Objectives: Observing the activities of the elderly in natural life is a crucial issue nowadays to better understand their potential behavioral changes and predict risks. To this end, a comprehensive hardware and software infrastructure has been designed by a multidisciplinary team of researchers and pre-tested in a smart flat lab. It enables to collect relevant data and develop algorithms to analyze activities and detect changes such as falls, wandering or other risky situations. This study was carried out in a shared house by 12 independent elderly people. The study focuses on episodes of falls in the house, and analyzes mobility behavior before and after falls to observe the person's rehabilitation in the home. Materials and Methods: Each resident's room and the two shared spaces were equipped with motion and magnetic contact sensors to record movements and entry/exit activities. 9 months of data were collected and analyzed, highlighting patterns of activity and changes in these behaviors, particularly when a fall occurred and then when the usual behavior returned, if at all. Two levels of analysis were implemented: the detection of deviation in activity indicators for each individual, and the detection of drift in the established behavior pattern over time. The classification technique used to extract the patterns is the Kmeans partitioning algorithm. We also used the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) method to validate the performance of the K-means method. Results: Data analysis was carried out on the 4 falls recorded during the observation period, involving 4 of the house's occupants. The results highlight the relationship between model conduct and events related to falls and returns from hospitalization. Detection was validated by share house carers' annotations, acting as a ground truth, on the days when falls occurred. The first results of pattern recognition with clustering methods show that the K-means method provides more convincing results than the DBSCAN method. In this study, by observing the movement signals of residents who fell during the course of the study, we were able to identify characteristic post-fall behaviors. (c) 2025 Published by Elsevier Masson SAS on behalf of AGBM.

Background/Objectives: Diagnostic accuracy studies are essential for the evaluation of the performance of medical tests. The risk of bias (RoB) for these studies is commonly assessed using the Quality Assessment of Diagnostic Accuracy Studies (QUADAS) tool. This study aimed to assess the capabilities and reasoning accuracy of large language models (LLMs) in evaluating the RoB in diagnostic accuracy studies, using QUADAS 2, compared to human experts. Methods: Four LLMs were used for the AI assessment: ChatGPT 4o model, X.AI Grok 3 model, Gemini 2.0 flash model, and DeepSeek V3 model. Ten recent open-access diagnostic accuracy studies were selected. Each article was independently assessed by human experts and by LLMs using QUADAS 2. Results: Out of 110 signaling questions assessments (11 questions for each of the 10 articles) by the four AI models, and the mean percentage of correct assessments of all the models was 72.95%. The most accurate model was Grok 3, followed by ChatGPT 4o, DeepSeek V3, and Gemini 2.0 Flash, with accuracies ranging from 74.45% to 67.27%. When analyzed by domain, the most accurate responses were for “flow and timing”, followed by “index test”, and then similarly for “patient selection” and “reference standard”. An extensive list of reasoning errors was documented. Conclusions: This study demonstrates that LLMs can achieve a moderate level of accuracy in evaluating the RoB in diagnostic accuracy studies. However, they are not yet a substitute for expert clinical and methodological judgment. LLMs may serve as complementary tools in systematic reviews, with compulsory human supervision.

The use of electrohysterogram (EHG) - uterine muscular activity signals - along with artificial intelligence models would help in the prediction of preterm delivery and thereby save the lives of many early-delivered infants through the necessary early medical care. The informative portions in the EHG recording which could aid in the prediction of this risky delivery are the ones related to the uterine muscular contractions. Hence, the more we perform a precise segmentation of these contraction signals, the more valuable and discriminative are the data provided to the AI models, which will in turn reflect a better learning process and prediction outcome. This paper presents a new algorithm called Slope of Tangent (SOT) for an enhanced segmentation of uterine muscular contraction signals. The method is further compared to the up-to-date uterine contraction automatic segmentation methods. The results showed that the method allowed a higher number of full detections ($F D=90$) and partial detections ($P D=263$) in comparison to the wavelet $H 2$ nonlinear correlation method (wavh2) ($F D=42$ and $P D=145$) and the sample entropy method ($F D=13$ and $P D=138$). Further work should be done in order to reduce the number of false detections given by the SOT method. In addition, the method should be further validated on open-source databases.

The robust prediction of the infant delivery term through the cooperation of artificial intelligence (AI) and electrohysterogram (EHG) would enable the appropriate early medication for possible premature delivery, thus avoiding death risk or sequels. This paper focuses on unraveling the best preprocessing scheme to be used when dealing with the classification of uterine muscular contraction signals. In addition, the study discusses the impact of several EHG denoising techniques on the prediction outcome. Hypergraph neural network (HGNN) is employed to evaluate the different preprocessing steps. The results show that it is better to start by segmenting the contractions and concatenating them, then standardizing the resulting signal or normalizing it between -1 and 1, and finally segmenting the contractions again in order to characterize them by a set of features that are used to finally train the classifier. The accuracy achieved by the HGNN given this methodology was 89.2%. Moreover, the use of the conventional EHG denoising methods such as canonical correlation analysis (CCA), empirical mode decomposition (EMD), and their combination (EMD-CCA) did not improve the prediction results. As a conclusion, proper preprocessing of the EHG signals would lead to a great improvement in the prediction process.Clinical Relevance— This study presents a new effective approach for preprocessing EHG signals to improve the prediction of delivery term.

The current epidemiology of the global population highlights an increasing number of aged individuals living with decreasing autonomy. Modern societies face the significant challenge of caring for persons with various disabilities, whether in institutions or at home. Health systems are ill-prepared in terms of staffing, economics, and public policies to manage the growing population of elderly individuals. This article discusses how technology can alleviate isolated lives, reduce hazards, and enhance human relationships, particularly the physician-patient relationship. Our interdisciplinary group focuses on developing innovative technologies to be implemented in the coming decades to improve the living conditions of elderly populations.

Stress detection is crucial for maintaining individual well-being and preventing potential health complications. In this paper, we explore the use of Convolutional Neural Networks (CNNs) for stress detection based on Heart Rate Variability (HRV) signals. The study leverages the Multi-Modal Dataset (MMSD), a comprehensive repository encompassing HRV data from 74 subjects across distinct phases: relaxation, stress induction, and recovery. We explore the effectiveness of CNNs compared to traditional machine learning algorithms in discerning stress states using HRV data. Results reveal promising performance of CNNs in accurately identifying stress, showcasing their potential for real-time stress monitoring applications. This research contributes to advancing the field of stress detection, offering insights into the application of deep learning techniques on physiological data for enhanced stress detection strategies.

BioImpedance Analysis (BIA) is a safe, simple, and noninvasive technology to measure body composition. By measuring the electrical impedance of biological tissues, BIA provides valuable biological insights such as body composition, hydration status, and some health conditions. The principle is to apply an electric current to body segments, which water content and conductivity are characteristics, and to determine the electric impedance depending on body tissues passed through. However, these measurements are indirectly related to body composition and intensively depend on limited and imprecise assumptions to estimate mathematical models. This is the source of methodological and experimental challenges. BIA is very promising to offer non-invasive and portable solutions to assess health status and well-being, but challenges must be considered: they impact technological limitations, methodological standardization, and data interpretation. Advancements in BIA require to address these hurdles to improve accuracy, reliability, and applicability in diverse settings. In this article, we reviewed in depth these challenges based on a systematic review of literature. Purpose: The objective of this systematic review is to identify key challenges of BIA to assess body composition to develop possible directions for improving this technology. Our review underlines clearly the need to reduce these challenges with the multiplication of biostatistical sources, the definition of personalized models, and the adjustment of mathematical assumptions, to improve BIA reliability and adoption in e -health or specific applications. Methodology: The objective of this systematic review from published literature was to answer the question: "How to assess whole body composition in the average human adult with BIA, what are the scientific challenges and limits for a wider adoption in medical practice?". We limited our research within Pubmed, ScienceDirect and IEEE complementary databases. Our research was carried out in English using the keywords "body composition" and "bioimpedance analysis" over a period from the included 1995 to 2022. We controlled inclusion criteria to collect only articles with average human adults' groups: age from 18 years, both males and females, mixed ethnics, BMI ranging from 18 to 30 kg/m2, either healthy or non -healthy status. We added the following exclusion criteria: athletics, malnourished, eating or mental disorders, pregnancy and menstrual period. Finally, we kept articles validated versus state-ofthe-art methods DEXA, or isotope dilution. Summary findings: Our literature review identified seven major challenges with BIA: Rheological modeling precision represent human body as an electrical circuit made of resistors and capacitors to reflect electrical properties of tissues; Body compartments to model human body as a combination of cylinders different tissues type and fluids volumes; Physiological approximations as anthropometric data used in body composition modeling refer to an ancient population from 1975 (ethnicity: Caucasian, body mass index: BMI=24, sex: male, height: 170 cm, age: 25 years, health status: healthy... ); Predefined constants to predict body composition were calculated on healthy subjects; Electrical stimulation frequency choice as the impedance depends on the value and the number of frequencies used for the measure; Flow of current inside the body may not be uniform nor following the same pathway crossing all body tissues and finally Standardization of measurement protocols and body position to minimize the interferences and factors affecting the accuracy of BIA measurement. Conclusion: BIA is simple, easy to use, and noninvasive technique integrated in portable, wearable, and connected health solutions. The complexity of rheological models cannot reflect precisely the complexity of the human body. The compartment numbers considered for tissues modeling are critical for results accuracy, the commonly used configurations are the 3-C and 5-C to predict body composition referring to standard methods. Numerous physiological assumptions introduce several factors of variability that must not be generalized, the assumptions should be applied on groups with similar characteristics as the population studied only and must include subjects specificities. The models assume the use of constant values that are generic, imprecise, and estimated on limited healthy groups, future work needs to customize population -specific equations. Multiplication of electrical stimulations at different frequencies is required to consider different types of tissues and to guarantee a response from all tissues. The measures are significantly influenced by electrodes positioning, gel and dry electrodes both imply trade-offs between accuracy, convenience, and mobility. There is no one -size -fits -all answer, nevertheless standardization of procedures is a step for BIA studies to move forward and subsequently improve accuracy and reduce the gaps when results from different devices are compared. From this review, it looks critical to improve BIA methods by developing novel electrodes designs that may improve electrical contact and reduce contact impedance or by exploring the use of smart textiles and wearable electrodes for continuous monitoring of body composition and hydration status. Acquiring more data, at several electrical stimulation frequencies and in different contexts (healthy and pathological status, ethnicities, ages, comorbidities...) to enrich references and adjust constant values. Analyzing large datasets to refine prediction models. These improvements are essential prerequisites so incorporation of machine learning and artificial intelligence algorithms can explore individual variability in the future and improve the potential benefits of BIA predictions in research and clinical practice. (c) 2024 AGBM. Published by Elsevier Masson SAS. This is an open access article under the CC BY license (http://creativecommons .org /licenses /by /4 .0/).

Objectives: Heart rate variability (HRV) is a valuable indicator of both physiological and psychological states. However, the accuracy of HRV measurements taken by wearable devices can be compromised by errors during transmission and acquisition. These errors can significantly affect HRV features and are not acceptable for precise HRV analysis used for medical diagnosis. This study aims to address this issue by investigating the effectiveness of four different interpolation methods (Nearest Neighbour -NN, Linear, Shape-preserving piecewise cubic Hermite -Pchip, and cubic spline) in tackling missing RR values in real-time HRV analysis. Materials and Methods: In this study, HRV signals were obtained from Electrocardiograms (ECG) through automatic detection and manually corrected by a specialist, resulting in high-quality signals with no missing or ectopic peaks. To simulate low-quality data acquisition, values were iteratively deleted from each HRV analysis window. The deleted values were then replaced using four different interpolation methods. Time and frequency domain features were computed from both the original and reconstructed signals, and the Mean Absolute Percentage Error (MAPE) was used to compare these features. Results: Results showed that as the percentage of missing values increased, some interpolation methods were more suitable for RR time-series with a greater number of missing data. Furthermore, the study suggests that the impact of interpolation on HRV features varied across different features and that SDNN is the least affected by interpolation. In the time domain, nearest neighbour interpolation gives the best results for up to 50% missing data. Beyond this threshold, it seems better not to use any interpolation for RMSSD. In the frequency domain however, the lowest errors of HRV feature estimation are obtained using linear or Pchip interpolation. To achieve maximum performance, it is recommended to adapt the interpolation method to both the percentage of missing values and the targeted HRV feature.Conclusion: Results highlight the importance of choosing the appropriate interpolation method to accurately estimate HRV features in real-time analysis. Overall, the Pchip interpolation seems to yield the best results on most HRV features as it preserves the linear trend of the data while adding very light waves. The findings can be beneficial in the development of more precise and reliable wearable devices for real-time HRV monitoring. (c) 2023 AGBM. Published by Elsevier Masson SAS. All rights reserved.

The experiment presented in this study aimed at eliciting two affective states at pre-determined periods including relaxation and stress using various stressors. We advance the hypothesis that it is possible to observe patterns and variations in physiological signals caused by mental stress. In this chapter, an exhaustive description of the experimental protocol for signal acquisition is provided to ensure both reproducibility and repeatability. Details are presented on the whole process from the choice of sensors and stressors to the experimental design and the collected data so that the potential user of our database can have a global view and a deep understanding of the data. Four physiological signals are recorded throughout the experiment in order to study their correlation with mental stress: electrocardiogram (ECG), photoplethysmogram (PPG), electrodermal activity (EDA) and electromyogram (EMG). A statistical analysis is carried out for preliminary results and for protocol validation before a deeper analysis using artificial intelligence algorithms in future work.

Drowsy driving is a major issue in road safety.In this paper, we propose a description of an experimental data collection to develop a drowsiness detection model.The objective of this data collection was mainly to gather physiological data of individuals in simulated driving situations.We designed a realistically annoying scenario to induce fatigue while staying close to real driving conditions.The experiment was run on an augmented reality platform called CAVE.The need for contextualization came early in the design of the experiment.Therefore, in addition to physiological data, we added much more data sources, from driving habits to driving behaviour in addition to self-assessment of fatigue levels and the gold standard (EEG).As a result, this experience helped us create a data set of physiological data completed by elements of context and driving behaviour.Thus allowing us to perform a very rich analysis of these physiological data.

As the French, European and worldwide populations are aging, there is a strong interest for new systems that guarantee a reliable and privacy preserving home monitoring for frailty prevention. This work is a part of a global environmental audio analysis system which aims to help identification of Activities of Daily Life (ADL) through human and everyday life sounds recognition, speech presence and number of speakers detection. The focus is made on the number of speakers detection. In this article, we present how recent advances in sound processing and speaker diarization can improve the existing embedded systems. We study the performances of two new methods and discuss the benefits of DNN based approaches which improve performances by about 100%.

Stress is an increasingly prevalent mental health condition across the world. In Europe, for example, stress is considered one of the most common health problems, and over USD 300 billion are spent on stress treatments annually. Therefore, monitoring, identification and prevention of stress are of the utmost importance. While most stress monitoring is carried out through self-reporting, there are now several studies on stress detection from physiological signals using Artificial Intelligence algorithms. However, the generalizability of these models is only rarely discussed. The main goal of this work is to provide a monitoring proof-of-concept tool exploring the generalization capabilities of Heart Rate Variability-based machine learning models. To this end, two Machine Learning models are used, Logistic Regression and Random Forest to analyze and classify stress in two datasets differing in terms of protocol, stressors and recording devices. First, the models are evaluated using leave-one-subject-out cross-validation with train and test samples from the same dataset. Next, a cross-dataset validation of the models is performed, that is, leave-one-subject-out models trained on a Multi-modal Dataset for Real-time, Continuous Stress Detection from Physiological Signals dataset and validated using the University of Waterloo stress dataset. While both logistic regression and random forest models achieve good classification results in the independent dataset analysis, the random forest model demonstrates better generalization capabilities with a stable F1 score of 61%. This indicates that the random forest can be used to generalize HRV-based stress detection models, which can lead to better analyses in the mental health and medical research field through training and integrating different models.

Medical data like physiological signals or others, are usually hard to collect, label and share. This is a huge problem because the unavailability of medical datasets will limit the development of machine learning models that can be of big benefit in the medical field. Electromyography (EMG) signals are a type of physiological signals that when available, can be used to train predictive models for motion recognition or muscle assessment. Collecting the EMG data can be hard due to the rarity of some diseases or the measuring being unachievable because of the condition of the patient (e.g. not being able to walk or properly contract their muscle). And finally, sharing these datasets is very limited due to privacy concerns where the identity of the source can be leaked which is a crucial problem. In this preliminary study, our aim is to present a solution to provide sEMG datasets that are big enough in size to train machine learning models. We explore a transformer based GAN model to create synthetic sEMG signals that can replace the real data and try to solve all the problems discussed above. In the first part of this work, we will discuss the generation process, and in the second part, the evaluation of the created ‘fake’ signals.