Electrical Impedance Spectroscopy (EIS) has emerged as a promising non-invasive technique for skin characterization and early detection of dermatological pathologies. However, the reliability and repeatability of EIS measurements are strongly influenced by the electrode-skin interface, particularly by electrode material and geometry. This work presents an experimental evaluation of the variability and repeatability of skin impedance measurements acquired using three different electrode types. Impedance spectra were recorded at three distinct locations on the biceps brachii of a healthy male subject under carefully controlled environmental conditions. For each electrode and measurement site, ten consecutive measurements were performed over a frequency range from 10 Hz to 100 kHz using identical acquisition protocols. Statistical analyses were conducted to assess intra-point repeatability, site-level consistency and inter-electrode variability, including tests for normality, homoscedasticity and non-parametric comparisons. The results showed weak statistically significant differences in the impedance modulus among the electrode configurations, resulting in a significance level of 0.05. These findings highlight the impact of electrode characteristics on measurement dispersion and support the importance of standardized electrode selection in EIS-based skin assessment to improve measurement comparability and reproducibility across experimental setups and future clinical or research-oriented EIS applications.
CANGURO software was developed to complete a workflow to obtain joint and pelvic angles time series as well as derived gait parameters. Original data is obtained from photogrammetry using Cortex 10 system to feed CANGURO software. CANGURO is an original tool developed at the MOVHUM laboratory to support the identification of gait abnormalities and rehabilitation planning. Data processing is based on the result of reviewing key methodological milestones. From the Cortex-10 derived coordinates time series of every segment, CANGURO calculates the heel strike and toe off events, based on the position and acceleration of vertical movement of the heel and toe. CANGURO then calculates the mean gait parameters over 22 strides of self selected walking speed: speed, stride and step lengths, cadence, and swing-to-stride-time duty factor. The clinical report by CANGURO includes also the hip, knee, ankle and foot progression angles time series compered to the selected reference range. CANGURO is multilingual using a central table with one column per language. Finally, representative results from a recorded patient are presented, and future lines of work are discussed. CANGURO is being used to obtain reference values for the population of Uruguay. The first gait parameters obtained for 8 heathy participants are speed $1.21 \mathrm{~m} / \mathrm{s} \pm 0.32$, stride length $127.8 \mathrm{~cm} \pm 18.8$, stride time $1.07 \mathrm{~s} \pm 0.14$, cadence $113.0 \pm 15.0$ step/min, step length $64.2 \mathrm{~cm} \pm 10.0$, step time $0.54 \mathrm{~s} \pm 0.9$, swing time $0.49 \mathrm{~s} \pm 0.08$ and duty factor $45.5 \mathrm{~s} / \mathrm{s} \pm 3.2$. Gait Analysis workflow is limited by the operator interactive time with Cortex 10 to reconstruct all missed markers. Workflow on average takes one hour per patient.
Accurate perforator selection is critical for successful deep inferior epigastric artery perforator (DIEP) flap breast reconstruction. While computed tomography angiography (CTA) provides high-resolution anatomical detail, it lacks functional perfusion assessment. Dynamic infrared thermography (DIRT), in contrast, captures superficial rewarming dynamics but lacks objective anatomical validation. This study presents a standardized multimodal measurement system enabling automated fusion of DIRT with CTA-derived perforator coordinates. The framework integrates controlled thermographic acquisition, realtime hotspot detection, metric spatial calibration, and global one-to-one assignment of thermal hotspots to CTA perforators within a predefined periumbilical region of interest. The workflow was evaluated in 12 patients. Of 46 CTA perforators within the region of interest, 84.8% were successfully linked to thermographic hotspots, with a median spatial offset of 14.74 mm. Automated assignment excluded 76.1% of detected hotspots due to lack of anatomical correspondence, thereby reducing subjective interpretation. The proposed system enables quantitative multimodal validation of functional and anatomical imaging and provides objective support for perforator selection in clinical practice.
Left ventricular elastance is a measure that characterizes the mechanical behaviour of the left ventricle while it pumps blood. Elastance is derived from pressure and volume measurements obtained via invasive catheterization. Non-invasive approaches to study elastance typically involve determining ventricular volume and pressure through imaging techniques such as magnetic resonance imaging or echocardiography or use mathematical models to derive these parameters. In this study, we propose a novel method for non-invasively estimating elastance using measured carotid artery distention data, ejection fraction, cardiac output, and seismocardiogram data. Parameters derived from the sensed outputs are carotid pressure, ventricular volume at the end of systole and diastole, and the timing of left ventricular valves’ opening and closing. These parameters are integrated into the Double Hill function model for computing left ventricular elastance. Using the proposed method, time-varying elastance curves were generated for 20 subjects (9 females and 11 males). Visual analysis of the normalised elastance plots indicates that the shape of the estimated elastance curve aligns well with expected physiological patterns. Additionally, the simulated elastance plots derived from the model exhibit strong beat-to-beat repeatability over 20 cardiac cycles of a single participant.
This paper presents Backpack EKG, a standalone, energy-autonomous three-lead electrocardiography (EKG) platform tailored to decentralized cardiovascular screening in offgrid and resource-limited environments. Unlike long-term ambulatory monitors or diagnostic 12-lead systems, the proposed device targets short-duration, repeatable acquisitions within outreach and first-contact workflows, where electrical infrastructure, connectivity, and external computing devices cannot be assumed. The platform integrates biopotential acquisition, embedded real-time processing, on-device waveform visualization, and fully offline data logging in a robust backpack-deployable form factor. A three-lead configuration is obtained from two measured leads plus a digitally derived lead (Einthoven relation), enabling rhythm inspection and heart-rate estimation while reducing hardware complexity and power demand. Energy autonomy is achieved through a rechargeable lithium-ion battery with photovoltaic-assisted charging, supporting up to 10 h of continuous operation and approximately 120 screening acquisitions per charge. Heart-rate estimation was verified using a calibrated patient simulator across $30-320$ BPM setpoints, achieving a mean absolute error of 0.19 BPM (mean relative error $\lt0.1 \%$) under steady-state conditions. Overall, Backpack EKG provides a measurement-oriented, operationally autonomous acquisition platform for screening campaigns where deployment logistics and power availability constrain conventional EKG instrumentation.
The world is facing a critical maternal health crisis, with more than 700 pregnant women dying daily in 2023 due to obstetric causes, underscoring the need for timely access to reliable clinical information. This study evaluates an orchestrated multi-agent system designed to answer clinical questions derived from electronic health records, special obstetric cases, and clinical practice guidelines. The system uses large language models (LLMs) from the Gemini family and relies on automated question classification and semantic routing across specialized obstetric categories to generate context-aware responses. Clinical data are structured using the OMOP Common Data Model to ensure interoperability. The platform is deployed on Google Cloud Platform and integrates RetrievalAugmented Generation (RAG) with clinical knowledge graphs. Performance was assessed using zero-shot and few-shot prompting using 35 expert-curated maternal-fetal clinical questions. Gemini 2.5 Flash and Gemini 2.5 Pro achieved 98.9% precision and 100% recall under both strategies. Gemini 3.0 ProPreview showed reduced precision to 53.1% in few-shot prompting while maintaining full recall. These results demonstrate the feasibility of orchestrated multi-agent systems for contextualized obstetric clinical support.
Binaural hearing is essential for sound localization and auditory awareness, and its decline is frequently associated with a range of conditions such as neurocognitive disorders, traumatic brain injuries, and ageing. Electroencephalography (EEG) offers high temporal resolution for studying neural processing during binaural hearing. However, clinical-grade EEG systems are cumbersome, while current consumer-grade devices lack the parietal and occipital coverage necessary to capture the later stages of binaural processing. We hypothesize that augmenting a consumer-grade EEG device with one or more supplementary electrodes may offer the spatial coverage of a clinical-grade system with the convenience of a headband. We here investigate the optimal placement for a single auxiliary electrode added to the MUSE S headband for binaural hearing assessments. We record EEG data from participants performing a horizontal sound localization task using a 32-channel clinical system to identify regions of peak activation. Through analysis of power amplitude fluctuations in delta and alpha bands, as well as Independent Component Analysis (ICA), temporal and occipital regions are identified as the most informative sites. Subsequent validation experiments, utilizing the MUSE S with a single auxiliary electrode placed at these identified coordinates, demonstrate that these locations captured substantially higher fluctuations in relevant frequency bands compared to a control site. These findings suggest that a modification to consumer EEG hardware can successfully capture additional signatures of spatial hearing, paving the way for accessible, portable auditory and neurological screening outside of clinical settings.
This work presents the development of a bioimpedance spectroscopy meter intended for in vivo rectal tissue measurements. The device is based on the Analog Discovery 2, a compact and robust platform that integrates a waveform generator, a two-channel oscilloscope, and dual power supplies, thereby simplifying the implementation of several essential stages required in a bioimpedance measurement system while enabling a flexible instrumentation architecture. The analog front-end, including a Howland current source and voltage and current measurement stages, was designed and implemented by the authors. The main design parameters, such as frequency range, current amplitude, and number of frequency points, were defined based on previous experimental studies conducted by our research group on colorectal tissue bioimpedance characterization. The developed device exhibited errors below one percent during resistor validation tests and minimal deviations in saline solution measurements after calibration, demonstrating adequate measurement stability and accuracy. Furthermore, the system demonstrated the capability to measure ex vivo porcine colorectal tissue and in vivo human rectal tissue, successfully identifying the characteristic dispersions typically observed in biological tissues. These results indicate that the proposed system provides stable impedance measurements and represents a compact instrumentation platform for bioimpedance spectroscopy studies of colorectal tissue under both experimental and preliminary clinical conditions.
The original DICOM [3] and its DICOMweb [4] evolution are a fundamental stumbling block in the engineering required to optimize the transmission of huge studies in complex air networks (mixed $\mathbf{5 G}$ /satellite/WIFI). Consequently, we created a new representation of DICOM objects that we called DECK (Dicom Exam Contextual Keys) and specialized a new WebTransport [5] protocol. These have the complete study, and any contextualized attribute of it as primary objects. This differs radically from the SOP instance of DICOM, and from the additional XML, JSON, HTTP, meta and bulk data of DICOMweb. We redefined both the original encoding (which becomes a flat list of key-value) and the DICOM Upper Layer (instead of which we use UDP duplex stream and asynchronous push of datagrams). REGEX [6] is used by us for attribute selection in the request to be executed on the server, so that we do not need XPATH [7] applied on a memory DOM representation of a SOP instance. Our response format is a very simple format, much more so than JSON, XML or DICM. These features were designed to facilitate javascript interaction with an HTTP/3 server from a web page running on the client. Such workspace gets encryption and complex network routing for free thanks to the QUIC engine of HTTP/3 WebTransport. This new DICOM communication stack provides the necessary building blocks to design a reactive mobile imaging workstation for medical web imaging.
Continuous monitoring of heart rate and respiration rate, crucial for intensive care is currently performed using skin contact sensors. While these offer reliable measurements, they are not convenient in the case of burn trauma patients or neonates, as contact sensors pose the risk of infection and can lead to additional skin damage. Non-imaging, noncontact monitoring using near-field sensing methods offers an attractive alternative without compromising the patient’s privacy. We have investigated the feasibility of strain gaugebased Ballistocardiography (BCG) for continuous monitoring of a subject’s heart rate and respiration rate in a hospital settings. We developed a BCG sensing system that could be easily integrated into a standard hospital bed without major mechanical modifications, and measure vitals without causing any discomfort to the subject. The designed system utilized strain gauges as sensors to capture BCG signals with 16-bit resolution. With an in vivo study on 17 subjects, we verified the ability of the system to capture reliable BCG signals and key vitals. The system measured heart rate with a root mean square error (RMSE) of 3.5 beats per minute (bpm) and a mean absolute error (MAE) of 2.8 bpm. Similarly, the respiratory rate was measured with an RMSE of 2.3 breaths per minute (brpm) and an MAE of 1.9 brpm.
The motion capture techniques based on markers have been used worldwide for gait analysis for years. However, despite the accuracy, this is a costly method that needs specialized equipment and personnel. That’s the reason why, nowadays, new low-cost methods such as the markerless techniques based on human pose estimation have been developed. The calibration stage in motion capture systems is an important procedure that clearly impacts on the system performance. However, it is not fully validated how the results are affected by variations in the calibration procedures. This work proposes to evaluate 2 different spatial configurations of the acquisition devices modifying their height (H1, highest and H2, lowest) and different calibration procedures (intrinsics and extrinsics parameters) in a markerless system based on 4 smartphones IPhone SE. To evaluate the impact of intrinsics parameters, the distance between calibration pattern and camera (DPC); and the number of registered frames (\#frames) are modified. Three different calibration patterns (extrinsics calibration) were evaluated to quantify the effect on markerless system performance. The final results of the trajectory of the markers are compared to the gold standard method, 8 VICON Bonita cameras, data which is captured simultaneously. Comparisons are realized considering two metrics widely used in bibliography (MPJPE and PCK). The results show that H1 configuration performs slightly better than H2. Regarding extrinsic calibration, patterns that consider a volume more than a surface are better for both configurations. About the intrinsic calibration, the best performance is given by the minimum DPC with pattern 2, with a negligible influence of \#frames. Respecting PCK performance for different thresholds, the results are worse (higher $\tau^{*}$ values) for $H 2$ than for $H 1$ in most of the markers. Comparing both methods, the most affected direction is the transversal while the vertical and anteroposterior were similar to each other.
This paper presents the LAMAS Mammography Image Database, a large-scale, BI-RADS-oriented temporal dataset designed to support reproducible and auditable artificial intelligence research in breast imaging. The database comprises 836,328 full-field digital mammography images from 139,388 examinations of 50,538 patients acquired between 2007 and 2024 at two specialized breast imaging centers. Ground truth is defined using an outcome-driven framework integrating longitudinal follow-up and histopathological confirmation, enabling explicit annotation of true positives, true negatives, false positives, and false negatives. Structured annotations include BIRADS assessments, lesion-level bounding boxes, and temporal linkage across examinations, supporting measurement of stability, progression, and retrospective error identification. A baseline deep learning experiment for BI-RADS breast density classification demonstrates dataset usability and label consistency. By combining standardized reporting, longitudinal verification, and controlled governance, the proposed database establishes a robust measurement framework for clinically meaningful development and validation of mammography AI systems.
Cardiovascular diseases (CVD) are the leading cause of global mortality, accounting for approximately 19 million deaths annually, with myocardial infarction (MI) being a critical manifestation. The 12-lead electrocardiogram (ECG) remains the gold standard for initial diagnosis; however, its manual interpretation is complex and prone to specialist scarcity. Deep learning models, particularly Bidirectional Long Short-Term Memory (Bi-LSTM) networks, have shown high accuracy in automating MI detection. Despite this, the inherent “black-box” nature of these architectures hinders clinical adoption due to a lack of transparency and auditability. This study proposes a comprehensive framework to audit an MI detection model by applying two distinct Explainable AI (XAI) techniques: Gradientweighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME). Using the PTB-XL dataset (21,799 ECGs), this study selected $11,601 \mathrm{MI}$ and Normal records sampled at 100 Hz, applying baseline-wander removal via median filters and high-frequency noise attenuation through 35 Hz Butterworth low-pass filtering. The three-layer Bi-LSTM ($\mathbf{3 2 / 1 6 / 4}$ units) reached 89.73% test accuracy. Auditability was evaluated using temporal and spatial (lead-wise) relevance across True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN) exemplars. Grad-CAM highlighted clinically meaningful regions (QRS and ST) in TP cases (Sample #514, $P_{\text {MI }}=0.9950$), where detections relied on recurrent patterns across multiple beats. Conversely, FP errors (Sample #2302, $P_{\mathrm{MI}}=0.6461$) were largely triggered by early signal artifacts and baseline drift. LIME provided a granular decomposition of these decisions, identifying specific lead contributions. In the FN case (Sample #1343, $P_{\text {MI }}=0.2410$), LIME revealed that although the model detected MI-related features in leads V1 and V2, these were numerically outweighed by competing local weights favoring the “Normal” class in leads aVL and aVR. This dual-XAI approach effectively transforms automated ECG interpretations into an auditable process, providing a necessary layer of verification for clinical decision-support systems.
Parkinson’s disease (PD) affects motor regulation, altering both limb movement and fine phonatory control. Although speech-based characterization of PD traditionally relies on acoustic features, these mainly reflect source-filter interactions and provide limited insight into the underlying laryngeal control processes. This study investigates PD-related changes in voice production by combining conventional acoustic descriptors with biomechanically informed control variables generated within a physiologically motivated laryngeal motor control framework. Sustained vowel /e/ phonation from the PC-GITA corpus was analyzed in 50 individuals with PD and 50 healthy controls to provide a controlled investigation of phonatory motor regulation under simplified phonation conditions. Acoustic descriptors capturing phonatory stability (e.g., perturbation and harmonicity measures), pitch, and intensity variability were extracted, together with model-derived trajectories of subglottal pressure and intrinsic laryngeal muscle activations. Group differences were assessed using non-parametric statistics and effect size analysis, and discriminative performance was evaluated using support vector machine classification. Statistical analyses indicate that acoustic instability observed in Parkinsonian phonation is coherently reflected in model-derived biomechanical control features. Classification analysis using biomechanical features yielded performance comparable to acoustic descriptors. Overall, this work provides research-level insights into Parkinsonian phonatory control mechanisms and highlights the potential of biomechanically interpretable descriptors as indicators of phonatory dysfunction.
Affordable point-of-care ultrasound technologies with automated capabilities for vessel analysis are integral to wide-spread adoption of assessing early vascular ageing, where A-mode technologies present simpler and cost-effective alternatives. Automating artery detection, a task that is intrinsic to such technologies, still suffers robustness issues due to low magnitude arterial pulsations, extreme heart rates, and motion artifacts in general, posing even higher challenge for A-mode 1D signals. In this work we present a hybrid deep-learning and classical signal processing approach for robust artery detection from A-mode ultrasound frames, that is feasible to deploy on an entry level edge-computing system. The method uniquely employs spatio-temporal features of the ultrasound frames recording arterial dynamics, by stacking temporally acquired 1D A-mode frames to form motion mode images on which detection is performed. It’s a hybrid method that uses deeplearning framework for the region of arterial walls and then classical signal processing for accurately locating the walls’ positions within the region, with a high resolution of 10 micrometer. We employ YOLO-based deep-learning models for the artery detection, benchmarking multiple variants for accuracy, model size, and inference speed. The models were deployed on an edge-computing system to ensure real-time feasibility. Among them, YOLO-11n achieved the best balance between performance and efficiency, with 2.6 million parameters, 83.8% precision, 93% recall, and an average inference time $\sim \mathbf{1 0 ~ m s}$ per image. This balanced attribute of 11n, makes it a potential choice for the deployment on edge systems. Further, 11n exhibited comparable high performance even on the edge cases where classical signal processing methods typically would fail, with a precision of 77.5% and a recall of 81.4%. These results establish the feasibility of an AI-powered, edge-computing A-mode ultrasound system for real-time vascular assessments.
Organ transportation remains one of the most crucial and susceptible stages of the transplantation process, as the physiological stability of donor tissues directly affects graft viability and post-transplant outcomes. During transport, organs are vulnerable to thermal fluctuations, mechanical stress, and metabolic deterioration, compromising their functional integrity. Conventional Static Cold Storage (SCS), although widely used due to its simplicity and low cost, suffers from uneven cooling, temperature fluctuations, a lack of biochemical monitoring, and mechanical vulnerability, which contribute to ischemia–reperfusion injury and increased organ discard rates. This study introduces Orgport, a next-generation intelligent organ transportation system that integrates active heat management, real-time monitoring, and data traceability. The system maintains stable hypothermic conditions ($4-8{ }^{\circ} \mathrm{C}$) using compressor-based cooling and multi-sensor monitoring to ensure uniform temperature distribution. Automated perfusion control, vibration isolation, and web-based visualisation are incorporated to enhance organ stability, minimise mechanical shock, and enable continuous remote monitoring during transport. Comparative evaluation of caprine hearts demonstrated that Orgport outperformed conventional SCS in terms of temperature uniformity, biochemical stability, and mechanical protection. Functional and histological analyses confirmed improved structural preservation, reduced edema, and enhanced post-reperfusion recovery. The proposed system represents a scalable and cost-effective approach toward intelligent, data-driven organ preservation for safer longdistance transplantation.
Quantitative gait analysis commonly relies on marker-based optoelectronic motion capture (mocap) systems, which provide high-resolution kinematic data and can be synchronized with force plates, electromyographic (EMG) sensors, and other biomechanical devices. Although the C3D file format is widely used to store these multimodal data, extracting, segmenting, and processing C3D content often requires proprietary software or manual procedures that limit reproducibility and hinder large-scale analyses. This work presents an open, script-based framework for processing locomotion trials, combining Vicon ProCalc models with two Python pipelines designed to automate stride segmentation and compute spatio-temporal and mechanical parameters of human gait. The first pipeline (“python_operations”) performs stride detection using kinematic criteria based on foot marker vertical velocity. It interactively identifies initial foot contacts, segments the trial into strides, and exports markers, model outputs, and EMG signals into portable CSV and NumPy (npy) formats. Additional utilities prepare EMG data for muscle synergy analysis using the open-source R package musclesyneRgies. The second pipeline (“energetics”) loads the stride-based npy files and computes key gait variables, including stride time, stride length, duty factor, center-of-mass energies, external and internal mechanical work, recovery indices, and congruence. The pipeline also generates diagnostic plots of kinetic, potential, and total energies for each stride. All scripts, Vicon models, and example datasets are provided in the public GitLab repository fromC3d_to_energetics, enabling full reproducibility and facilitating adaptation to different research needs. The proposed framework offers a transparent, customizable, and extensible solution for researchers requiring automated processing of C3D files and detailed mechanical analysis of human locomotion.
Markerless pose estimation is increasingly used for human motion analysis due to its low cost and ease of deployment. However, the measurement uncertainty and sensitivity under operating conditions remain incompletely explored. This study aims to evaluate the metrological performance of MediaPipe for knee flexion angle estimation during squat execution, with particular attention to the effects of acquisition geometry and clothing configuration. MediaPipebased estimates were compared against the BioVal (Biorescue, RM Ingenierie, France), an inertial measurement unit (IMU) used as reference. A preliminary static analysis was performed to identify the optimal camera-subject distance, which was then used for the dynamic protocol. Dynamic acquisitions were performed at viewing angles ranging from 45° to 90° and under three clothing conditions (shorts, long trousers, and tight-fitting trousers). Results show a systematic underestimation of the knee flexion angle by MediaPipe, with a residual of $(-13.9 \pm 15.1)^{\circ}$ (reported as mean ± standard deviation) and 95% limits of agreement of [$-43.5^{\circ}, 15.5^{\circ}$]. From our tests it results that the estimation performance is influenced by the acquisition geometry, with lateral views yielding the mean deviation and variance $\left((-2.3 \pm 6.4)^{\circ}\right.$ at $\left.90^{\circ}\right)$. Clothing configuration also contributes to measurement uncertainty, with tight-fitting trousers yielding more stable estimates (i.e., $(1.6 \pm 3.5)^{\circ}$ at $\mathbf{9 0}^{\circ}$) compared to shorts.
The assessment of pulse wave velocity (PWV) in the brachial artery could provide valuable insight into the mechanical behaviour underlying oscillometry, ultimately improving the accuracy of non-invasive blood pressure measurement. The wall displacement-based (WD) method is a widely adopted solution for PWV estimation in large arteries, but its application to smaller vessels such as the brachial artery remains challenging due to stringent temporal sampling requirements. In previous work, we showed that conventional ultra-fast ultrasound coherent compounding processing fails to provide reliable WD-based PWV estimates in the brachial artery, under typical acquisition settings. In this study, we investigate alternative processing strategies aimed at increasing the effective temporal sampling rate without increasing the transmission pulse repetition frequency. Three approaches are compared: conventional compound-first processing, per-angle motion-first estimation, and a time-aligned per-angle motionfirst strategy. Numerical simulations of the brachial artery are used to evaluate PWV estimation performance under different noise conditions and reference PWV values. The results show that time-aligned processing partially mitigates temporal undersampling effects for physiologically relevant PWV values, at the cost of increased sensitivity to noise. While all approaches remain affected by noise and temporal resolution limitations, the time-aligned strategy yields PWV distributions that extend toward higher values compared to conventional processing. These findings highlight the role of temporal sampling in WD-based PWV estimation and motivate further investigation toward noise-robust implementations.
The automatic classification of cardiac conditions—specifically Arrhythmia (ARR), Congestive Heart Failure (CHF), and Normal Sinus Rhythm (NSR)—from electrocardiogram (ECG) signals remains a critical challenge in clinical diagnostics. This paper presents the design of a Recurrent Neural Network (RNN) classifier using ECG recordings from PhysioNet-derived datasets. The proposed methodology relies on a multi-domain hand-crafted feature extraction strategy to capture the complex dynamics of cardiac signals. To ensure temporal consistency, ECG signals were segmented into 10 -second windows, from which a comprehensive feature set was computed across four key domains: time-domain statistics, Hjorth parameters, frequency-domain spectral descriptors, and nonlinear measures derived from Poincaré plots. A statistical feature selection pipeline—combining the Kruskal–Wallis test, Benjamini–Hochberg False Discovery Rate (FDR) correction, and effect size analysis—was applied to retain the most discriminative and non-redundant predictors. To identify the most effective temporal modeling strategy, the classifier incorporates a Bayesian hyperparameter optimization framework that automatically selects the optimal recurrent architecture among Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Units (GRU) topologies. This process simultaneously tunes model complexity and regularization parameters to mitigate overfitting. In addition, class imbalance is addressed through a cost-sensitive learning approach using weighted loss functions. The resulting classifier achieves strong discriminatory performance across the three rhythm classes, demonstrating that a Bayesian-optimized RNN architecture combined with domain-specific features provides a solution for cardiac rhythm analysis.
