BIOSIGNALS 2026 Abstracts

Full Papers Short Papers
Full Papers
Paper Nr: 51
Title:

Associations between Circadian Sleep Alignment, Sleep Regularity, and Wellbeing: Evidence from Wearable Device Data
Authors:

Swathi Senthil, Szymon Fedor, Boyu Zhang, Viswam Nathan, Robert Stickgold and Rosalind W. Picard
Abstract: Sleep timing and consistency are increasingly recognized as key predictors of both mental and physical wellbeing. In this study, we investigate how sleep regularity (SR) and circadian misalignment-the discrepancy between sleep timing and endogenous circadian rhythm-relate to daily and overall wellbeing and subjective sleep quality. During a sixteen-day period, 48 participants wore Samsung Galaxy smartwatches to passively track sleep and completed daily surveys that evaluated wellbeing and recorded their subjective bedtime and wake-up time. At the end of the study, saliva samples were collected under dim-light conditions to assess participants circadian phase. We computed overall SR across the study period, daily SR, and circadian misalignment, and analyzed their associations with next-day outcomes using t-tests and linear mixed effects models on both watch-recorded and self-reported data. Our findings indicate that sleep regularity and alignment with circadian rhythm are significantly associated with subjective sleep quality and next-day wellbeing. These results highlight the possible importance of maintaining consistent sleep patterns and circadian alignment for both daily functioning and overall mental and physical health.
Download
Paper Nr: 71
Title:

Phonation Encephalographic Bands in Parkinson’s Disease
Authors:

Pedro Gómez-Vilda, Andrés Gómez-Rodellar, Jiři Mekyska and Daniel Palacios-Alonso
Abstract: This study introduces a novel method using phonation-related encephalographic (phEG) features—specifically θ and β frequency bands—as biomarkers for distinguishing Parkinson’s Disease (PD) from control and normative groups. Neuromotor activity is indirectly estimated from laryngeal biomechanics during sustained vowel phonation, with group differences quantified via Jensen–Shannon divergence. Classification and Regression Trees (CART) enable interpretable stratification across male and female cohorts. Male participants showed more complex tree structures and higher sensitivity in normative detection, while both groups achieved strong performance using only two phEG-derived features: up to 88% accuracy, 94% specificity, and F1 scores of 0.80. These results align with known roles of β oscillations in motor maintenance and θ oscillations in adaptive control and sensorimotor integration. The approach demonstrates that minimal EEG-derived input can yield biologically plausible and diagnostically relevant classifications. Future work will expand the methodology to larger cohorts and additional phEG features to enhance precision in characterizing neuromotor impairments.
Download
Paper Nr: 79
Title:

Benchmarking and Enhancing PPG-Based Cuffless Blood Pressure Estimation Methods
Authors:

Neville Mathew, Yidan Shen, Renjie Hu, Maham Rahimi and George Zouridakis
Abstract: Cuffless blood pressure screening based on easily acquired photoplethysmography (PPG) signals offers a practical pathway toward scalable cardiovascular health assessment. Despite rapid progress, existing PPGbased blood pressure estimation models have not consistently achieved the established clinical numerical limits such as AAMI/ISO 81060-2, and prior evaluations often lack the rigorous experimental controls necessary for valid clinical assessment. Moreover, the publicly available datasets commonly used are heterogeneous and lack physiologically controlled conditions for fair benchmarking. To enable fair benchmarking under physiologically controlled conditions, we created a standardized benchmarking subset NBPDB comprising 101,453 high-quality PPG segments from 1,103 healthy adults, derived from MIMIC-III and VitalDB. Using this dataset, we systematically benchmarked several state-of-the-art PPG-based models. The results showed that none of the evaluated models met the AAMI/ISO 81060-2 accuracy requirements (mean error < 5 mmHg and standard deviation < 8 mmHg). To improve model accuracy, we modified these models and added patient demographic data such as age, sex, and body mass index as additional inputs. Our modifications consistently improved performance across all models. In particular, the MInception model reduced error by 23% after adding the demographic data and yielded mean absolute errors of 4.75 mmHg (SBP) and 2.90 mmHg (DBP), achieves accuracy comparable to the numerical limits defined by AAMI/ISO accuracy standards. Our results show that existing PPG-based BP estimation models lack clinical practicality under standardized conditions, while incorporating demographic information markedly improves their accuracy and physiological validity.
Download
Paper Nr: 125
Title:

Real-Time Predetermination of Excitation Frequencies for Optimizing Electrochemical Impedance Spectroscopy
Authors:

Norman Pfeiffer, Toni Wachter, Joan Bausells, Abdelhamid Errachid and Albert Heuberger
Abstract: Electrochemical impedance spectroscopy (EIS) commonly relies on fixed logarithmic sweeps that ignore which frequencies are most informative. This unnecessarily prolongs the measurement duration, a limitation that is particularly problematic for mobile applications or in instances of drifting electrochemical cells. We present a real-time predetermination of excitation frequencies (RTPEF) based on a multi-objective optimization (MOO) with the non-dominated sorting genetic algorithm II (NSGA-II) that jointly minimizes fitting error, measurement time, number of frequencies, and fitting time. Pareto-optimal frequency sets from 200 simulated Randles circuits are aggregated into a sector-specific, frequency-dependent probability distribution derived from phase-based curve segmentation. During measurement, sparse probing maps sectors before drawing additional frequencies from the distribution. These frequencies are then iteratively refined via complex non-linear least squares (CNLS) fitting until an average square deviation threshold is met. On 400 simulated spectra, RTPEF used 9 frequencies on average, reduced total measurement time by 17.5 % versus a 30-point logarithmic sweep (217.7 s vs. 263.9 s), and markedly increased the fitting success rate (98.3 % vs. 81.8 %), while maintaining comparable parameter estimation errors. Subsequent research might focus on the validation of the presented approach applied to real measurements.
Download
Paper Nr: 168
Title:

Evaluating the Vector Based Selection Algorithm for Sample Entropy Estimation on FHR Time Series with Missing Data
Authors:

Athanasios Georvassilis, Dimitrios Platakis and George Manis
Abstract: In physiological signal analysis, entropy measures provide valuable insight into the complexity of biological systems. Sample Entropy is a widely applied nonlinear index, yet its reliability is strongly affected by missing or corrupted data, which frequently occur in practice. This paper evaluates the impact of different missing-data strategies on the estimation of Sample Entropy, using heartbeat time series from fetal recordings. Four approaches are examined: preprocessing with deletion, linear interpolation and quadratic interpolation, and a newly proposed algorithm, the Vector Based Selection algorithm. Their performance is systematically assessed across multiple parameter settings and varying percentages of data loss. Experimental results demonstrate that the Vector Based Selection method consistently produces more stable and close to the expected entropy values, particularly under higher levels of missing data, whereas traditional strategies often result into significant distortions. These findings underline the importance of robust processing in entropy-based analysis and suggest that the Vector Based Selection approach can improve the reliability of complexity assessment in physiological signals.
Download
Paper Nr: 173
Title:

A Cross-Modality Comparison between Structural and Functional Brain Connectivity and Individual Differences in Intelligence
Authors:

Sue Si-Qi Gao, Tong Liang, Zhe Sun, Cesar F. Caiafa and Jordi Solé-Casals
Abstract: This study aimed to examine whether individual differences in cognitive intelligence can be explained by corresponding patterns of structural and functional brain connectivity within a common whole-brain parcellation. Data were obtained from the WU–Minn Human Connectome Project (HCP) dataset, restricted to 418 young adults aged 26–30. Participants were divided into a high and a low cognitive intelligence task performance group (HIG/LIG; each n = 30) according to the top and bottom deciles of the general intelligence (g) factor derived from ten cognitive tasks. Functional connectivity (FC) was computed from resting-state fMRI using the CONN toolbox (n = 58), and structural connectivity (SC) was estimated by probabilistic tractography on diffusion MRI (dMRI) data (n = 60). Group-level ROI-to-ROI analyses and graph-theoretical metrics were compared between groups. Significant FC differences were observed primarily in cortico–subcortical clusters, where HIG showed stronger frontal–subcortical connections and weaker parietal–cerebellar coupling than LIG. In contrast, SC analyses revealed no significant frontoparietal edges under the same parcellation. The results suggest that functional connectivity differences associated with intelligence cannot be fully explained by structural architecture. These findings highlight the necessity of integrating multimodal connectivity measures to better understand how structural constraints support functional coordination underlying cognitive ability.
Download
Paper Nr: 213
Title:

Improving Feature Extraction for Spike Sorting Using a Custom Loss Function for an Autoencoder
Authors:

Elizaveta Kompaniets, Steven Le Cam and Radu Ranta
Abstract: The aim of this paper is to introduce a novel approach for extracting discriminative features for spike sorting. The proposed model extends the autoencoder-based architecture presented in (Ardelean et al., 2023) by incorporating a dedicated loss function computed directly in the feature (code) space. This design enables the generation of feature representations specifically adapted to the classification algorithm employed at the model output. In particular, we propose a loss function that jointly minimizes intra-class variance and maximizes inter-class variance, combined with a standard K-means algorithm for clustering. Model training requires a dataset with known or estimated labels. We first evaluate the feature extractor using perfectly labelled data to validate the effectiveness of the proposed architecture and loss formulation. Subsequently, we explore a more realistic scenario in which the shallow autoencoder from (Ardelean et al., 2023) provides estimated labels. Our results demonstrate that the proposed approach significantly improves the final classification performance.
Download
Paper Nr: 217
Title:

CAAM-Net: A Chrono–Acoustic Attention MobileNet for Snore-Based Sleep Apnea Screening
Authors:

Himanshu Sharma and Pradip K. Das
Abstract: Obstructive Sleep Apnea (OSA) is a widely prevalent yet significantly underdiagnosed condition for which in-laboratory polysomnography (PSG) continues to be the diagnostic gold standard. The high costs, restricted access, and inconvenience for patients drive the need for scalable, non-contact screening alternatives. We present CAAM-Net, a lightweight snore-event classifier designed for at-home screening. The model employs a MobileNetV3-Large backbone, augmented with an explicit squeeze-and-excitation (SE) attention head, and is trained on stacked log-Mel and Mel-Frequency Cepstral Coefficient (MFCC) spectrograms derived from PSG-synchronized audio. In a standard segment-level random split, CAAM-Net achieves an F1 score of 0.876 for abnormal snores (related to apnea) and a macro-F1 score of 0.85, surpassing both plain MobileNetV3 and a more complex baseline VGG-19. Importantly, we also report performance under a clinically relevant, subject-disjoint Leave-One-Patient-Out (LOPO) protocol, achieving a macro-F1 of 0.78. Unlike recent studies that focus on whole-night Apnea-Hypopnea Index (AHI) regression or elaborate multimodal systems, we emphasize a compact, single-modal architecture, transparent reporting under LOPO conditions, and training efficiency relevant for deployment. Our findings underscore a notable trade-off between accuracy and efficiency, indicating a viable pathway for accessible, non-contact OSA prescreening.
Download
Paper Nr: 237
Title:

Differential Contributions of Physiological Features across Early, Mid and Late Sleep Phases to Next Morning Cognitive Performance
Authors:

Chijing Wang, Ryota Nitto, Yuki Ban, Miki Nakai, Jun'ichi Shimizu, Tomoyoshi Ashikaga and Shin'ichi Warisawa
Abstract: Sleep physiology is often summarized using whole-night metrics such as total sleep time and sleep efficiency, yet such macrostructural indices can miss the intra-night temporal dynamics through which recovery unfolds. We investigated whether phase-specific physiological features from early, mid, and late sleep are differentially associated with next-morning cognitive performance. Thirty-two healthy male adults underwent overnight recordings of EEG, ECG, electrodermal activity, and respiration, and completed morning assessments of psychomotor vigilance that included reaction time and a flying-delay index. Participants also reported subjective sleepiness using the Karolinska Sleepiness Scale and performed a 3-back working memory task at 6:00, 7:00, and 9:00, with performance evaluated relative to a 20:00 baseline. Sleep was segmented into three equal-duration phases defined on an individual-night basis, and physiological features were computed separately for each phase. For statistical analysis, we fitted feature-wise linear mixed-effects models with participant-specific random intercepts and controlled the false discovery rate across tests. Significant associations were concentrated in psychomotor vigilance outcomes. Mid- and late-sleep features, including parasympathetic-related heart rate variability indices and synchronized EEG activity, were associated with slower reaction times and greater response-control instability at 6:00 to 7:00, consistent with stronger sleep inertia. Subjective sleepiness showed only a single significant association with a mid-sleep autonomic feature, and no phase-specific predictors were identified for 3-back performance after correction. These findings indicate that time-within-sleep physiology provides sensitive markers of immediate post-awakening vigilance beyond whole-night averages, and highlight the functional relevance of phase-specific autonomic and cortical coordination for morning readiness.
Download
Paper Nr: 239
Title:

Supervised Human Activity Recognition in Office Environments Using Trapezius-Mounted Accelerometers
Authors:

Gonçalo Barros, Sara Santos, Phillip Probst and Hugo Gamboa
Abstract: We evaluate whether trapezius-mounted muscleBAN (mBAN) accelerometers alone suffice for Human Activity Recognition (HAR) in office contexts. Using a labeled dataset comprising ten sub-activities grouped into three main classes (sitting, standing, walking), we implement a rigorous pipeline: protocol-driven segmentation, gravity removal and filtering, and a bespoke quality-assessment stage to handle non-continuous mBAN streams. Feature engineering follows a TSFEL-based configuration (selected statistical and spectral descriptors), and class balance is enforced by trimming to the minimum instances per subject × sub-activity. We benchmark Random Forest (RF), SVM (RBF) and k-NN. Cross-validation identifies RF without feature normalization as best (mean accuracy 87.53%±1.55). Deployed as the production model on a 20% hold-out set, RF attains 86.58% accuracy; class recalls are 71.8% (sitting), 91.7% (standing), and 96.3% (walking), with walking precision =98.7%. Per-subject accuracies range 82.97%–89.65%. Errors concentrate in sittingstanding confusions, while walking is detected with high reliability. Results support ACC-only shoulder sensing as a practical basis for office HAR, with clear pathways for refinement.
Download
Paper Nr: 242
Title:

QT Interval Estimation by Denoising Electrocardiogram Signals
Authors:

Francisco Javier Maldonado Carrascosa, José Enrique Muñoz Expósito, Sebastián García Galán, Francisco Jesús Cañadas Quesada, Adam Marchewka and José Ranilla Pastor
Abstract: Accurate measurement of the QT interval in electrocardiogram (ECG) signals is critical for diagnosing cardiac conditions but remains challenging due to various noise artifacts that affect signal recordings. This paper presents a multistage approach for QT interval estimation using an advanced denoising technique. the proposed methodology involves applying bandpass filtering to isolate frequency components, detecting QRS complexes through an iterative energy-based algorithm, identifying T waves using signal interpolation to remove QRS complex interference and implementing the Pan-Tompkins algorithm for precise feature detection. The proposed method was evaluated using ECG samples from the PhysioNet database, achieving high accuracy. The algorithm shows particular effectiveness in handling noisy signals, addressing a significant challenge in automated ECG analysis. Quantitative evaluation using ECG samples from the PhysioNet database demonstrated the effectiveness of the proposed method, achieving a mean error of 0.0087, a median error of 0.0085s, and a low standard deviation of 0.0338, indicating high accuracy and consistency in QT interval estimation across diverse signal conditions. Classification-style metrics, considering estimations within ±5 ms of manual reference values, showed 94.2% accuracy, 93.5% sensitivity, 95.1% specificity, and an F1-score of 0.94. Statistical analysis confirmed reliability, with a paired t-test p-value of 0.27 and Pearson correlation of 0.982 (p<0.001). The error distribution was tightly centered around zero, with less than 1% outliers, demonstrating minimal bias and robustness under noisy conditions. These results indicate that the proposed approach provides accurate, consistent, and reproducible QT interval estimation, offering potential applications in clinical diagnostics, continuous cardiac monitoring, and automated ECG analysis systems.
Download
Paper Nr: 254
Title:

Probabilistic Detection of Interictal Epileptiform Discharges
Authors:

Inês A. Gonçalves, Inês Silveira, Sem Hoogteijling, Irene Heijink, Maeike Zijlmans, Luís Silva and Hugo Gamboa
Abstract: Approximately 30% of patients with epilepsy are considered for surgery after failing to achieve seizure con-trol with anti-epileptic drugs. During epilepsy surgery, intraoperative Electrocorticography (ioECoG) is used to delineate the epileptogenic zone, yet signal interpretation remains based on visual inspection, a subjective and time-consuming process prone to observer bias. Existing computational methods for detecting interic-tal epileptiform discharges (IEDs) often provide deterministic classifications without quantifying prediction uncertainty, limiting their clinical applicability. This work presents an IED detection model that classifies intracranial EEG segments as healthy or pathological while also outputting the probability of pathology, serving as an uncertainty measure to support intraoperative decisions. Statistical, temporal, and spectral features were extracted alongside descriptors derived directly from segment scalograms, capturing the joint time–frequency characteristics of epileptiform activity. These features were used to train several machine learning models, with the Random Forest–Logistic Regression combination achieving an area under the curve (AUC) of 0.975 on unseen data and a Brier score of 0.0499. Model transparency was reinforced by identifying key features influencing predictions, providing interpretable insight into epileptiform electrophysiological patterns. Overall, the proposed framework provides a transparent and uncertainty-aware approach for automated IED detection, representing a promising step toward objective and assistive tools in epilepsy surgery.
Download
Paper Nr: 281
Title:

Controlled Large Scale Synthetic Motion Dataset Generation Leveraging Text-to-Motion and Sample-Wise Quality Assurance
Authors:

Lourenço Abrunhosa Rodrigues, Markus Wenzel and Felix Putze
Abstract: Learning representations of complex health data, such as human motion over the evolution of impairments, requires highly structured datasets with given temporal dependencies and reliable ground truth information. Collecting such data from humans is a costly endeavor, and subject to noise and confounders that make it difficult to attain these properties. Synthesizing data using text-to-motion models offers a solution to these problems, but requires an efficient and effective quality assurance process at the sample level that, until this work, did not exist. In this work we evaluate how MoBERT sample quality metrics align with human evaluations of sample naturalness and faithfulness. We design an efficient and highly scalable quality assurance protocol that can be used to verify the validity of samples in large scale synthetic human motion datasets. We generate a synthetic dataset containing thousands of samples, swiftly identify the samples with unsuitable quality, and show, with an activity recognition model, that these poor quality samples are indeed different from their higher quality counterparts, observing a 10% point drop in performance when recognizing the activity represented by poor quality samples.
Download
Paper Nr: 302
Title:

Detecting Infant Cries in Noisy NICU Recordings: A Study on SSL Models
Authors:

Ákos Antal, Péter Földesy and Péter Mihajlik
Abstract: Preterm infants in Neonatal Intensive Care Units (NICUs) communicate pain, discomfort, and medical needs primarily through crying. Automated detection of infant cries could support caregivers by reducing response times and enabling continuous monitoring. However, recordings in real NICU environments present significant challenges due to background noise, sound attenuation of the closed incubators, and limited labeled data. In this study, we investigate the use of self-supervised learning (SSL) models for infant cry detection in noisy NICU recordings. Audio data were collected in collaboration with Semmelweis University, Hungary, under approved ethical protocols, with additional samples from the CryCeleb2023 dataset to address class imbalance, totaling 83,860 samples. Data were segmented, annotated, and processed using noise reduction techniques. We evaluated multiple SSL-based architectures, and conducted ablation studies on input window size, layer selection, and fine-tuning strategies. Results show that lightweight HuBERT model variants achieve over 90% accuracy and F1 scores up to 0.90, outperforming baseline models, demonstrating robustness to noise and deployment feasibility. These findings highlight the potential of SSL approaches for real-world infant cry detection in NICUs.
Download
Paper Nr: 336
Title:

Stress Detection from Consumer-Grade ECG Device Using GA Optimization and Segment-Wise Analysis
Authors:

Abu Saleh Khan, Areen Ramesh Patil, Nipun Verma, Prakrititz Borah and Sakshi Arora
Abstract: Wearable stress monitoring requires models that are both accurate and efficient for real-time use on edge devices, especially when trained on low-sampling-rate signals. This study develops and compares two stress-detection methods: a lightweight machine-learning (ML) pipeline using HRV features and a deep-learning (DL) architecture trained on raw ECG. ECG from 51 healthy adults was collected under a standardized four-phase protocol (Baseline, Stroop, Mental Arithmetic, Recovery), yielding 20 hours of data. From these signals, 43 HRV features were extracted, and a Genetic Algorithm (GA) was used to reduce the feature set by 48.8% by selecting the most informative subset. With GA-optimized features, XGBoost achieved ∼74% accuracy on 60-second segments, while the proposed hybrid DL architecture CTBFNet with GA features achieved 79%. Although CTBFNet offered slightly higher accuracy, the ML models required 99.91% less memory and 99.99% less computation time, making them more suitable for resource-constrained deployment. Overall, this study shows that GA-enhanced HRV features allow lightweight ML models to approach DL performance while retaining practical efficiency.
Download
Paper Nr: 337
Title:

Multimodal Estimation of Chronic Stress Using Physiological Signals and Cognitive Task Performance
Authors:

Jiayi Hu, Yuki Ban and Shin'ichi Warisawa
Abstract: Current chronic stress estimation methods often lack accuracy and interpretability due to their reliance on single-modality data. To address this limitation, we propose a multimodal approach that integrates physiological signals with cognitive task performance. Forty-three participants completed a 5-minute relaxation phase followed by a 5-minute auditory 2-back task while physiological signals (ECG, EDA, BVP, TEMP, RIP, NIRS) were recorded. Our results demonstrate that multimodal integration substantially outperforms single-modality baselines. Specifically, a balanced fusion strategy (FUSE-BAL) achieved the highest performance, with an F1 score of 0.791 and a balanced accuracy of 0.763. Notably, features capturing temporal dynamics-such as slope and variability-proved far more discriminative than traditional static metrics. By combining physiological regulation markers with cognitive flexibility indices, this framework offers a robust and interpretable solution for objective chronic stress estimation.
Download
Paper Nr: 377
Title:

A Comparative Study of Deep Learning Approaches for Leishmania Detection in Microscopic Images
Authors:

Eduardo Monteiro, Marcelo Nogueira and Elsa Ferreira Gomes
Abstract: Leishmaniasis is a neglected tropical disease caused by protozoan parasites that predominantly affects the world’s poorest populations. Global estimates indicate approximately 700,000 to 1 million new human cases each year. Climate change, together with increased dog travel to endemic regions, is contributing to a northward expansion of the parasite’s habitat, leading to emerging cases in previously non-endemic areas of northern Europe. This work proposes a machine learning–based approach for the automatic identification of Leishmania in microscopic images, supporting automated in vitro diagnosis. The study encompasses dataset adaptation, preprocessing, data preparation, model design, and hyperparameter optimisation, and compares the proposed method with state-of-the-art approaches. The best performing model is a transfer learning architecture based on ResNet50, in which the pre-trained convolutional backbone is kept frozen and a lightweight classification head, composed of global average pooling, dense layers, and dropout,is trained on top. This model achieved 99% accuracy, a loss of 0.02, and an F1-score of 99% on the test set, demonstrating the effectiveness of leveraging pre-trained deep networks for robust and efficient Leishmania detection.
Download
Paper Nr: 435
Title:

Modular Framework for Comparative Analysis of EMG Detection Methods: Application to Wearable Interfaces for Persons with Motor Neuron Diseases
Authors:

Carolina Amante, Catarina Consolado, Gabriel Pires, Ana Rita Londral and Cláudia Quaresma
Abstract: Motor neuron diseases, particularly amyotrophic lateral sclerosis (ALS), progressively impair motor control, limiting communication and autonomy and reinforcing the need for assistive technologies capable of operating reliably as function declines. These interfaces can also support health data collection that is relevant for disease monitoring, in particular, the extent of the neurodegenerative process. This work introduces a modular framework for developing and refining electromyography (EMG)-based human–computer interfaces (HCIs), centred on muscular contraction detection. Using EMG data from 64 healthy participants and 11 persons with ALS, the framework applies a Modular Analysis Approach and Design Space Exploration to assess detection strategies under ALS-specific signal heterogeneity and limited data availability, aiming for adaptability and generalisation. Three methods were examined: an amplitude-based threshold, an energy-based threshold using wavelet-domain multi-band decomposition, and a deep learning model combining convolutional neural networks and long short-term memory layers. Results showed that standard signal conditioning stabilized EMG signals but slightly reduced model performance. Incorporating ALS data improved generalisation to pathological signals and a median-based energy threshold enables dynamic, user-adaptive calibration. Multi-band energy combination did not yield meaningful improvements, indicating that total-band energy already captures the most relevant activity information. This framework offers a reproducible basis for testing and integrating detection modules and supports future exploration of multimodal or threshold-free learning-based detection strategies for wearable assistive technologies.
Download
Paper Nr: 444
Title:

Fall Risk Classification in Older Adults Using Wearable IMU Data: A Comparative Study of Validation Strategies and Model Complexity
Authors:

Judith Urbina-Córdoba, Shuning Han, Zhe Sun, Ryutaro Himeno, David Arcos and Jordi Solé-Casals
Abstract: Fall prevention in elderly populations is closely related to the detection of gait abnormalities, which may indicate fall risk and underlying health issues. Recent advances in wearable IMU sensors and machine learning have enabled automatic fall risk assessment from gait data. In this work, we analyze IMU signals recorded from the left foot of 18 elderly volunteers during a 10-meter walking task. Two validation strategies are explored. First, a leave-one-subject-out (LOSO) cross-validation strategy is applied to an LSTM-based model for binary fall risk classification. Second, 157 gait features are extracted and ranked using ANOVA and mRMR, and several machine learning models, including logistic regression and MLPs, are evaluated using a leave-two-subjects-out strategy combined with LOSO cross-validation. The results show that the LSTM model performs poorly under LOSO due to dataset size and subject variability, whereas simpler models achieve better and more stable performance. Logistic regression yields the best results, with a balanced accuracy close to 0.69. Overall, this study highlights the limitations of deep learning models under strict cross-validation when applied to small datasets and emphasizes the importance of appropriate validation strategies for reliable fall risk assessment.
Download
Short Papers
Paper Nr: 50
Title:

Consensus Ranking via Entropy-Based Multi-Objective Optimization: Application of Selection of Biomarker Candidates
Authors:

J.-P. Conge, V. Vigneron, H. Maaref and L. T. Duarte
Abstract: Aggregating multiple ordered lists (or score scales) into a single consensus ranking is a recurrent need in decision support, information retrieval, and biomedical feature selection. In EEG biomarker screening, for instance, candidate features may be judged under several clinical and signal-quality criteria, which often leads to partially conflicting rankings. We propose an entropy-based multi-objective aggregation framework that converts heterogeneous inputs (total orders or scales) into a unified optimization problem. The method minimizes a discrepancy objective that can be instantiated with entropy-related divergences (e.g., Kullback–Leibler, Hellinger, or Pearson χ 2 ) and can incorporate rank disagreement through Kendall-type distances, while enforcing feasibility constraints on the aggregated distribution. The resulting continuous solution is then mapped to a discrete consensus permutation through a linear assignment formulation. We demonstrate the approach on an EEG biomarker candidate set evaluated by multiple criteria and show how the proposed aggregation produces a single ordered shortlist that is consistent with the global trade-offs visible in the PCA biplot (Fig. 1). To situate the contribution, we additionally compare against standard rank-aggregation baselines (e.g., Borda and median/Kendall aggregation) using agreement and stability metrics. We finally discuss computational complexity, scalability to larger candidate sets, and practical limitations, with guidance for realtime or large-scale deployments.
Download
Paper Nr: 59
Title:

Wavelet Analysis of non-Stationary Heart Rate Variability During Functional Tests
Authors:

S. V. Bozhokin and I. B. Suslova
Abstract: This article proposes a method to study heart rate variability (HRV), involving a new model of non-stationary heart rate tachogram as a frequency-modulated signal and the use of the double continuous wavelet transform (DCWT) for signal processing. This approach allowed us to obtain new characteristics of non-stationary HRV and to study how breathing test and tilt tests reveal the adaptive abilities of subjects. The rhythm characteristics obtained for all subjects made it possible to conduct static and dynamic clustering based on the strength of the functional test's influence on their physiological state under stress. The potential applications of the proposed method for HRV analysis during various functional tests are discussed, as well as the possibilities of influencing the heart rate using biofeedback.
Download
Paper Nr: 76
Title:

A Two-Stage Cascaded Ensemble Based on CRNN and Markov Chain for Sleep Apnea Detection Using ECG
Authors:

Faustine Faccin, El-Hadi Djermoune, Pauline Guyot and Laurent Bougrain
Abstract: Sleep-related irregular breathing and apnea involve periodic and cyclical decreases or interruptions in airflow, which may occur with or without obstructions of the upper airway. With nearly a billion people affected by this sleep disorder worldwide, its screening represents a major medical issue. Its early diagnosis is all the more important as numerous studies have highlighted the correlation between the presence of an untreated sleep apnea syndrome (SAS) and neurocognitive and cardiovascular consequences. Thus, in order to speed up and improve the diagnostic management of patients, researches have been conducted towards noninvasive and portable screening methods. Some of the latter are based on the patient’s cardiac activity, which is closely linked to the respiratory signal and easily recordable. In this paper, a new approach based on cascaded false-prediction-correcting ensemble using a hybrid deep model and Markov Chain is presented to detect sleep apnea events from nighttime long-term single-lead electrocardiograms (ECG), taken from the Apnea-ECG Database. The effectiveness of this approach is demonstrated through its capability to detect pathological ECG segments with a sensitivity, specificity and accuracy of 95.8%, 80% and 86%, respectively.
Download
Paper Nr: 78
Title:

Tensor Decomposition-Driven Variational Autoencoder: Biomarker-Aware OCT Classification
Authors:

A. H. Ahamada, A. Hazan, H. Maaref, T. Q. Syed and V. Vigneron
Abstract: We propose convolutional variational autoencoders (VAEs) whose encoder/decoder are structured by matrix and tensor decompositions (NMF, Tucker/NTD, and CPD). Factorization–low-dimensional representations–replaces large weight tensors by low-rank factors, reducing parameter counts while preserving reconstruction quality and improving interpretability. On a retinal optical coherence tomography (OCT) biomarker task for neovascular age-related macular degeneration (nAMD) activity, the symmetric CPD-NN variant reduces reconstruction error by about 49% relative to a compact ConvVAE baseline while maintaining high IoU. On geometric shapes, we achieve up to ≈ 35% parameter reduction with comparable IoU, highlighting the efficiencyperformance trade-off. Classification on latent features remains competitive, supporting the usefulness of factorized latent representations. Code and configurations are available at Github URL for reproducibility.
Download
Paper Nr: 81
Title:

Fully Convolutional Denoising Autoencoder with Skip Connections for Noise Reduction in (Simulated) Respiratory RIP Signals
Authors:

Redona Brahimetaj, Sofia Yfantidou, Maarten Gijssel, Vangelis Lympouridis, Silvia Zaccardi and Bart Jansen
Abstract: Chronic Obstructive Pulmonary Disease (COPD) is a life-threatening condition that impairs airflow, causing chronic respiratory symptoms. Continuous monitoring is essential for effective management, but methods like spirometry are impractical for long-term use. Respiratory Inductance Plethysmography (RIP) belts offer a non-invasive alternative but are highly susceptible to noise. This study proposes a fully convolutional denoising autoencoder (FCN-DAE) with skip connections to denoise RIP belt signals while preserving their integrity and key respiratory features. We first simulate clean (ground-truth) RIP signals and then apply several augmentation techniques to introduce noise. To evaluate the FCN-DAE performance, we analyze four signal variations: (a) clean ground-truth signals, (b) noisy augmented signals, (c) Butterworth-filtered signals (as a baseline filter), and (d) FCN-DAE denoised signals.We assess FCN-DAE performance using mean absolute and mean squared reconstruction error. Additionally, we extract several physiological features from all corresponding signal variations and compare features from the noisy, filtered, and denoised signals against those from the clean ground-truth ones to determine how well key respiratory features are preserved. The proposed FCN-DAE with skip connections demonstrates significant improvements in RIP belt signal denoising, highlighting its potential for wearable applications. To the best of our knowledge, this is the first study to apply a FCN-DAE with skip connections for this purpose.
Download
Paper Nr: 101
Title:

Temporal Dynamics of Functional Connectivity During Neuroaesthetic Experience: Effects of Art Expertise and Psychological Traits
Authors:

Pietro Tarchi, Lorenzo Frassineti, Lucrezia Bigagli, Valeria Pitterà and Antonio Lanatà
Abstract: In recent years, neuroaesthetics has highlighted how the experience of art engages brain networks related to perception, cognition, and emotion. In this study, electroencephalography (EEG) was recorded from 35 participants (22.86 ± 1.96 years) while they viewed 70 artworks. Functional connectivity was estimated with the Phase Lag Index across theta, alpha, beta, and gamma bands, and graph-theoretical metrics were derived. Analyses were conducted separately for early (0-1 s) and later (1-2 s) post-stimulus epochs, comparing experts and non-experts, as well as high- and low-empathy individuals (defined based on Interpersonal Reactivity Index scores). Results revealed that expertise exerted its strongest influence during the first second of processing, characterized by reduced theta transitivity and betweenness and increased alpha participation, particularly for familiar artworks. In the second epoch, expertise effects became more selective, with differences in beta and gamma metrics. Empathy effects, by contrast, were weaker in the first second, whereas in the second epoch empathic individuals exhibited more robust modulations, including higher path length, modularity, and altered clustering and centrality across alpha, beta, and gamma bands for high-valence and high-arousal artworks. These findings suggest that art expertise predominantly shapes early, familiarity-driven connectivity, whereas empathy modulates later affective stages of processing.
Download
Paper Nr: 106
Title:

Auto-Encoder Powered Genetic Algorithm-Based Dimensionality Reduction and Classification of in Situ Mid-Infrared Markers for Strawberry Maturity
Authors:

Mageda A. A. Sharafeddin, Adam Abdel Karim, Nour Harajli, Marwa Zeineddine and Abbass Rammal
Abstract: Rapid, low-cost, non-destructive monitoring of strawberry chemistry is key for precision harvest and postharvest quality Portable MEMS-MIR spectrometers provide rich spectra, yet many wavelengths are redundant We propose a two-stage pipeline: first, dimensionality reduction via PCA or AE or VAE compresses spectra spanning 14 maturity stages into 32-D latent vectors that preserve salient chemical cues while suppressing noise; second, a Genetic Algorithm (GA) performs feature selection directly in the latent space GA chromosomes encode subsets of latent variables; fitness combines cross-validated accuracy and parsimony using Random Forest (RF), Naive Bayes (NB), and Support Vector Machine (SVM) classifiers, encouraging compact, generalizable models On an independent test set, the best VAE+GA configuration achieved 98.9% accuracy with macro-precision, macro-recall, and macro-F1 all equal to 0 98 In head-to-head comparisons, the pipeline exceeded PSO (94.38%), ABC (92.13%), and BGWO (94.24%), and matched GSA (98%) Our model is interpretable for robust, accurate, fine-grained maturity classification for real-time use.
Download
Paper Nr: 122
Title:

Waveform Similarity Analysis of the Pupil Light Reflex for Detecting Dementia Symptoms
Authors:

Minoru Nakayama, Wioletta Nowak and Anna Żarowska
Abstract: A procedure which can predict the level of dementia was developed using metrics of waveform similarity of the pupillary light reflex (PLR) in response to chromatic light pulses. Metrics such as cross-correlation, coherence, and phase differences were used to characterise waveform similarities between PLR responses in three distinct phases in order to emphasise the features of chronological pupillary responses during PLR. Using only the extracted features, the prediction model could not extract a sufficient number of dementia patients, but the accuracy of estimation improved significantly when these features were combined together with data from the participant’s age category. The contribution of each feature was evaluated, and the findings corresponded with known physiological characteristics of PLRs.
Download
Paper Nr: 145
Title:

Counting without Seeing: Toward Acoustic Population Estimation from Unsupervised Audio Features
Authors:

Aysenur Arslan-Dogan and Aki Härmä
Abstract: Estimating the number of vocalizing individuals from audio recordings remains a central challenge in passive acoustic monitoring (PAM). A conceptual framework for acoustic population size estimation is proposed, based on two complementary strategies: a top-down approach using regression to predict group size from acoustic features (Direct Estimate of Size, DES), and a bottom-up approach employing unsupervised clustering to recover recurring individual vocal signatures (Count of Detected Individuals, CDI). Proof-of-concept results are demonstrated for both strategies: DES models trained on synthetic flamingo choruses are shown to accurately estimate group size, while CDI is used to recover individual counts in a zoo aviary with discrete callers. When tested on real zoo data, the DES model is found to differentiate between low- and high-density environments, although saturation occurs beyond the training range. These findings, complementary to the framework, suggest a promising path toward scalable, label-efficient population estimation from biosignals.
Download
Paper Nr: 165
Title:

Feasibility Study of Spirometric Parameter Estimation from Exhaled Breath Sounds Using Time–Frequency Representations and Deep Learning
Authors:

A. Salvador-Navarro, J. De la Torre-Cruz, A. Muñoz-Montoro, P. Revuelta-Sanz, J. M. Cruz-Molina, R. P. Paiva and F. J. Canadas-Quesada
Abstract: Respiratory diseases are a major global health concern, requiring accessible and reliable tools for lung function assessment. Although spirometry remains the gold standard, its use is often limited by equipment availability and patient cooperation. Recent advances in machine learning (ML) enable the estimation of spirometric parameters from respiratory sounds, offering a non-invasive and low-cost alternative. This work investigates the feasibility of estimating forced vital capacity (FVC), forced expiratory volume in one second (FEV1) and peak expiratory flow (PEF) from exhaled breath sounds. Several time–frequency (TF) representations and configurations are benchmarked to determine the most effective approaches, including hybrid combinations that integrate complementary information. A deep learning (DL) framework based on pre-trained convolutional neural networks (CNNs) is developed to automatically extract features and perform parameter regression. Preliminary results confirmed the feasibility of estimating spirometric parameters from respiratory sounds, obtaining R2 = 0.51 for FVC using STFT, R2 = 0.38 for FEV1 using Mel-spectrograms and R2 = 0.47 for PEF using CQT, encouraging further research in this emerging and clinically relevant field.
Download
Paper Nr: 167
Title:

Formal Grammar Based Syntactic Analysis of Heart Sounds
Authors:

George Manis
Abstract: Formal grammars have been extensively used in theoretical computer science, compiler construction technology, linguistics, mathematics, as well as in many other fields, including description and syntactic recognition in biosignal analysis and bioinformatics. Heart auscultation is a fast, non-invasive, cost-effective examination method, giving invaluable information about the heart. Each cardiac cycle consists of a series of sounds, some of which are characterized as normal and some indicating abnormal heart function. In this paper, we express the heart sounds sequence, employing a formal grammar description enhanced with attributes. The attributes are used to encode signal and diagnosis related information. The proposed grammars can syntactically recognize series of normal heart sounds S1 and S2, the heart sounds S3 and S4, as well as the presence of S1 and S2 split, whilst it can indicate a normal heart function or possible underlying pathology. The grammars have been developed and tested with the ANTLR meta-compiler tool.
Download
Paper Nr: 170
Title:

Investigating the Applicability of ECG Signals for Neonatal Seizure Detection Using ECG Spectrograms and CNN-ViT
Authors:

Kimia Rezaei, Sean R. Mathieson, Gordon Lightbody, Geraldin B. Boylan and William P. Marnane
Abstract: This study evaluates the effectiveness of ECG signals for automatic seizure detection by leveraging a large neonatal dataset and two different approches. The first approach, proposed in this study, directly extracts time-frequency information from ECG signals, ommiting traditional HRV feature computation to preserve more signal information. A CNN-ViT model is then designed to apply on these spectrogram for seizure detection. Second approach provides a basis for comparison with existing studies by extracting HRV features and using Random Forest and Multi-Layer Perceptron models for seizure detection. Finally, when evaluated on a fully unseen, patient-independent dataset, the proposed CNN-ViT approach achieved an AUC of 61.57, slightly outperforming the classical methods. Additionally, both approaches are evaluated on different subsets of the dataset to analyze the results with respect to patient-specific data and the distribution of seizure events in each subset.
Download
Paper Nr: 181
Title:

Detecting Cardiac and Respiratory Rhythms in Few-Second Auscultation Audio with Group Monte Carlo Singular Spectrum Analysis
Authors:

Sandra Ranilla-Cortina, Luciano Sánchez, José Ranilla, Alejandro Salvador-Navarro and Jose R. Gutierrez
Abstract: Estimating heart rate (HR) and respiratory rate (RR) from stethoscope audio is valuable in triage, bedside monitoring, and telemedicine, but accuracy drops when only a few seconds of noisy data are available. We propose a subject-specific method based on Group Monte Carlo Singular Spectrum Analysis (Group MC–SSA): multiple short recordings from the same subject are embedded in a common lagged space to learn an interpretable basis, and candidate oscillations are tested against per-recording first-order autoregressive (AR(1)) colored-noise surrogates for finite-sample, calibrated detection. On ten auscultations (3–7 s), we detect cardiac rhythm in 10/10 subjects and respiratory rhythm in 8/10, with MAE 8.18 bpm (HR) and 2.88 rpm (RR; detected cases). Compared with non-negative matrix factorization (NMF) baselines, Group MC–SSA achieves severalfold lower error without pre-trained dictionaries or heuristic thresholds, providing reproducible decisions in the few-cycle regime.
Download
Paper Nr: 187
Title:

Validation of an Imagined Speech Stimulation Protocol and EEG Investigation of Neurophysiological Differences between Rest and Imagined Speech
Authors:

Francesco Iacomi, Alessandro Elia, Gabriele Furnari, Harman Kaur, Andrea Farabbi, Riccardo Barbieri and Luca Mainardi
Abstract: This study addresses the topic of Imagined Speech (IS), a cognitive process through which an individual mentally pronounces one or more words without producing sounds or performing articulatory movements. In particular, this work aims to validate an experimental protocol describing its practical application during the data acquisition phase, analyzing the collected data in order to investigate neurophysiological differences between IS and REST states. The protocol validation is conducted through a dual approach: a qualitative analysis, based on participant feedback, and a quantitative analysis, focused on attention-related indices, including both spectral indices and eye blink metrics. The neurophysiological investigation relies on a visual analysis of the spatial distribution of EEG power across characteristic frequency bands, using topographic maps. The study involved 36 participants, each completing multiple sessions involving the mental repetition of 110 words presented in randomized order. Results indicate a higher effectiveness of the proposed protocol, which introduced a new structure division with more frequent pauses and a progressive reduction in the number of words, in order to help maintain attention levels and preserve EEG signal quality. The neurophysiological analysis revealed increased power in the delta, alpha, and theta bands during rest, and higher beta power during the IS condition. A pronounced activation was observed in frontal, central, and parieto-occipital regions, with a clear hemispheric asymmetry predominantly in the left hemisphere.
Download
Paper Nr: 241
Title:

Influence of Multi-Wavelength Sensor Array Design on Signal Stability and Application in Blood Pressure Detection
Authors:

Yuke Zheng, Long Yu, Ruishi Zhou, Yuchen Hu, Hongwei Wang, Hui Liu, Xuesong Ye and Congcong Zhou
Abstract: Hypertension has emerged as a critical global public health challenge, underscoring the pressing need for continuous, non-invasive, and reliable blood pressure (BP) monitoring technologies. Wearable devices have garnered increasing attention as a promising solution to meet this demand. This paper analyses the influence of multi-wavelength sensor design in wearable systems on the accuracy and broader applicability of BP measurement techniques, drawing upon principles from photoplethysmography (PPG), optical spectroscopy and signal processing methodologies. The integration of multi-wavelength sensing represents a paradigm shift in wearable BP monitoring, offering enhanced noise suppression, improved physiological specificity, and greater adaptability across diverse populations. Despite ongoing challenges related to device miniaturization, power efficiency and algorithmic robustness, recent advances in sensor engineering and data-driven modeling approaches hold substantial potential for enabling the next generation of wearable BP monitors. These innovations may ultimately transform hypertension management in both clinical and home-based healthcare settings.
Download
Paper Nr: 256
Title:

Open-Source Control PCB for Automated Behavioral Biomarker Acquisition
Authors:

Saumarí Negrón-Santos, Charlie Montes-Rivera, Juan González-Sánchez and Keven J. Laboy-Juárez
Abstract: High-throughput screening has transformed drug discovery, yet psychiatric research still lacks scalable platforms capable of capturing behavioral biomarkers with precision and reproducibility. We present an integrated electronics system for automated rodent training and monitoring that consolidates all control and acquisition functions into a compact, modular PCB architecture, eliminating external DAQ hardware and PC-hosted control; a low-power Raspberry Pi supervises video and logging. The platform couples a Raspberry Pi 5 supervisory unit for orchestration and video capture with an Arduino Nano 33 IoT and real-time clock (RTC) for deterministic event timing. Modular sensor-actuator boards control levers, lick ports, solenoids, speakers, and lighting arrays. Hardware validation demonstrates event-to-actuation latencies of 3.08 ± 0.27 µs, sufficient for alignment with neural recordings. Dual levers and lick ports enable independent reward types and schedules (e.g., water vs. sucrose) with calibrated 0.0208 ± 0.0029 mL per-pulse delivery. The firmware supports automated task scheduling, synchronized dual-camera recording, and continuous data logging. These results establish a reproducible, open-source infrastructure for high-throughput behavioral biomarker acquisition in preclinical research.
Download
Paper Nr: 266
Title:

DA-NN: A ResNet-Based Architecture with Domain Adaptation for Automatic EEG Artifact Identification
Authors:

Emmanuel de Jesús Velásquez-Martínez and Miguel Ángel Porta-García
Abstract: Electroencephalography (EEG) is a fundamental tool for studying brain activity; however, EEG signals are frequently contaminated by physiological and non-physiological artifacts that affect interpretation and diagnostic accuracy. Accurate artifact classification during preprocessing is therefore essential for reliable EEG analysis. While this problem has been addressed using both traditional and deep learning methods, many deep learning models rely on either synthetic data or real expert-labeled datasets and often exhibit degraded performance when evaluated on data differing from the training set, indicating limited generalization due to insufficient domain adaptation. To address this limitation, this study proposes a Domain Adaptation Neural Network (DA-NN) that generalizes EEG artifact identification across synthetic and real signals. The proposed semi-supervised architecture reduces the domain gap and achieves robust performance, reaching 91.59% classification accuracy.
Download
Paper Nr: 275
Title:

Neuro-Acoustic Fusion: EEG Biomarkers and Deep Audio Embeddings for Music-Evoked Emotion Recognition
Authors:

Amulya J. Naik, Anagha Madhusoodhana, Anagha Vyakarnam, Aritro Maiti and Gowri Srinivasa
Abstract: Emotion recognition during music perception is challenged by inherent inter-subject variability in neural responses. To this end, we propose a multimodal deep learning framework incorporating physiologically interpretable EEG biomarkers with stimulus-derived acoustic descriptors. The model fuses Frontal Alpha Asymmetry (FAA) and Theta–Beta Asymmetry (TBA) with high-level YAMNet embeddings extracted from the music stimuli. A specially designed EEGNet-based architecture performs the fusion and captures effectively the temporal-spectral EEG dynamics. The experiments on the EREMUS dataset demonstrate that the proposed framework outperforms EEG-only and audio-only baselines consistently and reaches an accuracy of 88.3% in binary valence and discrimination. Inclusion of asymmetry-based EEG features has a strong interpretability and computational efficiency while deep audio embeddings contribute to complementary emotional states.
Download
Paper Nr: 277
Title:

Gender- and Channel-Specific Electroencephalography Markers for Attention Deficit Hyperactivity Disorder Assessment
Authors:

Muhammet Aksakal and Beren Semiz
Abstract: Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental condition characterized by heterogeneous neurophysiological patterns. Electroencephalography (EEG) offers a non-invasive tool for objective ADHD assessment; however, existing findings vary substantially across genders and brain regions. This study investigates gender- and channel-specific EEG characteristics associated with ADHD using a pub-licly available dataset comprising five electrodes (Cz, F3, F4, Fz, O1) recorded from 79 adults during multiple resting-state and cognitive tasks. A total of twenty-two temporal, spectral, and statistical features were extracted. Feature significance was assessed using the Mann–Whitney U test, and classification was performed via a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel using five-fold cross-validation. When all channels and tasks were included, the model achieved high classification performance, with accuracies of 0.97 for all subjects, 0.98 for males, and 0.93 for females. Among individual channels, the frontal F4 electrode yielded the highest accuracy (0.99 for males, 0.97 for females), outperforming central (Cz) and occipital (O1) channels. Results revealed global alterations in low-frequency (delta and theta) and alpha-band features for males, while females exhibited localized deviations confined to frontal regions. These findings confirm that ADHD-related EEG patterns are influenced by both gender and spatial channel selection. Overall, the results emphasize the potential of channel-specific, gender-aware EEG analysis for enhancing the precision and personalization of ADHD classification.
Download
Paper Nr: 290
Title:

Multi-Method Signal Decomposition Reveals Neural Network Disruption in Claustrophobia: A Multi-Scale Connectivity Analysis
Authors:

Seyed Amir Tabatabaei Hosseini, Dunya Moradi and Beren Semiz
Abstract: Background: Traditional EEG spectral analysis fails to detect consistent neural alterations in claustrophobia due to inability to capture non-stationary, multi-scale dynamics. Advanced signal processing offers promise to uncover subtle neurophysiological signatures of phobic stress. Objective: To characterize neural network disruption in claustrophobia using Fourier Decomposition with Hilbert Transform-based EEG Signal Analysis (FHESA), Phase Lag Index (PLI) functional connectivity, and fractal dynamics assessment. Methods: EEG was recorded from 9 claustrophobic and 13 control participants during open space (R0), moderate (T1) and tight (T2) confinement. Instantaneous features were extracted using FHESA from Fourier Intrinsic Band Functions; beta-band connectivity was quantified using PLI across fronto-temporal, fronto-parietal, and inter-hemispheric pathways; fractal scaling was evaluated via Hurst exponent (H) and power-law exponent (β). Results: FHESA revealed significant group differences with large to exceptional effect sizes (Cohen's d=1.38–3.69, all p<0.05) at central and parietal electrodes. PLI showed progressive connectivity loss in claustrophobics, with significant fronto-temporal (p=0.023, d=−0.54) and inter-hemispheric (p=0.008, d=−0.61) reductions at T2. Fractal analysis revealed significant, state-dependent H–β coupling exclusively in claustrophobic participants, indicating systematic deviations from canonical linear fractal scaling. Conclusions: These findings establish state-specific, multi-scale alterations in neural integration in claustrophobia and validate these methods as novel EEG biomarkers for phobic anxiety.
Download
Paper Nr: 291
Title:

Differentiating between Multiple Sclerosis and Neuromyelitis Optica Spectrum Disorder Using IMU-Measured Spatiotemporal Gait Characteristics
Authors:

Dunya Moradi, Kardelen Akar, Hussein Youssef, Ayse Altintas, Atay Vural and Beren Semiz
Abstract: Multiple Sclerosis (MS) and Neuromyelitis Optica Spectrum Disorder (NMOSD) are two distinct neurological disorders that often present similar clinical symptoms. However, due to their differing pathogenesis, treatment protocols, and long-term outcomes, accurate distinction between these two disorders is crucial for appropriate medical interventions. This study introduces a non-invasive, wearable sensor-based framework that integrates inertial measurement unit (IMU) data with machine learning (ML) models to distinguish between patients with MS and NMOSD. IMU-based gait parameters were collected from 45 participants during a standardized 500-meter walking (500MW) task. A comprehensive set of 70 spatiotemporal gait features were extracted from triaxial accelerometer and gyroscope signals. Principal Component Analysis (PCA) reduced feature dimensionality, with the first five components explaining 70% of the total variance. Seven supervised ML classifiers were then developed using 5-fold cross-validation, with Support Vector Machine (SVM-RBF) achieving highest performance: accuracy of 85.5%, F1-score of 85.0%, and area under curve (AUC) of 0.91. Statistical analysis using the Kruskal-Wallis test revealed significant inter-group differences in gait parameters, particularly in cadence and terminal double support duration. For the best performing model (SVM-RBF), different numbers of PCA components (2-20) were also experimented. Although the performance increased up to 5 components, minimal improvement was observed beyond 5 components, confirming substantial information redundancy in the original feature space. These findings suggest that wearable IMU sensors, coupled with machine learning, offer a promising diagnostic tool for capturing disease-specific gait patterns, potentially enhancing early detection, differential diagnosis, and personalized treatment planning in neuroinflammatory care.
Download
Paper Nr: 299
Title:

Opportunistic Teacher Forcing: Training RNNs on Wearable Signals with Inherent Missing Data
Authors:

Sutashu Tomonaga, Andres Hernandez-Matamoros, Haruo Mizutani and Kenji Doya
Abstract: Wearable devices enable longitudinal monitoring of human physiology but suffer from inherent missing data, hindering analysis. Traditional recurrent neural network (RNN) training assumes complete datasets, while imputation introduces biases that corrupt the signal’s underlying dynamics. We propose Opportunistic Teacher Forcing (OTF), an extension of Scheduled Sampling that inherently handles missing data without imputation or architectural modifications. OTF switches probabilistically between teacher-forced and autonomous modes, critically computing loss only on available timesteps. We evaluate OTF using a Multi-Timescale RNN (MT-RNN) on surrogate data (Lorenz 63 system with up to 70% bursty missing) and real-world wearable data (ECG-like signals with ∼30% missing). Results demonstrate a robust reconstruction of the system’s geometry (via time-delay embeddings) and spectra (via Hellinger distance), validating OTF as a promising method for dynamical system reconstruction from incomplete biomedical data.
Download
Paper Nr: 312
Title:

Anomaly Detection in cECG Data Using Unsupervised Learning
Authors:

Lavinia Goldermann, Pascal Raabe, Markus Lüken, Steffen Leonhardt and André Stollenwerk
Abstract: Unobtrusive monitoring of vital parameters enables continuos and long-term health monitoring in everyday settings, but is challenged by frequent artefacts. Capacitive electrocardiography (cECG) allows unobtrusive ECG acquisition, but is highly susceptible to motion and environmental influences. In this study, we apply the neural-network based Anomaly Detection (AD) model DeepAnT to multi- and single channel cECG and optical photoplethysmography (PPG) data from the UnoViS bed dataset. To address the scarcity of medical labels, the model is trained fully unsupervised using unlabelled data, while a manually annotated dataset is utilized for validation. We evaluated univariate and multivariate configurations using both raw and manually cleaned training data. Results indicate that DeepAnT effectively detects anomalies with F1 scores ranging from 60.2 % and 100 %. Notably, multivariate models combining cECG and PPG consistently outperformed univariate approaches, demonstrating that sensor fusion compensates for single-channel signal loss.
Download
Paper Nr: 318
Title:

Multimodal Analysis of Human Movement Patterns Using a Smart Wearable T-Shirt and Force Platform
Authors:

Laura Valente, Cláudia Quaresma and Carla Quintão
Abstract: Studying human movement is essential for understanding behavior and motor functions, allowing the distinction between healthy and pathological patterns. Multimodal approaches, integrating data from multiple sensors, have been proven promising, enabling holistic characterizations. However, sensor integration challenges, the lack of standardization of experimental protocols and the small size of normative databases still limit the generalization of many studies. The present study sought to adapt methodologies from previous studies to characterize the movement patterns of 40 healthy participants, integrating physiological, kinetic, and kinematic data, collected using a Hexoskin smart shirt, a PLUX force platform, and inertial and electromyography sensors. Beyond continuous signal analysis, specific parameters were extracted and assessed for both their consistency throughout the protocol and the interrelationships between them. An illustrative application of this protocol was also included to compare movement patterns across two different age groups. The proposed protocol and methodology proved feasible, with successful integration of all measurement systems. Parameter correlations were confirmed, showing moderate associations between center of pressure instability, angular variability of the center of mass, and physiological measures in both static and dynamic conditions. Additionally, muscle activation patterns revealed co-contraction of homologous muscles predominantly the multifidus and greater activation of contralateral muscles during reaching tasks.
Download
Paper Nr: 354
Title:

Differentiating Pre-REM and Post-REM N2 Sleep Stage Using Machine Learning Models
Authors:

Miriam Allalouf, Tali Bitan, Lior Ketz, Eva Kimel, Safa Natsheh and Ilana S. Hairston
Abstract: Sleep is key to physical and mental health. While the regulation of sleep stages has been extensively studied, the dynamics of state transitions remain less well understood. The present study examined the ability to identify the transition into Rapid Eye Movement (REM) sleep by differentiating stage N2 immediately before REM (pre-REM) from N2 immediately after REM (post-REM), using computational analysis of polysomnographic (PSG) data. For this purpose, data from 28 participants were analyzed using XGBoost and Fully Connected Network models. The analysis was performed both at the individual participant level and across all participants combined. The results indicate that, at the individual level, the models distinguished between pre-REM and post-REM with accuracies above 88% for half of the participants and between 65% and 86% for the rest. Training and testing on data from a group of participants yields the average accuracy of the group's individuals. Evaluating the generalization of the pre-REM substage across participants resulted in a significant performance drop, often nearing chance level, indicating limited cross-subject transferability. However, a hybrid approach, training on multiple participants’ data plus a small portion of the target participant’s data, maintained high accuracy even when only 20% of personal data was used.
Download
Paper Nr: 365
Title:

Comparison of the Accuracy and Reliability of Artificial Intelligence and Human Assessment in the Diagnosis of Flatfoot in the Younger Population
Authors:

Jovana Kuzmanovic Pficer, Zorana Bukvic, Lazar Krstic, Filip Milanovic and Dejan Nikolic
Abstract: This cross-sectional study aimed to assess the accuracy and reliability of footprint analysis by researchers, by training the algorithm to recognize the same anatomical landmarks and calculate the Chippaux–Smirak index (CSI). This research included 200 children aged from 7 to 11 years from schools in the Niš district, Serbia. After obtaining the original print from the pediatrician, four copies were made. The three copies were distributed to three observers who performed manual measurements. The fourth copy was scanned and processed using an artificial intelligence (AI) algorithm specifically developed to analyze footprint images. The results showed that AI is a highly reliable and clinically relevant method for assessing flatfoot in children. AI demonstrated strong agreement with measurements from three examiners, as evidenced by high correlations (r>0.90) and intraclass correlation coefficients (ICC > 0.95). Balancing the dataset and aligning it with medical experts during annotation significantly improved the model's stability and precision, and enhanced the reliability of anatomical landmark detection. These findings highlight the critical importance of training data quality for the performance of AI-based systems in clinical settings. This is a very efficient, objective, and reproducible tool for automated assessment of flatfoot in children.
Download
Paper Nr: 406
Title:

Estimating Brain Activity of Neonates Using Physics Informed Networks
Authors:

Aleksandar Jeremic, D. Nikolic, G. Djuricic, N. Milcanovic and Z. Jokovic
Abstract: Source localization of electrical activity in newborn infants is important from two standpoints. From an academic standpoint such insights can enable better understanding of brain development and from clinical standpoint localization of electrical activity can identify regions of the brain with higher than usual activity and possibly improve possible treatment outcomes. The electrical activity and the corresponding electroencephalography (EEG) measurements are dependant on electrical properties of brain and skull tissue i.e. corresponding conductivities and geometry. In this paper we propose a technique for estimating brain activity of the neonates using physics informed neural networks. We assume that the brain activity can be modelled using a set of spatially distributed dipoles and train the PINN using finite element method by enforcing physics constraints at EEG sensor locations. To illustrate the applicability of the proposed method we use numerical simulations with realistic geometry obtained from magnetic resonance imaging of the neonatal skull.
Download
Paper Nr: 441
Title:

A Comparative Evaluation of Open-Source Gait Event Detection Methods Using a Single Sacrum-Mounted IMU
Authors:

Elena Botti, Redona Brahimetaj, Eva Swinnen and Bart Jansen
Abstract: Inertial measurement units (IMUs) provide a potential reliable alternative for out-of-the-lab gait analysis, particularly when the setup is simple, as in the case of a single sacrum-mounted sensor. However, estimating spatio-temporal gait parameters from inertial signals requires accurate detection of gait events (GEs), which allow to subdivide gait into cycles. This study benchmarks two open-source IMU-based GE detection algorithms (SensorMotion and SciKit-Digital-Health) using a public dataset of healthy adults walking at slow, comfortable, and fast speeds. Reference timings were obtained by comparing marker-derived events to force plate data using two vertical ground reaction force thresholds (20 N and 100 N), highlighting the influence of this parameter on GE timings. Across all trials, both IMU based methods managed to detect initial contacts within 40 ms, with detection rates up to 83%. In contrast, final contact (FC) detection was substantially less reliable, particularly at higher speeds, with lower detection rates and larger mean absolute errors. Although no algorithm was consistently superior across all metrics, our results favoured SKDH overall. These findings confirm that IMU-based algorithms are still suboptimal, and FC detection remains the primary limitation for sacrum-mounted IMU GE detection.
Download
Paper Nr: 56
Title:

Possibilities to Improve Dementia Classification through Data Balancing Using Augmentation by Synonym Replacement
Authors:

Attila Zoltán Jenei, Dávid Sztahó and Gábor Kiss
Abstract: Dementia is one of the most common neurodegenerative diseases, which becomes more prevalent with age. Early detection is crucial for starting the proper treatment as soon as possible to mitigate the symptoms. Recognition in the early stages is difficult because the symptoms can be confused with signs of normal aging. On this basis, several biomarkers have been researched, including speech and its transcription. In our research, the Pitt dataset is utilized to investigate the correction of data imbalance through sample augmentation via synonym replacement. Our target groups were the dementia and the healthy control groups. CountVectorizer, BERT, and miniLM models were utilized for feature extraction in conjunction with SVM classifiers, employing nested cross-validation. The large language models were also examined with a fine-tuning setup. The results showed that the use of augmentation for class balance can improve dementia detection in healthy populations compared to cases where augmentation was not used. For the BERT models and CountVectorizer, augmentation yielded the best results in terms of balanced accuracy compared to the original dataset. Furthermore, it was found that augmentation also helped to balance sensitivity and specificity.
Download
Paper Nr: 67
Title:

SIMUS-Based Modeling for Radial Artery Ultrasound Characterization
Authors:

Aurélia Leandri, Louis Lecrosnier, Adel Ghazel and Bastien Faure
Abstract: This paper presents a novel and comprehensive methodology for configuring the MATLAB UltraSound Toolbox (MUST), specifically its SIMUS (SIMulator for UltraSound) module, to achieve high-fidelity ultrasound modeling of the human wrist, with an emphasis on radial artery characterization. Three types of anatomical model were used as simulation input : schematic anatomy, ultrasound image, and MRI-derived. They were evaluated to assess their influence on simulation realism and measurement precision. A cross-correlation algorithm was applied to estimate artery position and diameter from simulated signals. The 2D version of the algorithm, particularly with MRI-based input, achieved the highest accuracy, with localization errors of 1 transducer width and diameter errors of 23% and 4% respectively in lateral and axial direction. These findings highlight the value of high-fidelity input for ultrasound simulation and multidimensional processing for enhancing vascular measurements.
Download
Paper Nr: 70
Title:

Predicting Signs of Problematic Internet Use to Analyze Children’s Health Using Machine Learning
Authors:

Fahad Layth Malallah and Kamran Iqbal
Abstract: Recently, the increasing use of the Internet by children has raised issues regarding their Internet addiction, which impacts their health. This paper aims to predict the impairment severity on children's health by building a data model to analyze children's physical activity and fitness to identify early signs of problematic internet use (PIU). The prediction is reported as four categories of impairments: none, mild, moderate, and severe. These are called the Severity Impairment Index (SII). Extensive experiments have been conducted using the dataset provided by the Healthy Brain Network in New York City, with the support of the Child Mind Institute. This dataset contains 3961 records, in which each child's information is represented as a data record. Furthermore, each record contains 80 features related to the child's behavior, which were collected by a questionnaire. Several deep learning algorithms have been applied to this dataset. It turns out that Gradient Boost Regression (GBR) is the best algorithm for training and testing, as it obtained 100% accuracy when training 70% of the dataset, and the remaining was used for the unseen testing set, a 10-fold cross-validation technique. To sum up, assessing the health of children who are addicted to Internet use can be made possible to help clinicians quickly and accurately by exploiting artificial intelligence technology.
Download
Paper Nr: 153
Title:

Fully Automated Recanalization Prediction in Ischemic Stroke Patients Using Multimodal MRI
Authors:

Rafia Ahsan, Sofía Vargas-Ibarra, Vincent Vigneron, Hichem Maaref and Sonia Garcia-Salicetti
Abstract: Acute ischemic stroke remains a leading cause of death and long-term disability worldwide, and timely, effective reperfusion is critical to improving patient outcomes. However, prospective prediction of treatment success remains challenging: current assessments of recanalization often rely on retrospective or manual evaluations–e.g., the Arterial Occlusive Lesion (AOL) score–and image interpretation is highly dependent on expert radiologists working under time pressure. To address this, we propose an AI-based, fully automated pipeline to predict recanalization success one hour after intravenous thrombolysis using multimodal MRI. Our approach combines automated segmentation of ischemic lesion and thrombus regions, extraction of radiomic features across multiple MRI modalities, and feature selection with ANOVA and Fisher scores, followed by binary classification. Evaluated on a single-center cohort of 288 patients, the method achieved a test accuracy of 75.9% using automatically generated masks, supporting the clinical potential of standardized radiomic features for automated recanalization prediction.
Download
Paper Nr: 191
Title:

Assessing Stimuli Detectability and Pleasantness for Auditory BCI
Authors:

Lenaïg Guého, Laurent Bougrain, Cyril Plapous, Patrick Hénaff and Rozenn Nicol
Abstract: Brain-Computer Interfaces (BCIs) enable device control by analyzing brain activity. In reactive auditory BCIs based on steady-state auditory evoked potentials, users are exposed to amplitude-modulated sine waves at given frequencies that encode information (i.e. the type of action expected), while their brain activity is analyzed to infer the intended action based on the frequency retrieved. However, listening to sine-wave may be perceived as unpleasant over time. This study compares the use of pure-tones with alternative sounds, including artificial stimuli (such as a Brownian noise) and natural sounds (such as cicada song and cat’s purr) by measuring brain responses of 48 subjects to these different stimuli, all amplitude-modulated at 40 Hz. The Signal-to-Noise Ratio (SNR) (i.e. the ratio between the power spectrum of electroencephalographic signals in response to the target stimulus and that in response to a silence stimulus) is computed at 40 Hz for each type of stimulus. It reveals that the 40-Hz modulation frequency is clearly more identifiable when carried by a pure tone than when carried by the other sounds, with an SNR increase up to more than 5 dB. The cicada song stimulus is a promising alternative, still requiring improvement to achieve the level of detectability observed for pure tones. The experiment is conducted at two different sound levels to assess whether increasing the listening level increases the SNR, but the opposite trend is found. Questionnaires indicate that more than half of the participants find pure tones annoying and prefer other sounds, confirming that this study is worth pursuing.
Download
Paper Nr: 271
Title:

Automated Diagnosis of Respiratory Diseases from Respiratory Sounds: A Reproducibility Case Study
Authors:

Diogo Pessoa, João Garcia, Juan De La Torre-Cruz, Francisco Cañadas-Quesada and Rui Pedro Paiva
Abstract: The automatic diagnosis of respiratory diseases using lung sound recordings has attracted growing attention due to advances in machine learning and the increasing availability of open-access respiratory databases. However, many studies in the field report near-perfect results that are difficult to reproduce and seldom translate to real-world clinical contexts and applications. In this work, we present a reproducibility case study in which we replicate a published deep learning model for pulmonary disease classification based on convolutional and recurrent neural networks, by reproducing the original methodology and correcting its methodological flaws-most notably, the presence of data leakage arising from patient overlap between training and testing sets-we demonstrate that previously reported results were overly optimistic. By enforcing patient-level data separation, we observed a significant drop in the model’s performance, suggesting limited generalization. This study highlights the importance of transparent and reproducible research practices, rigorous experimental evaluation setups, and the development of cross-database and domain-adaptive models to ensure clinically reliable and generalizable computer-aided diagnostic systems for respiratory sound analysis.
Download
Paper Nr: 276
Title:

Quantitative EEG Biomarkers for Cognitive Decline: A Comparative Study of Frontotemporal Dementia and Alzheimer’s Disease
Authors:

Aisa Sadat Takyar and Beren Semiz
Abstract: Frontotemporal Dementia (FTD) and Alzheimer’s Disease (AD) are the two most prevalent forms of earlyonset dementia, both associated with marked alterations in cortical electrophysiology. While neuroimaging offers anatomical insight, electroencephalography (EEG) provides a direct and temporally precise window into neuronal dysfunction. This study proposes a standardized quantitative EEG analysis framework to identify spectral and temporal biomarkers that differentiate FTD, AD, and healthy controls. Resting-state EEG recordings from 88 participants were preprocessed and analyzed to extract spectral and temporal features. Statistical analysis (Kruskal–Wallis with Dunn’s correction) confirmed significant group-level differences in Peak Alpha Frequency (PAF), relative alpha/theta power, and cognitive performance (MMSE). Machine learning models trained on these features successfully distinguished dementia patients from controls (accuracy of 0.78 ±0.03) and achieved comparatively strong differentiation between AD and FTD (accuracy of 0.75 ± 0.14). Feature importance and SHAP analyses identified PAF, MMSE, Hjorth complexity, and gamma power as key discriminants, reflecting characteristic cortical slowing and altered signal complexity across subtypes. These findings demonstrate the potential of EEG-derived biomarkers for supporting differential dementia diagnosis and highlight the importance of standardized, interpretable analysis pipelines for clinical translation.
Download
Paper Nr: 279
Title:

Comparative Evaluation of Rule-Based and Machine Learning Models for EEG-Based Psychological Stress Detection
Authors:

Elvin Güngör and Beren Semiz
Abstract: Psychological stress represents a fundamental physiological response that influences cognitive, emotional, and behavioral regulation. This study investigates the potential of electroencephalography (EEG) to identify neural activity patterns associated with stress using both rule-based and machine learning approaches. A publicly available EEG dataset was analyzed through time-domain and spectral features, including the Theta (4–7 Hz), Alpha (8–13 Hz), and Beta (14–30 Hz) frequency bands. Two classification models were employed: a Boolean threshold model and an Extreme Gradient Boosting (XGBoost) classifier. The XGBoost model demonstrated superior performance, achieving an accuracy of 78% compared to 71% from the Boolean model in distinguishing stress from resting states. Feature-importance analysis revealed that increased frontal Beta activity and reduced parietal Alpha power were the most discriminative indicators of stress. Overall, EEG-based modeling of stress responses offers promising insights into the neurophysiological mechanisms underlying emotional regulation and can support the development of data-driven tools for mental health monitoring and early intervention.
Download
Paper Nr: 313
Title:

Anomaly Detection in Intensive Care Unit Data: A Comparative Analysis
Authors:

Simon Fonck, Lavinia Goldermann, Sebastian Fritsch, Lena Olivier and André Stollenwerk
Abstract: We present a comparative analysis of anomaly detection (AD) methods applied to vital sign time series data from intensive care units (ICUs). Our study evaluates five unsupervised deep learning-based and three statistical or rule-based AD algorithms with respect to their ability to identify anomalous data points in a binary anomaly classification setting. We have access to physician-annotated datasets of 44 ICU patients. Of these, 38 data sets were used for training and six for evaluation. In addition, selected deep learning models were retrained on a larger dataset comprising 3,000 patients to assess the effect of training set size. Performance is assessed using the F1 score by comparing algorithm outputs with expert annotations. Among all methods, OmniAnomaly and DeepAnT achieved the highest F1 scores, reaching up to 98.38 % and 78.07 %, respectively.
Download
Paper Nr: 366
Title:

A Machine Learning-Based System for Human Movement Analysis
Authors:

Mauro Aloisio Smaniotto, Fabio Kurt Schneider and Heitor Silvério Lopes
Abstract: The analysis of human movement has gained prominence in sports and physical assessment applications, driven by computer vision tools capable of identifying and quantifying biomechanical patterns. Although advanced video-based systems with markers exist, their high cost and complexity limit clinical use. In this context, markerless solutions based on Deep Learning models, such as convolutional networks for detecting body points, offer low-cost alternatives, suitable for home use and requiring less preparation. This work presents a real-time motion capture and analysis system built upon MoveNet, a CNN-based pose estimation model, capable of identifying joints, generating biomechanical reports containing range of motion and cadence, providing immediate feedback through RAG (Red–Amber–Green) indicators, and allowing analysis with simultaneous cameras or pre-recorded videos. The angle validation against Kinovea software (reference method), using Bland-Altman analysis with 100 samples, demonstrated excellent agreement: mean bias of -0.036°, narrow LoA (-0.408° to 0.336°), and minimal variability, indicating an absence of systematic error and precision exceeding the limits considered excellent in the literature. Applications for this software includes rehabilitation, post-injury follow-up, exercise prescription, sports monitoring, and support for clinical decision-making, strengthening the use of decentralized and accessible motion capture.
Download
Paper Nr: 367
Title:

Image-Based Evaluation System for Paraná State Military Physical Fitness Test: Pull-Ups
Authors:

Andrio Ramos dos Santos, Mauro Aloisio Smaniotto and Fabio Kurt Schneider
Abstract: This paper presents the use of a software for real-time motion capture and analysis, capable of identifying joints, detecting the appropriate execution of an exercise that is part of a typical Physical Fitness Test (PFT) of the Paraná State Military Force. The exercise evaluated in this paper is the Pull-Up where along with other details, the elbows must go through three main states: full extension, flexion, and full extension again. The software can provide immediate feedback through retrivial-augmented generation (RAG) indicators and real time counting. The application was validated using Bland-Altman analysis with 18 people and software (SW) and 16 human-based evaluations. The software (SW) demonstrated excellent agreement: The system validation was conducted using videos from 18 participants, comparing three methods: the reference, the proposed automated system, and the average of trained evaluators, through descriptive statistics and Bland–Altman analysis. The results showed excellent agreement between the proposed and the reference methods, with zero bias, low variability, and narrow limits of agreement, while human evaluations exhibited greater inconsistency. The system demonstrated a strong ability to reduce subjectivity and provide an objective, standardized alternative for counting repetitions.
Download
Paper Nr: 368
Title:

Image-Based Evaluation System for Paraná State Military Physical Fitness Test: Push-Ups
Authors:

Tarcisio Maico Goncalves Da Silva, Lucas Daniel Ortiz Arias, Mauro Aloisio Smaniotto and Fabio Kurt Schneider
Abstract: This paper evaluates the PUSH-UP exercise that is part of typical Physical Fitness Test (PFT) applied to the Military Forces. Here, the study is applied to the Paraná State Military Force considering their protocol for validation of repetitions. Although it is a simple exercise, some features such as body alignment in a plank position, chest approximation to a fixed reference point 10 cm from the ground, full extension of the elbows at the end of the movement need to be evaluated simultaneously. A system using Machine Learning models based on Convolutional Neural Networks is used for detecting body joints and angles are computed along with the sequential execution of the human movements in order to support the evaluators in counting repetitions. The application was validated using Bland-Altman analysis with 15 people and SW and human-based evaluations. The push-up counting using the proposed SW method against the CEFID method (reference method) demonstrated excellent agreement, with a mean bias of 1.267 and limits of agreement (LoA) from −9.882 to 12.415. Among all the comparisons analyzed, these LoA were the narrowest, indicating the lowest variability between methods. The statistical consistency and reduced dispersion observed between CEFID and SW make this comparison the most reliable and precise among the evaluated methods.
Download
Paper Nr: 384
Title:

Gender and Age Modulate the Diving Reflex: Quantifying Bradycardia Dynamics and Autonomic Shifts via HRV and Exponential Modeling
Authors:

O. Barquero-Pérez, R. Goya-Esteban, A. Luque-Casado, D. Grassi and F. Suárez
Abstract: The diving reflex, a protective mechanism enabling apnea through bradycardia, remains incompletely understood in its demographic modulation. We investigated gender and age-related differences in 44 healthy volunteers (17 men, 30.05±13.30 yrs; 27 women, 24.62±10.71 yrs). After a 5-min baseline, participants performed three cold-water facial immersion apneas, with HR monitored via Firstbeat®. Bradycardia was modeled using f (t) = a+ce(aut), where au quantifies RR-interval increase rate. Women showed significantly higher au during the first apnea (auf emale = 0.035±0.008 vs. aumale = 0.025±0.006, p < 0.05), indicating faster bradycardia onset. This difference diminished in later apneas, suggesting adaptive dampening. LF/HF ratio was higher in women at baseline (3.1 vs. 2.9), but reversed during apnea (peak: 5.0 in men vs. 2.9 in women), revealing greater sympathetic drive in men. Aging significantly attenuated the response: participants >59 yrs showed near-zero au in the third apnea and elevated LF/HF (>4.0), indicating reduced vagal tone and sympathetic dominance. These findings demonstrate that both sex and age critically shape autonomic dynamics during the diving reflex, supporting the need for personalized physiological models.
Download
Paper Nr: 442
Title:

An Exploratory Study on SVM-Based Fall-Risk Classification Using Hybrid IMU-Derived and Clinical Features
Authors:

Judith Urbina-Córdoba, Shuning Han, Zhe Sun, Ryutaro Himeno, David Arcos and Jordi Solé-Casals
Abstract: Falls in older adults are a major public health problem, as they are a primary cause of injury and accidental death, with substantial impact on healthcare systems. There is interest in fast fall notification and in developing reliable, easy-to-use, and cost-effective methods to prevent them. One approach uses accelerometer signals from wearable IMU sensors to classify movement patterns into fall-risk categories. Support Vector Machines (SVMs) are commonly used for this task, but their generalization and integration with traditional clinical tests remain open challenges. This work presents a case study on 182 Japanese adults aged 65 to 89, describing a fall-risk classification criterion, a feature extraction method from acceleration signals, a feature importance analysis comparing gait-derived and clinical-test variables, and five SVM models with different kernels. Models were trained using 5-fold cross-validation, and testing results showed that the best-performing model was the linear-kernel SVM trained with the six highest-ranked Minimum Redundancy Maximum Relevance (MRMR) features. The feature extraction method produced many missing values, and feature importance analysis indicated that clinical-test variables were generally more informative than gait-derived metrics. The final model achieved 83.33% accuracy, but per-class metrics showed poor detection of the high fall-risk class (F1-score: 40%). In conclusion, this study contributes an incremental step toward automating fall-risk prediction using wearable sensors and machine learning.
Download
Paper Nr: 443
Title:

Resting Surface Electromyography as an Indicator of Motor State in Parkinson’s Disease
Authors:

Ester Muñoz del Campo, Carlos Pérez-López, Jordi Solé-Casals, Miriam Grande Gordon, Irene Gonzalo Asenjo, Lúa Márquez López, Marta Montesinos Terceño and Patricia Rodríguez Sánchez
Abstract: Gait alterations are among the most informative indicators of motor state and disease severity in Parkinson’s disease (PD); however, gait-based monitoring is inherently limited to periods of movement. This study investigates the feasibility of using surface electromyography (sEMG) recorded during resting conditions to discriminate PD motor states in the absence of voluntary movement. Multimodal wearable data were collected from 74 PD patients during ON and OFF medication states within a semi-free-living protocol. sEMG signals from bilateral quadriceps were segmented into 30-second windows corresponding to periods without leg movement and labeled using the MDS-UPDRS rigidity score. Binary classification models were trained to distinguish absence of rigidity (score 0) from moderate to severe rigidity (score ≥3). Feature-based machine learning models achieved robust performance with ROC-AUC values up to 0.84, while a hybrid Convolutional Neural Network (CNN) + Long-Short Term Memory (LSTM) deep learning architecture reached an AUC of 0.98. These findings demonstrate that clinically relevant motor state information is encoded in resting sEMG activity. Resting sEMG therefore represents a valuable complementary modality to gait analysis, enabling continuous motor state assessment during sedentary periods in Parkinson’s disease.
Download