<0.001). In conclusion, the IPCV represents a distinct strategy for organising inspiratory energy delivery that can significantly diminish MPmax and consequently the risk of VILI, with effects influenced by RR and respiratory system mechanics.

Conflict of Interest: All authors have no conflicts of interest.

Acknowledgement: This research was fully funded by the National Science Centre, Poland, under Grant No. 2023/50/A/ST7/00498.

ABSTRACT. Stereotaxic surgery on rodents is a delicate procedure essential for research on central nervous system physiology and pathology. Surgeons generally rely on post-mortem-based atlases that only represent the cerebrospinal fluid spaces in collapsed state. This synchrotron radiation-based in vivo tomography study with 12-week-old C57BL/6J mice enabled the extraction of the cerebrospinal fluid space and relevant stereotaxic landmarks to optimize the intracerebroventricular infusion. For this purpose, we injected a contrast agent into the right lateral ventricle. Its temporal distribution could be followed by a series of tomograms acquired at the BMIT-ID beamline of the Canadian Light Source. The segmented cerebrospinal fluid spaces and the semi-automatically determined stereotaxic landmarks were rigidly aligned to an anatomically oriented reference frame. For 14 mice, the 3D coordinates within the left lateral ventricle with the maximum distance to its surface were correlated to the animal’s weight and sex. The stereotaxic coordinate frame showed an overlapping volume of 0.85 mm³ of the left lateral ventricles. The optimized coordinates were 90 µm deeper for males with respect to females. Therefore, separate sets of stereotaxic coordinates for ventricular infusion should be considered for female and male mice with the aim to improve the success of the procedure.

ABSTRACT. CT perfusion can be applied to detect CCD in acute stroke by analysing cerebellar perfusion abnormalities. This study presents a method for CCD assessment in the acute phase of ischemic stroke, enabling detailed characterization of cerebellar asymmetry through a novel perfusion normalization framework applied to CTP maps. CT Perfusion images were registered to standard space using an ad hoc procedure and global and regional asymmetry indices in Sensorimotor, Cognitive and Vestibular areas, were derived from standardized MTT maps. Voxel-wise analysis was performed to characterize the spatial distribution of cerebellar perfusion asymmetries using a seed-based region-growing algorithm. Patients were stratified according to admission stroke severity. Global and regional asymmetry indices were compared. Associations between asymmetry intensities and neurological deficit at admission and clinical outcome were investigated. Global asymmetry was observed in 76.5% of patients. Voxel-wise analysis revealed specific spatial patterns reflecting cerebrocerebellar organization. Patients with more severe stroke showed higher global and regional asymmetry, particularly in Sensorimotor cerebellar region, which represented the most affected territory. Increasing asymmetry was associated with worse neurological status and lower odds of favourable outcome. This study advances the characterization of CCD in acute ischemic stroke by introducing an approach allowing comprehensive assessment of cerebellar perfusion abnormalities.

ABSTRACT. Objective assessment of upper limb function remains a critical challenge in monitoring cerebellar ataxia, as traditional clinical scales often lack the sensitivity required to detect subtle changes in fine motor control. This work introduces an innovative Mixed Reality (MR) framework that integrates an ecological functional task—inspired by gardening activities—with an automated kinematic analysis pipeline. Developed for the Microsoft HoloLens 2 platform, the system employs event- triggered data acquisition to segment the task into three distinct operational phases: digging, sowing, and watering. This approach enables the extraction of granular metrics for each step, allowing for a clear distinction between motor planning latency and executive efficiency. Preliminary results from two healthy subjects demonstrated the system's capability to map specific error patterns, such as apical trajectory deviations and proximal-distal coordination instability during rotational movements. The extracted metrics (RMS accuracy, target stability, and drift) confirm the robustness of the pipeline and its capacity to generate interpretable quantitative indicators. The framework is designed to complement standard clinical scales, providing therapists with objective data to personalize rehabilitation interventions and support the longitudinal monitoring of motor performance in patients with coordination disorders.

ABSTRACT. Biohybrid engineered tissues (BET) hold great promise in the pursuit of generating functional biological tissues for either research and/or clinical applications. These constructs involve biologically active material, therefore a high variability in the progress of their in vitro maturation process is expected. Traditionally, the evaluation of tissue engineered constructs relied on a multitude of specific metrics tailored to the tissue type being engineered. While these metrics can be highly informative, they have limitations, including a lack of generalizability and comparability across different constructs, leading to an incomplete reference values that make difficult to compare results across different studies and laboratories. Starting from this paradigm, a quantitative virtual profile (qVP) based on histological and metabolomic parameters was developed for an advanced monitoring of 3D in vitro bone extracellular matrix mineralization. This approach revealed a specific pattern during the maturation process of a bone tissue engineering-based product, suggesting that integrating data from various analytical techniques can yield a reliable tool for establishing unambiguously when a BET construct has reached a state of maturity. In this context, machine learning algorithms can be employed to uncover patterns, correlations, and trends that may not be apparent via traditional statistical analysis, enabling the development of a dimensionless predictive index. This idea new work is expected to unleash the full potential of tissue engineering, marking a groundbreaking advancement in the evaluation of TEed constructs and a significant milestone in the field. As the use of dimensionless indices becomes more widespread, the potential for innovative breakthroughs will bring us closer to the realization of the promise of TE in clinical applications.

ABSTRACT. Otosclerosis is a complex middle-ear pathology characterized by subtle bone remodeling that can lead to severe hearing loss if not detected early. Although Computed Tomography (CT) is the gold standard for clinical evaluation, diagnosing early-stage otosclerosis remains highly challenging due to low radiological contrast and high operator dependency, particularly for non-specialist clinicians. Building upon prior convolutional neural network (CNN) benchmarks that achieved an initial baseline accuracy of 81%, this study aims to develop and validate an enhanced diagnostic framework designed to improve classification performance and robustness in clinical decision support. The primary contribution of this research is the integration of multi-head self-attention mechanisms and Vision Transformers (ViTs) specifically optimized for high-resolution temporal bone CT patches. Unlike traditional CNN approaches that focus strictly on local pixel-level features, this hybrid architecture captures long-range spatial dependencies and global contextual changes across complex ear anatomy, thereby drastically minimizing false negatives in early fenestral and retrofenestral lesions. The expanded framework was trained and evaluated using an enriched dataset of temporal bone CT images from patients with confirmed otosclerosis. We implemented a hybrid CNN-ViT architecture in which convolutional layers serve as feature extractors and transformer blocks process the embedded tokens to model macro-structural relationships. Model validation was performed using a multi-cycle hold-out strategy. Furthermore, gradient-weighted class activation mapping (Grad-CAM) was integrated into the interface to ensure diagnostic explainability for radiologists. Preliminary evaluations of the transformer-based framework demonstrate a substantial performance leap, achieving 89% predictive accuracy in discriminating pathological states from normal cases. This represents an 8% improvement over our previous CNN-only baseline. Expert radiological validation confirmed that the model’s self-attention maps accurately align with true hypodense otosclerotic foci, enhancing both diagnostic precision and clinician trust. The proposed AI framework offers a highly accurate, explainable, and non-invasive tool for automated otosclerosis screening. By shifting from local convolutions to global attention mechanisms, this system will provide reliable diagnostic support, potentially democratizing access to specialized otological expertise and facilitating timely surgical or medical interventions to preserve patients' hearing.

ABSTRACT. Usability and co design are increasingly recognised as essential drivers for the development of safe, effective, and equitable medical and digital health technologies. This paper summarises the contributions from a thematic symposium on medtech usability, presenting an integrated framework that combines methodological innovation, patient centred design, and regulatory alignment. After a brief introduction to the topic, section 2 introduces the Usability Room, a hybrid physical–virtual infrastructure that enables realistic simulation of use environments, structured observation, and iterative formative and summative evaluations in accordance with IEC 62366-1 and MDR 2017/745. This approach supports early identification of use related risks, strengthens safe by design development, and fosters multi stakeholder consensus. Section 3 presents the Digital Fragility Questionnaire, a rapid and multidimensional tool, co designed to assess patients’ cognitive, technological, and psychological readiness for telemedicine. The fourth section reports on co design as a practical advocacy mechanism, addressing cognitive and communicative exclusion and promoting inclusive digital health ecosystems. The fifth section highlights the role of Notified Bodies in assessing usability evidence across the device lifecycle, emphasising the importance of post market data for continuous risk management. Finally, in section 6, the Pausetiv case-study exemplifies how clinical co design and human factors engineering can translate complex health content into a safe and accessible Software as a Medical Device for menopausal health. Together, these contributions outline a comprehensive pathway for transforming users into active co creators of medical technologies.

ABSTRACT. High tibial osteotomy (HTO) is a widely established surgical intervention for realigning varus knee kinematics; however, the resulting biomechanical redistribution is highly sensitive to the morphological variants of the lateral meniscus. This study quantifies morphology-dependent load redistribution following HTO using patient-specific finite element models categorized into non-discoid (non-DLM), incomplete discoid (iDLM), and complete discoid (cDLM) meniscus types. Under simulated gait (75% BW) and running (80% BW) conditions across a parametric range of weight-bearing line (WBL) increments, the 95th percentile Von Mises stress was evaluated to assess the mechanical demand on the joint. The FE analysis revealed divergent biomechanical behaviors: the non-DLM model maintained relatively stable meniscal stress (27.8–32.3 MPa) despite significant stress concentration at the tibial hinge (112.2 MPa at 62.5% WBL), while the iDLM model exhibited negligible load transmission (<0.6 MPa), indicating a functional bypass of the lateral compartment. In contrast, the cDLM variant demonstrated pronounced alignment-dependent stress amplification, with magnitudes escalating from 19.2 to 69.9 MPa during gait and from 20.0 to 73.9 MPa during running. These findings suggest that post-HTO load redistribution is not a uniform mechanical phenomenon but is governed by meniscal morphology, necessitating morphology-specific alignment strategies to ensure biomechanically optimized outcomes.

ABSTRACT. Balance impairment is a common cause of falls and functional decline among neurological patients and aging populations. Wearable assistive devices employing angular momentum exchange – specifically, reaction wheels – remain an underexplored intervention modality for trunk stabilization. Our team developed a novel, wearable, power-autonomous prototype device for active balance assistance, based on a reaction wheel (project Sentinel). This preliminary study uses the prototype device as a data-acquisition platform to assess postural sway in two healthy participants across progressively challenging standing tasks. The aim was to assess the device's measurement capability and to derive a subject-specific dead-band threshold. Both participants wore the device and completed six standing tasks involving sensory, mechanical, and cognitive challenges. Postural sway was quantified using roll root-mean-square, a 95% confidence ellipse area, and sample entropy. A provisional dead-band threshold of 0.281° was derived from participant S00’s balance quiet-standing data. Across the tested conditions, the device demonstrated the ability to capture differences in sway amplitude among participants. The consistency of findings across the three metrics points toward reliable device measurements, but a larger cohort is required to confirm it. The posterior mass of 4.8 kg introduced by the device primarily shifts the Center of Mass (CoM) in the sagittal plane and is expected to have minimal influence on the frontal plane in the current experimental analysis. The conclusions from these initial experiments motivate evaluation of the device using closed-loop, threshold-based control, pending cohort expansion and formal validation.

ABSTRACT. The assessment of osteogenic maturation remains a major challenge in bone tissue engineering, due to high variability across cell culture experiments and the lack of a recognized standard for osteogenic maturation. Although extracellular matrix (ECM) mineralization is commonly considered an indicator of differentiation, it does not necessarily reflect actual maturation of the construct. This project aims to develop a more standardized and quantitative approach for assessing ECM mineralization by integrating an improved Alizarin Red assay with longitudinal metabolic profiling of cell cultures. Cultures were maintained in osteogenic conditions and monitored during the early stages of commitment. Free calcium and other metabolites in the culture medium were measured over time using a laboratory metabolite analyzer to investigate extracellular biochemical changes associated with osteogenic progression. To improve the assessment of mineral deposition, a modified Alizarin Red quantification protocol is proposed. After staining, the bound dye was extracted using acetic acid, neutralized with ammonium hydroxide and then quantified using a spectrophotometer. A first calibration curve was generated using Alizarin Red standards prepared in the same extraction solution to estimate the molar concentration of recovered dye. To further relate the Alizarin signal to the calcium deposited in the ECM, a second calibration was performed using calcium phosphate standards processed with the same staining and extraction procedure. This dual calibration strategy was expected to provide a reproducible method for estimating calcium deposited in the ECM. Finally, ECM calcium measurements were compared with longitudinal free calcium levels in the culture medium to identify potential correlations between extracellular metabolic dynamics and matrix mineralization. By combining metabolic profiling with dual-calibrated Alizarin Red quantification, this framework provides a more standardized and non-destructive strategy for monitoring osteogenic cultures over time. If validated, it may support the identification of early biomarkers of osteogenic progression and reduce reliance on endpoint destructive assays.

ABSTRACT. Graphene oxide–silver nanoarchitectures (GO–AgNPs) have emerged as promising platforms for biomedical applications due to their synergistic antibacterial properties. However, controlling their biological performance while maintaining cytocompatibility remains a major challenge. In this work, we investigate the combined influence of graphene oxide chemical state—modulated through pH—and the phytochemical composition of green synthesis extracts on the formation and biological response of GO–Ag nanoarchitectures. Graphene oxide was synthesized and subsequently conditioned at different pH values in order to modulate its surface chemistry and functional group availability. Silver nanoparticles were then nucleated and anchored onto the GO sheets using plant-derived extracts rich in phytochemicals, which served as both reducing and stabilizing agents. The resulting nanoarchitectures were characterized to evaluate their structural and chemical features, and their antibacterial performance was assessed against representative Gram-positive and Gram-negative bacteria, including resistant strains. Additionally, cytocompatibility studies were conducted to identify concentration ranges suitable for biomedical applications. The results indicate that both the chemical state of graphene oxide, governed by pH, and the phytochemical composition of the synthesis medium play a key role in controlling nanoparticle anchoring, dispersion, and overall biological response. Certain nanoarchitectures showed enhanced antibacterial performance while maintaining cytocompatibility, underscoring the importance of controlling synthesis parameters to achieve biologically effective and safe materials. These findings contribute to a better understanding of the structure–property–biological response relationships in GO–Ag nanoarchitectures and support their potential use in antibacterial platforms for wound healing and related biomedical applications.

ABSTRACT. Background Desmoid tumors (DTs) are rare mesenchymal neoplasms characterized by local infiltrative growth and unpredictable clinical behavior. Although they lack metastatic potential, they may be clinically aggressive because of local recurrence and invasion of surrounding anatomical structures. Current international guidelines recommend active surveillance as the initial management strategy for most newly diagnosed, asymptomatic patients since many lesions remain stable or even regress spontaneously. However, up to 30% of patients experience early, clinically relevant progressive disease (PD) and ultimately require systemic therapy or other interventions. To date, no validated clinical, biochemical, or imaging biomarkers reliably predict DT behavior at diagnosis. Radiomics-based machine learning (ML) may provide quantitative imaging biomarkers to support early risk stratification and personalized follow-up. Objective This preliminary study aimed to assess the feasibility of radiomics-based ML for predicting early PD versus stable disease (SD) or spontaneous regression (SR) of DT undergoing active surveillance. Methods A preliminary cohort of 8 patients with histologically confirmed DT was retrospectively analyzed. Follow-up response was defined according to RECIST-based clinical-radiological assessment. Patients with SD or SR were grouped into the non-progressive class, while patients with PD were assigned to the progressive class. The dataset included 4 SD/SR and 4 PD patients. Manual tumor segmentations were used to extract 102 radiomic features from each two-dimensional tumor slice, yielding 147 slices. Due to the limited cohort size, the analysis was performed at slice level; however, all train-test splits were performed at patient level to avoid data leakage. In each external split, 6 patients were used for training and 2 for testing, considering all 16 balanced patient-level combinations. Hyperparameters were optimized on the training set using inner 3-fold stratified group cross-validation. Several supervised classifiers were evaluated, including Logistic Regression, Support Vector Machine, Random Forest, AdaBoost, Gradient Boosting, and Gaussian Naive Bayes (GNB). Models were tested with and without ANOVA F-test-based feature selection. Final performance was reported as mean ± standard deviation across the 16 external test combinations. Results Overall, the tested models showed moderate ability to discriminate PD from SD/SR, with substantial variability across patient-level splits. GNB achieved the best overall performance without features selection, with accuracy 0.658 ± 0.280, balanced accuracy 0.677 ± 0.262, F1-score 0.684 ± 0.258, recall 0.766 ± 0.308, and ROC-AUC 0.683 ± 0.287. The aggregated out-of-sample confusion matrix showed 169 true negatives, 135 false positives, 67 false negatives, and 217 true positives, suggesting good sensitivity for PD but moderate specificity. Features selection did not provide a consistent performance improvement. Conclusion This preliminary feasibility analysis suggests that MRI-derived radiomic features may contain useful information for predicting DT evolution during active surveillance. GNB emerged as the most promising exploratory model, particularly for identifying patients with PD. However, the moderate performance, high variability across splits, limited sample size, and relevant number of false positives indicate that these results should be interpreted as exploratory. Larger cohorts, patient-level validation, and robustness analyses are required before clinical translation. If confirmed, radiomics-based machine learning may enable personalized surveillance strategies, facilitating early treatment for patients at risk of rapid disease progression while reducing unnecessary imaging, patient anxiety, and overtreatment among those likely to experience SD or SR.

ABSTRACT. The treatment of superficial wounds remains a major challenge in healthcare due to the increasing prevalence of multidrug-resistant microorganisms and their ability to form biofilms, which significantly delay the healing process. In this context, the development of antimicrobial biomaterials has emerged as a promising strategy for improving infection control and promoting tissue regeneration. This work explores the functionalization of vinegar mother-derived bacterial cellulose with silver nanoparticles (AgNPs) through two different nucleation routes. The first approach consists of the in situ synthesis of AgNPs directly on the cellulose matrix, promoting nanoparticle formation and growth in the presence of the biopolymer. The second approach involves ex situ synthesis, where AgNPs are produced separately by conventional chemical reduction and subsequently incorporated into the cellulose network. Additionally, bacterial cellulose obtained from vinegar mothers cultivated using two different fruit substrates will be evaluated in order to determine whether the cultivation source influences the structural characteristics of the cellulose and its interaction with silver nanoparticles. The objective of this study is to investigate the combined effect of cellulose origin and nanoparticle nucleation route on the physicochemical and biological performance of the resulting biocomposites. Fourier Transform Infrared Spectroscopy (FTIR), X-ray Diffraction (XRD), and Scanning Electron Microscopy (SEM) will be employed to assess structural modifications, crystallinity, nanoparticle morphology, distribution, and anchoring within the cellulose matrix. Furthermore, biological assays will be conducted to evaluate antimicrobial activity and biocompatibility. The results are expected to provide valuable insights into the influence of cellulose microstructure and nanoparticle incorporation pathways on the development of advanced antibacterial biomaterials for wound healing and infection-control applications.

ABSTRACT. Brain organoids have emerged as powerful in vitro models for investigating human brain development, neurodegenerative diseases, and therapeutic responses. However, conventional histological evaluation relies on tissue fixation, sectioning, and staining, making longitudinal and rapid assessment challenging while introducing potential processing artifacts. In this study, we developed a label-free ultraviolet photoacoustic microscopy (UV-PAM) platform for high-resolution structural imaging of intact brain organoids. By exploiting the intrinsic ultraviolet absorption of nucleic acids, UV-PAM enables direct visualization of cell nuclei without exogenous contrast agents or histological staining, providing histology-like images while preserving the native tissue architecture. The proposed approach achieves subcellular-resolution imaging of whole brain organoids and enables quantitative assessment of nuclear morphology and spatial organization through automated image analysis. Morphometric features extracted from UV-PAM images were used to characterize structural alterations associated with biological perturbations, demonstrating the capability of the platform to detect subtle changes in tissue architecture. The label-free and non-destructive nature of the technique further enables repeated imaging of the same specimen, providing opportunities for longitudinal studies that are difficult to achieve using conventional histopathology. To validate the biological relevance of the acquired images, UV-PAM-derived structural information was compared with standard fluorescence histology, demonstrating strong agreement in the visualization of nuclear distributions and tissue morphology. Furthermore, the platform was applied to chemically perturbed brain organoids to demonstrate its utility for quantitative phenotypic analysis of neurotoxic responses. These results highlight the potential of UV-PAM as a rapid and robust imaging modality for organoid research. Overall, this work establishes UV-PAM as a powerful label-free histological imaging technology for brain organoids, enabling high-resolution, quantitative, and minimally invasive characterization of three-dimensional neural tissues. The proposed platform has broad potential for applications in developmental neuroscience, disease modeling, drug screening, and regenerative medicine, where rapid and repeatable assessment of organoid morphology is increasingly required.

ABSTRACT. Automated fluid management is a critical yet underserved requirement in the development of functional three-dimensional tissue constructs, where static diffusion is insufficient to sustain metabolically active cells beyond the 100–200 µm diffusive limit. This paper describes a modular microperfusion platform that addresses existing gaps in portability, automation, and reproducibility of perfusion-based bioreactor systems. The device integrates a dual-pathway hydraulic architecture driven by Actuonix L16 linear actuators and controlled by optoisolated three-way solenoid valves, mounted on a custom aluminum frame housing hybrid brass-stainless steel fluidic chambers with PTFE plungers and AISI 316L internal sleeves for biocompatibility and sterilizability. System intelligence is centralized on a custom PCB hosting an Arduino UNO microcontroller, a DS3231 TCXO-based real-time clock, a TB6612FNG dual H-bridge motor driver, and a switching buck converter for thermal-stable autonomous powering. Automation logic is implemented as a non-blocking finite state machine incorporating hydraulic safety delays to prevent pressure spikes and retrograde flows. A Python-based graphical user interface provides full protocol customization and real-time telemetry logging. Gravimetric validation (n = 50 cycles) yielded a post-calibration coefficient of variation of 0.20% and 0.24% for the feeding and discharging channels, respectively. Kinematic regression models confirmed linear actuator response across unloaded, dry-friction, and hydraulic-load conditions (R² > 0.996). Prolonged operation at 37°C over 48 hours demonstrated complete functional integrity and sterility. The all-inclusive prototype cost is under 1200 €, representing a cost-effective, reproducible solution for autonomous long-term perfusion in 3D cell culture applications.

ABSTRACT. Neurological disorders including epilepsy impose a disproportionate burden on sub-Saharan African (SSA) populations, yet access to electroencephalography (EEG)-based diagnostics remains critically limited across the region. This gap is not solely a consequence of resource scarcity; it reflects an unresolved mismatch between how neurological diagnostic tools are designed and the clinical environments in which they are needed most. The EEG systems and AI-assisted diagnostic tools currently available across SSA are predominantly developed within high-income clinical contexts, embedding assumptions around power infrastructure, specialist availability, connectivity, and dataset composition that do not reflect SSA healthcare realities. This paper proposes that context-aware neuroengineering, treating the specific conditions of SSA clinical environments as primary engineering design inputs rather than post-hoc implementation barriers, is a necessary reorientation for the field. Through a structured decomposition of SSA-specific engineering constraints including power infrastructure variability, offline-first connectivity requirements, cost and maintenance parameters, clinician capacity considerations, and the representational absence of SSA patient populations in published EEG datasets, we define what context-aware EEG diagnostic design demands in practice. These constraints are examined through a comparative analysis of four low-resource EEG diagnostic systems developed or deployed in SSA settings: the SEEDS-derived convulsive epilepsy diagnostic aid, a tablet-based EEG system deployed in the Republic of Guinea, a graph attention network system tested on recordings from Nigeria and Guinea-Bissau, and a published TinyML-based embedded AI seizure detection system developed for resource-constrained deployment but not yet validated within SSA clinical environments. Analysis of these cases surfaces consistent design principles that distinguish contextually grounded systems from adaptations of high-resource tools and reveals persistent gaps that the field is only beginning to address. The paper concludes by proposing directions toward a more comprehensively contextualised neuroengineering practice oriented to the realities of SSA healthcare systems.

ABSTRACT. Biomedical signal generation and calibration are essential components of healthcare quality assurance. Among the vital signs continuously monitored in clinical settings, cardiac electrical activity recorded by the electrocardiogram is of primary clinical importance, particularly in cardiology, anesthesia, intensive care, and emergency medicine. ECG device calibration requires a reference signal generator capable of reproducing both Normal and arrhythmic waveforms at controlled amplitude and timing parameters. However, comprehensive FPGA-based implementation covering multiple arrhythmic signal types on the Altera FPGA board platform is still a lap. In this study, ECG signals corresponding to Ventricular Tachycardia (210 bpm), Ventricular Paced (75 bpm), and Atrial Flutter (150 bpm) will be mathematically modelled, designed, and implemented on the Altera Cyclone V SoC Field Programmable Gate Array (FPGA) using VHDL in the Altera(Quartus) design environment. Each ECG waveform will first be modelled as a piecewise mathematical function and the resulting amplitude values will be pre-computed and stored within the FPGA Programmable Logic fabric. MATLAB-based ECG signals will serve as the reference, and the FPGA-generated outputs are validated by computing the Mean Squared Error (MSE) between MATLAB and Quartus simulation amplitude values at corresponding time points. This study demonstrates that FPGA-based ECG signal generation is achievable using VHDL on the Altera Cyclone V SoC platform, providing a hardware-implemented ECG calibration system applicable in low-resource healthcare settings.

ABSTRACT. In vitro models currently used to study cardiomyopathies fail to reproduce the structural and functional complexity of the human heart. To address this challenge, we are developing a three-dimensional heart chamber organoid based on a fibrin/Matrigel hydrogel molded around a Foley catheter to mimic a ventricular cavity. The construct, supported by 3D-printed structures and compatible with multiwell culture plates, encapsulates cardiomyocytes derived from human-induced pluripotent stem cells. The chamber is designed to accommodate a pressure sensor for investigating pressure-volume relationships, which cannot be obtained with conventional in vitro models. This approach lays the groundwork for the development of patient-specific cardiac organoids for the study of cardiomyopathies and the evaluation of personalized therapies.

ABSTRACT. Predictive models for continuous glucose monitoring in Type 1 Diabetes are often developed under a global paradigm, although recent work increasingly recognizes substantial inter-subject heterogeneity in glycemic dynamics and the benefits of personalized or patient-specific modeling. In this work, we investigate whether personalization can emerge through sparse, patient specific neural pathways within a shared architecture, opening the way to more parameter efficient and personalization oriented adaptive glucose prediction models. To test this hypothesis, we formulate a three class prediction task on the OhioT1DM dataset, targeting 30 minute ahead transitions toward hypoglycemia, normoglycemia, and hyperglycemia, and extract class specialized subnetworks through an iterative pruning strategy inspired by the Lottery Ticket Hypothesis. Beyond predictive performance, we analyze the resulting subnetworks through complementary structural and functional similarity measures, based respectively on spectral similarity and Gromov-Wasserstein distance. The results show that substantial sparsification can be achieved while preserving, and even improving, predictive performance, with optimal sparsity often concentrated around moderate pruning levels and reaching much higher reductions in favorable subjects. Structurally, hypoglycemia exhibits the strongest inter subject variability and the greatest deviation from cohort level behavior, whereas normoglycemic and hyperglycemic transitions remain more architecturally shared. Functionally, hyperglycemia emerges as the most subject dependent class, indicating that shared structural substrates may still support individualized processing. These findings suggest that personalization in glucose prediction may be more effectively achieved through a selective combination of shared sparse pathways and class specific subject dependent specialization, rather than through fully individualized models.

ABSTRACT. Protein-bound uremic toxins (PBUTs), such as indoxyl sulfate (IS) and p-cresyl sulfate (pCS), are poorly removed by conventional hemodialysis because of their strong affinity for plasma proteins, which reduces the effective diffusive driving force across the dialyzer membrane. Existing mechanistic models of PBUTs kinetics during hemodialysis rely on population-averaged parameters, thereby limiting their predictive capability at the individual patient level. In this work, a patient-specific extension of a previously validated mechanistic model is proposed. The model couples a three-compartment representation of the patient (plasma, interstitial, and intracellular compartments) with a one-dimensional dialyzer model, accounting for convection, diffusion, and reversible protein binding kinetics. Model parameters were individualized by integrating clinical data with in vitro measurements. The model was applied to a cohort of 20 hemodialysis patients treated with two different dialyzers, Solacea 21H and Cordiax FX100 with the online hemodiafiltration technique. Predicted total IS and pCS plasma concentrations showed good agreement with clinical measurements collected at multiple time points during dialysis sessions. Predictive performance was stable across patients, toxins, and treatment conditions, with absolute prediction errors below 20 mg/L and no significant differences between dialyzers or across time points. Overall, the proposed framework accurately reproduces patient-specific PBUTs kinetics and highlights the value of integrating experimental and clinical data for personalized modelling. This approach represents a step toward precision dialysis and may support the optimization of treatment strategies based on individual patient characteristics.

ABSTRACT. Type 1 diabetes (T1D) is a chronic disease characterized by elevated blood glucose (BG) levels caused by the absence of endogenous insulin secretion. In T1D management, corrective insulin bolus (CIB) strategies are particularly important to reduce hyperglycemia without leading to hypoglycemia. Most existing approaches for determining their timing and dosing remain reactive and rely on fixed heuristics that do not account for individual glucose–insulin dynamics.

In this work, we leverage the ReplayBG digital twin (DT) framework to compare three recent literature-based CIB rules with three proactive and personalized algorithms that incorporate a patient-specific inter-bolus timing, and prediction-based triggers: drCORRECT, rAR-CIB, and LSTM-CIB. All methods are evaluated in a unified DT-based in-silico environment using real-world free-living data from 30 adult patients, including continuous glucose monitoring data, meal logs, and insulin administrations. Performance is assessed using standard glucose control metrics, including time in range (TIR), time above range (TAR), and glycemia risk index (GRI). The results show that all CIB strategies improve glycemic control compared to baseline (TIR: 52.29%, TAR: 46.57%, GRI: 52.71), with the highest benefits obtained for proactive approaches. In particular, LSTM-CIB achieves the best overall performance (TIR: 65.29%, TAR: 33.26%, GRI: 33.64) while maintaining low hypoglycemia exposure. Compared to the literature-based methods, the considered proactive strategies consistently yield a higher TIR and a lower GRI.

Overall, integrating personalized timing with predictive triggers improves glycemic outcomes compared to baseline and rule-based literature approaches, highlighting the value of DT–enabled personalization for next-generation CIB strategies in T1D.

ABSTRACT. Access to digital healthcare remains a persistent barrier for older adults and individuals with disabilities, a challenge the Red Cross and TECSOS directly address through Orientatech, a curated assistive technology (AT) platform designed to foster independent living. Despite the availability of life-changing tools such as screen readers, cognitive support aids, and communication devices, people with disabilities frequently struggle to identify solutions matching their health and functional needs, risking exclusion from essential digital healthcare services. This paper presents Ori, a web-integrated conversational agent designed to promote independent living and improve quality of life, grounded exclusively in curated platform resources via a Retrieval-Augmented Generation (RAG) pipeline. Ori ingests platform data in PDF format, segments content into semantically coherent chunks, encodes them using SentenceTransformer embeddings (all-MiniLM-L6-v2), and retrieves context through a ChromaDB vector store. Retrieved content is passed to a locally hosted Ollama LLM, preserving data privacy and eliminating dependency on external cloud services. The system was evaluated across four task categories, product explanation, needs-based discovery, fallback handling, and personalized recommendation, achieving an average retrieval precision of 88.1%, response consistency of 93.7%, and mean latency of 1.85 s. A System Usability Scale (SUS) study with ten participants (ages 30–63, including users with visual, cognitive, and motor impairments) yielded an overall score of 88.7/100, with 90% rating the system as intuitive. Domain-specific design choices, including a constrained knowledge scope, multi-turn context accumulation, and accessibility-first interface principles, distinguish Ori from generic RAG deployments and align it with EN 301 549, WCAG 2.1, and digital accessibility requirements.

ABSTRACT. Cardiovascular diseases (CVDs) remain the primary cause of global mortality, highlighting the need for improved early detection strategies. While conventional risk calculators rely on standard clinical parameters, growing evidence suggests associations between retinal microvascular alterations and cardiovascular pathologies. Retinal imaging has emerged as a promising, non-invasive tool for cardiovascular risk assessment. In this study, we investigated the diagnostic utility of optical coherence tomography angiography (OCTA) combined with radiomics to identify individuals with CVDs, including congestive heart failure, stroke, and vascular diseases. The publicly available RASTA dataset was used, comprising 814 OCTA images from 491 participants. Multiple feature selection techniques and machine learning classifiers were evaluated to optimize performance. The final model, based on correlation filtering, elastic net regularization and logistic regression, achieved a balanced accuracy of 0.782 and an area under the receiver operating characteristic curve (AUC) of 0.808 on the test set. The model demonstrated high sensitivity (0.941) and successfully identified all cases of congestive heart failure and stroke. These findings support the potential of OCTA-based radiomics as a non-invasive screening approach for CVD detection, warranting further validation in larger cohorts.

ABSTRACT. Abstract. Obstructive sleep apnea (OSA) is a highly prevalent and heterogeneous sleep-related breathing disorder still primarily characterized using global severity indices derived from polysomnography, such as the apnea-hypopnea index (AHI). However, these indices only partially capture the underlying physiological variability across subjects. This study proposes a multimodal phenotyping framework based on subject-based correlation networks constructed from physiological features derived from pulse oximetry (SpO2) and heart-rate variability (HRV). Data from 2,641 subjects from the Sleep Heart Health Study were analyzed using a novel pipeline combining bootstrap resampling and consensus clustering for stable phenotype identification, and community detection based on modularity optimization for phenotypic discovery. Four clinically and physiologically distinct phenotypes were identified: (1) younger, lean individuals with mild OSA and preserved autonomic function; (2) older subjects with intermediate severity and a hyperreactive, potentially compensatory autonomic response; (3) younger, obese patients with severe hypoxia but paradoxically blunted autonomic activity; and (4) older, obese individuals with extreme hypoxic burden, high comorbidity, and maladaptive autonomic patterns. These phenotypes showed significant differences across 43 of 73 clinical variables and were predicted with a Cohen’s kappa of 0.878 on an independent test set. These results demonstrate that integrating SpO2- and HRV-derived features enables the identification of clinically relevant and highly separable OSA phenotypes.

ABSTRACT. Motor engagement during robot-assisted gait training (RAGT) is a key factor influencing rehabilitation outcomes, yet its objective and quantitative assessment remains challenging. Traditional evaluation approaches are mainly based on motor performance and provide limited information about the patient’s level of participation. In this study, we propose a non-invasive method to estimate motor engagement by combining visual behavioral features and biomechanical signals acquired during RAGT sessions. Joint torque data were collected from a Lokomat system, while gaze direction was extracted from frontal video recordings using computer vision techniques. An event-based framework was adopted, defining events as time windows in which gaze orientation exceeded a given threshold, indicating attention toward the task. Results showed that task-oriented gaze was associated with lower joint torque values, suggesting higher levels of active participation. This relationship was consistently observed across sessions. These findings support the use of simple video-based features combined with robotic measurements as a practical and objective proxy for engagement during RAGT.

ABSTRACT. Reverse osmosis (RO) membranes are extensively employed in medical water purification systems, particularly for hemodialysis, where the production of ultrapure water is key to patient safety. This paper provides a comprehensive analysis of RO applications in dialysis facilities, describing the multistage purification process and evaluating its effectiveness in removing chemical and microbiological contaminants. A case study from a dialysis unit in Morocco demonstrates the system's stable performance. Current challenges encompass membrane fouling, high energy consumption, and complex maintenance requirements. To mitigate these issues, the integration of Internet of Things technologies enables real time monitoring and predictive maintenance, while renewable energy can be used to power high-pressure pumps. These strategies support the development of resilient, efficient, and sustainable dialysis infrastructures, especially in regions facing water scarcity. The paper underscores the importance of continued advancements in membrane technology and system optimization to meet the growing demand for safe and environmentally responsible dialysis treatment.

ABSTRACT. Abstract: In this work, we study the physical phenomena and the application of HIFU (High Intensity Focused Ultrasound) using both simulation and experiment. The experiment was done with a bowl-shaped focused ultrasound system. High speed videos of the generation of cavitation bubbles at the focal point of the HIFU are captured. We observed interesting bubble cloud structures, bubble movements and stationary bubble oscillations. The system is then applied for dentistry applications. Firstly we use the HIFU to drive antibacterial nanoparticles into dentinal tubules for disinfection in root canal treatment. Initial results show delivery of these nanoparticle deep into the dentinal channels which are a few microns in size. Next we cultivate E. Faecalis biofilm, a common bacteria colony found in the mouth and teeth, on petri dish and in human tooth. Then we subject them for removal under strong HIFU for a period of time. We obtain positive results in the biofilm removal with the increasing HIFU sonification time. Separately, we simulate the interaction of a bubble with pulsed ultrasound and in an ultrasound field near bio-materials such as fat, muscle and bone. The simulation shows extreme growth and collapse of the bubble under certain conditions. It is found that the formation and direction of the water jet during bubble collapse is highly dependent on the properties of the bio-materials nearby. These studies provide a foundation for better understanding of HIFU and its uses in medical treatment.

ABSTRACT. A reliable, continuously updated medical equipment (ME) inventory is the foundation of any modern Healthcare Technology Management (HTM) system. The WHO 2022 Global Atlas of Medical Devices reports that most countries lack such inventories at national scale. We describe the implementation of a Unified Medical Equipment Management System (MEMS) covering the Greek public healthcare sector, delivered as part of a national digital transformation programme. The work covers two coupled processes: a standardised on-site registration of every electrically powered, maintenance-relevant device (room-by-room tagging, photographic documentation, mobile data capture), and a structured data-entry workflow that converts image-based field records into clean inventory entries on the web-Praxis platform. To accelerate data entry we developed an AI-assisted matching tool that proposes manufacturer–model–GMDN “Triplets” directly from device photographs, with a three-tier human review pyramid for quality control. The project covered 129 public hospitals, registering 170,860 devices, and assigned GMDN codes to 26,156 distinct models drawn from 1,143 GMDN Terms (197 Collective Terms). The approach is reproducible and scalable, and provides an evidence base for centralised contracting, harmonised maintenance and equipment redistribution.

ABSTRACT. Medical device maintenance is essential to ensure patient safety and the continuity of healthcare services; however, in developing countries such as Peru, it is often affected by fragmented technical information, lack of structured training processes, and high turnover of biomedical interns. These limitations reduce learning efficiency and increase the risk of operational errors. This study presents the design and preliminary evaluation of a digital training platform aimed at improving knowledge transfer in clinical engineering environments. The proposed system integrates structured audiovisual modules and centralized access to technical documentation, organized according to the institutional inventory of biomedical equipment. A pilot implementation was conducted in a real clinical setting, focusing on ultrasound equipment as a case study. The evaluation involved biomedical interns and assessed usability, access to information, and perceived support for learning. Results indicate a positive user perception, with all participants reporting that the platform is accessible and easy to use. These findings support the feasibility of the proposed approach as a tool for structuring training processes and improving access to technical knowledge. Although the results are limited to a preliminary evaluation, the platform provides a foundation for future studies aimed at assessing its impact on learning outcomes and operational performance.

ABSTRACT. Mechanical ventilation is a cornerstone of advanced critical care, providing essential respiratory support to patients experiencing respiratory failure, severe pulmonary conditions, or neurological impairment. Despite its clinical significance, access to ventilatory support remains severely limited in resource-constrained healthcare settings, particularly in low- and middle-income countries. Conventional mechanical ventilators are prohibitively expensive, bulky, and require specialized infrastructure and trained personnel, rendering them inaccessible in district hospitals, ambulances, rural health centers, and during mass casualty events. The COVID-19 pandemic further exposed a critical global shortfall of approximately 880,000 ventilators, underscoring the urgent need for affordable and portable alternatives. This study presents the design and development of a low-cost automated ventilation system intended for emergency respiratory support in resource-limited environments. The system employs a motor-driven mechanism to automate the compression of an Ambu bag (self-inflating resuscitation bag), thereby eliminating reliance on manual bag-mask ventilation, which is inherently imprecise and operator-dependent. The design prioritizes portability, affordability, and scalability through the integration of lightweight, readily available, and cost-effective components, making the device suitable for deployment across a broad range of clinical and pre-hospital settings. Performance testing demonstrated the system's capability to deliver controlled breaths through precise timing mechanisms and current regulation. The device is designed to be adaptable for future upgrades, including the integration of patient monitoring modules. Performance testing was conducted to validate the system's functionality and reliability as an emergency ventilatory aid. The outcomes of this project demonstrate the feasibility of producing a portable, automated ventilation device that significantly reduces production costs while maintaining clinical efficacy. This system has the potential to bridge the critical gap between the high demand for emergency ventilatory support and the limited availability of conventional ventilators in under-resourced healthcare systems, including those in sub-Saharan Africa. The scalability of the design further positions it as a viable solution for mass production in response to public health emergencies.

ABSTRACT. Forcecardiography (FCG) is a novel cardiomechanical monitoring technique based on a non-invasive, small monolithic force sensor that captures respiration, heart walls motion, infrasonic heart valves vibrations, and heart sounds, all simultaneously from a single contact point onto the chest. This work presents an overview of the principles, performance, and clinical significance of FCG. It also discusses the relationship between FCG and other well-established cardiomechanical signals. Forcecardiography could pave the way for a comprehensive evaluation of cardiac mechanics via wearable devices for continuous long-term monitoring of many clinical indices like cardiac time intervals, stroke volume, cardiac output, ejection fraction, myocardial contractility, heart valves performance. Therefore, FCG could provide critical improvements in patients follow-up for a variety of chronic cardiovascular and pulmonary diseases, especially in non-clinical settings.

ABSTRACT. Heart rate variability (HRV) has traditionally been analyzed from electrocardiograms, but recent progress in microelectromechanical (MEMS) sensors and signal processing has made it increasingly feasible to estimate HRV from mechanocardiographic signals, including ballistocardiograms, seismocardiograms, and gyrocardiograms. This narrative review discusses 18 studies in which cardiomechanical signals were used for HRV-based health- and state-related assessment. The reviewed literature was grouped into four main areas: cardiovascular diseases, posture and activity, music-related effects, and alcohol-related impairment assessment. The most mature evidence focuses on cardiovascular applications, especially atrial fibrillation detection, where arrhythmia is directly reflected in mechanically derived beat-to-beat timing series. Mechanocardiographic HRV was also informative for distinguishing healthy subjects from patients with valvular heart disease, although more specific disease phenotyping generally required additional morphology-based or multimodal features. In posture-, activity-, and sleep-related applications, interval-derived HRV indices appeared more robust than raw morphology-sensitive descriptors, particularly under low-motion or controlled conditions. Music- and alcohol-related studies suggest that mechanocardiographic HRV can also reflect subtler physiological or behavioral state changes, but the evidence in these areas remains preliminary. Overall, HRV analysis based on mechanocardiograms is a feasible extension of conventional ECG-based HRV analysis, but broader adoption will require more reliable ECG-independent heartbeat detection, larger and better-balanced cohorts, public benchmark datasets, standardized processing pipelines, and stronger external validation.

ABSTRACT. Morphic sensors based on electro-resistive bands (ERBs) represent a robust and viable solution for ubiquitous, non-invasive physiological monitoring. These sensors, which function as distributed strain-gauges, can be easily integrated into standard wearable garments providing continuous surveillance of cardio-respiratory signals without the need for direct skin contact or obtrusive electrode arrays. Our design tested against medical gold standards show that clinically relevant parameters like tidal volume, heart rate and respiratory rate can be extracted with sufficient precision during sleep as well as light physical activity

ABSTRACT. Forcemyography (FMG) has recently emerged as a promising technique to monitor muscle activity level by sensing the mechanical changes related to contraction. Traditionally, surface electromyography (EMG) has been widely adopted for such applications; however, this technique requires reliable electrical contact and is subject to various issues, such as electrode stability, electromagnetic interference, and the need of raw EMG data processing to extract the envelope. This study presents a comprehensive overview of the evolution of FMG, from muscle contraction monitoring to its application in real-time Human–Machine Interfaces (HMIs). The FMG signal is typically obtained using force or pressure sensors placed on the skin near the muscle. During contraction, the muscle cross-section increases and the tendons tighten, exerting a radial force that can be recorded onto skin surface. Several investigations demonstrated a strong correlation between FMG and the EMG linear envelope. Although electrical phenomena always precede mechanical contraction, the FMG signal does not exhibit a noticeable delay relative to the EMG envelope, which requires processing that inevitably introduces a time delay in real-time applications. Obviously, FMG cannot capture the complexity of individual motor unit recruitment but generates an overall signal of muscle contraction. Many real-time applications, such as HMI control, rely exclusively on this overall signal and can benefit from the FMG technique, which offers the advantages of extreme ease of use, the availability of very thin sensors, immunity to electromagnetic interference, and a signal that requires minimal or no processing. Novel experiments on isometric contractions have further demonstrated that the FMG is capable of accurately estimating muscle force through regression models. Finally, the applicability of FMG in HMIs is reviewed through representative implementations, including gesture recognition interfaces, prosthetic hand control, and wearable hand exoskeletons. Overall, FMG demonstrates significant potential as a robust, low-cost alternative to EMG, particularly suited for real-time interfaces.

ABSTRACT. Designing reliable AI pipelines for lung cancer screening requires addressing reproducibility, scalability, and methodological pitfalls. Variability due to non-deterministic training, heterogeneous computing environments, and differences in data handling can substantially affect model performance, making results difficult to replicate. Moving from single-server experimentation to high-performance computing —combined with containerization and dataset versioning—enables controlled, scalable, and more reproducible workflows. At the same time, computational efficiency remains a limiting factor in large-scale CT analysis, and data leakage can artificially inflate reported performance if not rigorously controlled. We present experimental results on reproducibility and computational efficiency using the Sybil deep learning model on the National Lung Cancer Screening Trial (NLST) dataset (train: 9,646 patients, 28,053 samples [26,621 negative, 1,432 positive]; test: 2,203 patients, 6,277 samples [5,978 negative, 299 positive]). We observe high robustness to class imbalance on the negative class: using only 20% of negative samples reduces training time by over 50% with negligible performance loss. In contrast, reducing positive samples degrades performance without computational benefits due to oversampling effects. We further analyze Sybil’s predictive behavior across clinically relevant factors. While risk estimates are consistent across reconstruction kernels, performance varies significantly with cancer visibility: at a 3-year prediction window, AUC reaches 0.936 for prevalent nodules, decreasing to 0.758 for pseudo-incident and 0.508 for incident cases, with similar trends up to 6 years. Overall, this work emphasizes that achieving clinically meaningful AI systems requires treating reproducibility and scalability as design objectives, rather than retrospective checks, integrating them directly into pipeline development and validation.

ABSTRACT. Lung cancer screening using low-dose computed tomography (LDCT) has been shown to reduce mortality significantly. Furthermore, integrating automated artificial intelligence (AI) models for classifying lung nodule malignancy can provide radiologists with valuable second opinions faster than histopathologic malignancy confirmation. However, these deep learning (DL) models frequently struggle to generalise across different datasets and clinical institutions due to domain shift caused by differences in scanners, acquisition protocols, hospitals, and patient demographics. In addition to limited access to data for model training, challenges originate from heterogeneous data sources, with inconsistent annotation practices, label semantics, and limited harmonisation of protocols across datasets. In this talk, we examine how these challenges manifest in real-world settings, drawing on examples from both public and private datasets, and show how inconsistent annotations and heterogeneous data sources further degrade model performance. Additionally, we will discuss the importance of data harmonisation and domain-generalisation strategies in creating generalisable, robust, and scalable AI solutions well-suited to the realities of lung cancer screening using LDCT scans.

ABSTRACT. The clinical implementation of artificial intelligence (AI) in low-dose computed tomography (LDCT) lung cancer screening is frequently hindered by heterogeneous data and the absence of standardized reference space. While neuroimaging has long benefited from established spatial templates and rigorous data standards, thoracic imaging has historically lacked these foundational tools. This presentation explores the strategic adaptation of proven neuroimaging best practices in lung cancer screening. First, we addressed data standardization by introducing a proof-of-concept extension of the Brain Imaging Data Structure (BIDS) to CT imaging, applying it to the National Lung Screening Trial (NLST) dataset, and developing a BIDS-App. Second, we tackle anatomical standardization by constructing a high-resolution, 3-D lung template and probabilistic lobar atlas derived from the NLST cohort, utilizing the Advanced Normalization Tools (ANTs) ecosystem and FSL, a staple in neuroimaging. This atlas enabled voxel-wise analysis to map specific deformation patterns associated with varying emphysema severities. Furthermore, it facilitated the creation of a lung nodule atlas, highlighting the distinct spatial locations of malignant versus benign pulmonary nodules within the cohort. By bridging the gap between neuroimaging methodologies and thoracic AI, we introduce robust, open-source tools to advance and standardize lung imaging research.

ABSTRACT. Lung cancer screening programs can leverage Deep Learning (DL) models as a supplementary tool for radiologists to detect and classify lung nodules. Various studies in the literature have applied high-performance DL models to the LIDC-IDRI dataset; however, the robustness of these models is affected by the quality of the dataset annotations. This study seeks to deepen our understanding of how label noise and concept shift affect the cross-dataset generalisation of DL models for lung nodule malignancy classification. The focus was on using LIDC-IDRI, which was labelled by radiologists with a "likelihood of malignancy" ($Y_{rad}$) rather than biopsy-confirmed labels ($Y_{bio}$), and which presents annotations with high interobserver variability. The findings revealed that models trained on $Y_{rad}$ labels struggle to generalise to datasets with $Y_{bio}$ labels. To address this issue, we propose a model-driven relabelling strategy for LIDC-IDRI that uses an ensemble of 2.5D~\textit{ResNet-18} trained on the LUNA25 dataset with biopsy-confirmed labels. This strategy demonstrated a notable improvement in cross-dataset generalisation for the classification models, achieving a 10\% increase in AUC on an external dataset while also improving performance on the internal dataset. In summary, the results underscore the critical importance of addressing annotation quality and developing strategies to enhance the generalisation of DL models across diverse clinical data sources in the context of lung cancer screening.

ABSTRACT. “AI” refers to applied, system-level use of machine learning in healthcare environments, where it intersects with the following: • Medical devices • Hospital infrastructure • Data systems • Safety and compliance 1. Intelligent Health Technology Management (HTM) • Predictive Maintenance of Medical Equipment o AI is used to monitor and predict equipment failures.  Inputs: Sensor data (temperature, voltage, usage cycles)  Outputs: Alerts before breakdowns happen • HTM Workflow & Operational Tools o Digital systems that improve hospital efficiency.  Asset tracking (RFID, RTLS)  Maintenance management systems (CMMS)  Post-market device surveillance (PMS) 2. Distributed Care • Telehealth & Remote Care Platforms o Digital tools that deliver care outside the hospital.  Video consultations, remote diagnostics, e-prescriptions  Integration with hospital systems (EHR, PACS) o Remote Patient Monitoring (RPM) & Wearables  Connected devices that continuously collect patient data. • Vital signs: heart rate, SpO₂, ECG, blood pressure • Chronic disease monitoring (diabetes, cardiac conditions) o Examples of device types (not brands):  Smartwatches with ECG  Home-use pulse oximeters  Connected glucose monitors o Mobile Health (mHealth) Applications  Patient-facing apps for health management. • Medication reminders • Rehabilitation tracking • Mental health monitoring 3. Clinical Applications • Cancer a) Health promotion, i.e., keeping people healthy and therefore preventing cancer b) Early detection &amp; diagnosis before the presence of clinical signs, via screening and biomarker-based technologies c) Treatment of cancer and its follow-up d) Support if cancer is no longer curable • Imaging AI models (often deep learning) analyze: o CT, MRI, X-ray images o Ultrasound scans • Use cases: o Tumor detection o Fracture identification o Stroke diagnosis (rapid triage) o Mammography o Colonoscopy o Pancreatic cancer • Clinical Decision Support Systems (CDSS) o AI assists clinicians in making decisions. • Examples: o Sepsis prediction o Drug interaction alerts o Risk scoring for ICU patients iv. Digital Diagnostics o Software tools that assist in diagnosis or care planning. o Symptom checkers o AI-assisted triage systems o Scheduling and bed management tools 4. Digital Health (DH) Tools and Interoperability • Electronic Health Records (EHR) &amp; Data Integration Core digital infrastructure for managing patient data. o Centralized patient records o Integration across departments (lab, imaging, pharm.) • Workflow Optimization & Hospital Operations AI helps optimize hospital efficiency. o Patient flow prediction o Operating room scheduling o Resource allocation • Smart Medical Devices &amp; IoMT (Internet of Medical Things) AI is embedded directly into devices. A network of connected medical devices and systems. Examples: o Smart infusion pumps o Wearables (heart rate, oxygen monitoring) o AI-enabled ventilators o Bedside monitors, imaging systems o Real-time data sharing across hospital networks “Digital Health” refers to the ecosystem of connected technologies that support care delivery, including: • Devices (hardware) • Platforms (software) • Networks (infrastructure) • Data systems (integration & analytics) Cybersecurity & Data Privacy AI + Cybersecurity in Healthcare Systems • Since dataset explicitly links AI with cybersecurity, this is important. • AI is used to: o Detect unusual network/device behavior o Identify potential cyberattacks on medical devices o Digital health greatly expands the attack surface. o Protection of patient data o Securing connected devices and networks o Compliance with regulations (HIPAA, FDA guidance)

6. Regulatory, Risk, Safety and Forensics • Regulatory, Safety, and Quality Issues A major theme in CE publications: o Ensuring AI systems meet standards (e.g., FDA, IEC) o Validating algorithms on real-world data o Managing risks (false positives/negatives) with embedded AI and System of Systems

ABSTRACT. Artificial Intelligence (AI) is rapidly transforming healthcare, with most attention focused on clinical applications such as diagnosis, decision support, precision medicine, and predictive analytics. While these developments are undoubtedly important, they represent only part of the value that AI can generate within healthcare systems. A less explored, yet equally impactful domain concerns the application of AI to clinical engineering, healthcare operations, and medical device management.

Healthcare organizations are increasingly challenged by growing complexity, resource constraints, workforce shortages, and the need to manage large fleets of interconnected medical technologies. Clinical engineers play a pivotal role in ensuring the safety, effectiveness, availability, and sustainability of medical devices and healthcare infrastructures. However, many of the processes supporting these activities remain labor-intensive, fragmented, and reliant on manual decision-making.

This presentation will explore how AI can support clinical engineering functions beyond traditional clinical applications, highlighting real-world proof-of-concept (PoC) solutions developed within European research and innovation projects. These initiatives demonstrate how AI can enhance operational efficiency, improve resource allocation, support decision-making, and contribute to safer and more resilient healthcare systems.

The first group of applications focuses on process mining and healthcare workflow optimization. By combining AI techniques with data extracted from hospital information systems, it becomes possible to identify inefficiencies, bottlenecks, and deviations from intended workflows, enabling evidence-based redesign of clinical and technical processes.

A second area concerns crowd management and patient-flow optimization. AI-driven models can analyze movement patterns, predict congestion, and support dynamic allocation of resources in complex healthcare environments. Such approaches have demonstrated potential to improve patient experience, increase operational efficiency, and enhance preparedness during periods of high demand.

The presentation will also discuss AI-supported infection prevention and control strategies. Machine learning algorithms can integrate heterogeneous data sources to identify risk patterns, support surveillance activities, and provide early warning mechanisms that contribute to reducing healthcare-associated infections and improving patient safety.

Another important application area involves the micro-logistics of third-party managed medical devices. Hospitals increasingly rely on externally owned or managed devices, creating challenges related to traceability, availability, utilization, and inventory management. AI-based solutions can support real-time monitoring, demand forecasting, and optimized allocation of these resources, reducing waste and improving service continuity.

The role of AI in advanced document management will also be presented. Clinical engineering departments manage large volumes of technical documentation, maintenance records, regulatory evidence, procurement documents, and manufacturers’ information. Recent advances in natural language processing and large language models enable intelligent document retrieval, automated knowledge extraction, compliance support, and enhanced decision-making throughout the device lifecycle.

Finally, the presentation will examine emerging applications of AI for workforce training and competency development. Intelligent tutoring systems, adaptive learning platforms, and AI-powered simulation environments can support continuous education of healthcare professionals and clinical engineers, facilitating knowledge transfer and improving training effectiveness.

Drawing on practical experiences and lessons learned from European projects, this presentation aims to broaden the discussion on AI in healthcare by highlighting the significant opportunities that exist outside the traditional clinical domain. The examples presented demonstrate how AI can become a strategic enabler for clinical engineering, contributing not only to technological innovation but also to operational excellence, sustainability, and resilience across healthcare systems.

ABSTRACT. A presentation reviewing major developments in Biomedical Engineering over the past forty years through the perspective of personal participation in international projects and initiatives spanning research, education, regulation, clinical engineering, and health technology assessment. From the early European policy initiatives of the 1980s, including the “New Approach” that paved the way for the Medical Devices regulatory framework, to the growth of transnational Biomedical Engineering education from the 1989 European postgraduate MSc program in BME, to the application of AI in BME education, and from the strengthening of IFMBE structures in Clinical Engineering and HTA, to the establishment of EAMBES and more recently the GCEA, the field has progressively expanded into a globally connected, multidisciplinary domain. The presentation links these milestones with direct experience in European and international activities, including participation in the European Commission Medical Devices Committee, leadership roles in IFMBE, the creation of INBIT in 1991 that has been designated this year (2026) as a WHO Collaborating Centre for Medical Equipment Management and Assessment. The aim is to illustrate how the evolution of Biomedical Engineering has been shaped not only by scientific and technological progress, but also by international collaboration, educational innovation, institutional development, and policy engagement.

ABSTRACT. The second edition of the *Encyclopaedia of Medical Physics* aims to provide free, high‑quality educational resources for medical physics students and educators worldwide through an open‑access website. Building on its 2010 first edition, the updated Encyclopaedia expands its scope with substantial new material on Clinical Engineering for imaging and radiotherapy equipment, developed in collaboration with IFMBE specialists. Between 2019 and 2021, the entire online content underwent major revision, incorporating about 25% new material reflecting technological advances since 2010. The current edition includes 3,300 peer‑reviewed, cross‑referenced articles, over 600 abbreviations, around 1,000 updated entries, and approximately 2,000 illustrations across more than 4,000 printed pages. Contributions came from 150 experts across 30 countries, supported by a dedicated Clinical Engineering subgroup. The Encyclopaedia serves both as a graduate‑level teaching resource and a comprehensive reference, with strong emphasis on CE–MP collaboration throughout the equipment lifecycle. Its multilingual Scientific e‑Dictionary translates 4,000 terms into 32 languages, supporting global education. The redesigned website, accessible on mobile devices, offers advanced search modes and has attracted up to 40,000 monthly global enquiries. The project represents the first freely available, large‑scale educational platform of its kind in medical physics.

ABSTRACT. This presentation will examine how large language models and other AI tools are reshaping the creation and evaluation of scholarly content in Biomedical Engineering and Medical Physics. It will map where AI now enters the research and publishing cycles, highlight the implications for article and book development, and identify the emerging challenges for quality control, originality, and accountability.

ABSTRACT. As the "trusted link" between disruptive laboratory innovation and resilient clinical care, Biomedical Engineering (BME) has evolved into the essential pillar of European healthcare modernization and strategic autonomy. No longer confined to a purely technical background role, the profession now stands at the decisive intersection of medicine and technology, guiding innovation to ensure it serves humanity ethically, safely, and sustainably. This talk will explore the trajectory of the European Alliance of Medical and Biological Engineering and Science. Representing half of the global BME community within the IFMBE, EAMBES has successfully positioned the profession as a vital interlocutor in European policy-making. Key achievements to be discussed include the establishment of the European Parliament Interest Group on BME (EPIG BME) and the alliance's proactive role in shaping landmark regulations such as the Medical Device Regulation (MDR), the AI Act, and the European Health Data Space (EHDS). Special focus will be given to "Concept to Impact" through EAMBES projects, including but not limited to, the EMBE Innovation Consortium (EMBEic): Accelerating the internationalization of BME research through expert matchmaking. The TRUST Project: Leveraging trustable AI to automate regulatory compliance for medical research, reducing the administrative burden on biobanks and clinical units. Professionalization Mandates: The urgent EAMBES-led initiatives for the formal certification of BME as a regulated profession and the harmonization of clinical engineering curricula across European hospitals to ensure patient safety and technical authority. By leveraging the Quintuple Helix model—linking science, industry, policy, society, and the environment—EAMBES is defining a future where BME serves as the central nervous system of healthcare, translating scientific discovery into practical, sustainable, and universally accessible well-being.

ABSTRACT. The rapid evolution of diagnostic imaging technologies continues to reshape modern healthcare, demanding comprehensive professional guidance and harmonized educational curricula. The new edition of the IAEA Diagnostic Radiology Physics - a Handbook for Teachers and Students provides updated content covering both classical and advanced resources for medical physicists, radiologists, technologists, and related disciplines. The second edition of the handbook extends beyond the classical applications and imaging modalities like radiography, mammography, interventional, computed tomography, dental, ultrasound and magnetic resonance imaging. Enhanced coverage of emerging technologies, advanced imaging systems, image quality assessment, radiation protection, information technology, and equipment management. The handbook also expands on digital technologies, reconstruction algorithms, and the integration of artificial intelligence into imaging workflows, reflecting the field’s shift toward data driven and automated solutions. Additionally, the handbook provides practical insights into clinical implementation and foundation of diagnostic radiology, and the basics of imaging anatomy, bridging the gap between physics principles and diagnostic practice. In general, the new edition of the Diagnostic Radiology Physics handbook serves as a timely and forward looking reference that encapsulates the technological progress, safety frameworks, and clinical application related to the contemporary diagnostic radiology.

ABSTRACT. The European regulatory landscape for medical devices has undergone profound transformation with the full application of Regulations (EU) 2017/745 (MDR) and 2017/746 (IVDR), which introduced more stringent requirements on safety, clinical evidence, performance evaluation, post-market surveillance, and traceability through the UDI system. In parallel, Regulation (EU) 2021/2282 on Health Technology Assessment (HTAR), applicable since January 2025, has introduced a European coordination mechanism through Joint Clinical Assessments (JCAs), progressively extending from oncology drugs and advanced therapies to class IIb/III medical devices and class D in vitro diagnostics from 2026. This presentation will illustrate the regulatory frameworks of medical devices and the new related challenges to obtain the CE marks. The speaker will also discuss how Italy has transposed this evolving framework through Legislative Decrees 137/2022 and 138/2022, redefining national governance of medical devices by introducing new obligations regarding vigilance, traceability, and multidimensional assessment of health technologies.

ABSTRACT. Health systems across Europe have developed markedly different approaches to the assessment of health technologies, each shaped by distinct institutional traditions and governance priorities. This presentation will provide a comparative analysis of consolidated European HTA systems, examining NICE in the United Kingdom, the dual G-BA/IQWiG model in Germany and the HAS in France. The speaker will highlight shared features underpinning these mature systems, over two decades of operational continuity, high productive capacity, robust integration of HTA into reimbursement and procurement decisions, scientific independence, and methodological transparency, while also discussing their structural differences, from NICE's pragmatic cost-per-QALY thresholds to the German model's rapid hospital access pathways. Drawing transferable lessons from these European experiences, the speaker will examine Italy's current position with the National HTA-MD Programme (PNHTA-DM), discussing its governance architecture and funding mechanism, and identify which elements of these European systems, in terms of institutional design, evaluator plurality, and integration with procurement, could inspire a more responsive and scalable Italian HTA model.

ABSTRACT. Faced with a growing structural gap between the demand for HTA of medical devices and Italy's current evaluative capacity, SIHTA has developed a novel governance model, the Subsidiary HTA system, designed to complement and strengthen the current national PNHTA-DM framework. This presentation will discuss the five founding pillars that guide the definition of the Subsidiary HTA model: subsidiarity with respect to the public governance system, institutional authority guaranteed by central accreditation and oversight, timeliness aligned with medical device life cycles, a liberal approach based on regulated competition among accredited evaluators, and accountability towards all stakeholders. The speaker will then present the new Subsidiary HTA system based on the strategic distinction between HTA Doers, the new Subsidiary Bodies for HTA of medical devices responsible for producing HTA assessments, operating under the strategic supervision of the Ministry of Health and the technical coordination of AGENAS, and the HTA Users, namely, healthcare organizations, contracting authorities, and other decision-making entities that use HTA assessments for appraisal and adoption decisions. The presentation will outline the rationale, objectives, operational mechanisms, system architecture, roles, responsibilities, and functional relationships among the key actors of the HTA-MDs ecosystem and expected benefits of the proposed model.

ABSTRACT. Once HTA evidence has been produced, its value depends on how effectively it is translated into procurement, adoption, and monitoring decisions at national, regional, and local level. This presentation will focus on the perspective of HTA Users within the Subsidiary HTA system proposed by SIHTA. The speaker will illustrate how validated HTA reports inform technical evaluation criteria in public tenders under Legislative Decree 36/2023, enabling value-based procurement models in which clinical outcomes, organisational impact, and total cost of care are translated into objective scoring criteria, moving decisions beyond price alone. The presentation will further address the adoption phase, in which Regions and healthcare organisations plan procurement volumes and integrate HTA evidence into care pathways, and the subsequent monitoring phase, where real-world clinical, organisational and economic outcomes feed back into national information systems coordinated by the Ministry of Health. The speaker will discuss how this real-world data supports periodic reassessment of technologies, including the review of disruptive innovation status, and will argue for strengthening HTA appraisal capacity within healthcare organisations as a prerequisite for translating national evidence into responsive, context-sensitive local decisions.