Wearable sensors with local data processing can detect health threats early, enhance documentation, and support personalized therapy. In the context of spinal cord injury (SCI), which involves risks such as pressure injuries and blood pressure instability, continuous monitoring can help mitigate these by enabling early deDtection and intervention. In this work, we present a novel distributed machine learning (DML) protocol for human activity recognition (HAR) from wearable sensor data based on gradient-boosted decision trees (XGBoost). The proposed architecture is inspired by Party-Adaptive XGBoost (PAX) while explicitly preserving key structural and optimization properties of standard XGBoost, including histogram-based split construction and tree-ensemble dynamics. First, we provide a theoretical analysis showing that, under appropriate data conditions and suitable hyperparameter selection, the proposed distributed protocol can converge to solutions equivalent to centralized XGBoost training. Second, the protocol is empirically evaluated on a representative wearable-sensor HAR dataset, reflecting the heterogeneity and data fragmentation typical of remote monitoring scenarios. Benchmarking against centralized XGBoost and IBM PAX demonstrates that the theoretical convergence properties are reflected in practice. The results indicate that the proposed approach can match centralized performance up to a gap under 1% while retaining the structural advantages of XGBoost in distributed wearable-based HAR settings.

The autonomous nervous system (ANS) response in neurological disorders is a direct modifiable risk factor for cardiovascular health, however, difficulty in remote monitoring and objective assessment has made it underrepresented in preventive healthcare. Particularly, autonomic dysreflexia (AD) is a dangerous hypertensive emergency, potentially life-threatening in people with spinal cord injury (SCI), yet detection outside clinical settings remains reactive and episodic. This study presents an interpretable and scalable framework for creating a digital biomarker from multimodal wearables in data scarcity through vital sign attribution analysis in multiple body locations, evaluated with 27 subjects undergoing clinical examination with objective blood pressure measurements. Our framework learns from diverse biosignals—lectrocardiography (ECG), photoplethysmography (PPG), bioimpedance, skin temperature, heart rate, and respiratory rate—proving robustness to sensor failure and a pathway to remote monitoring. This study identified heart rate and ECG as dominant predictors, with PPG providing complementary value, under simulated single-modality failure or noisy channels. This work advances a feasible path to equitable, remote monitoring of the ANS response, reducing dependence on intermittent BP measurements and enabling earlier intervention.

The pervasive integration of robots into daily life necessitates advanced human-robot interaction (HRI) capabilities, particularly the accurate understanding of human physiological and cognitive states. The current state of the widely used Robot Operating System (ROS2) lacks standardized mechanisms for representing and communicating human states. This paper introduces ROS 4 Healthcare (ROS4HC), a comprehensive open-source framework designed to standardize the acquisition, representation, and integration of human sensing data into robotic systems. ROS4HC provides unified message types, modular sensor drivers, signal processing libraries, and visualization tools for physiological, biological, and physical signals. This framework is validated through empirical case studies in healthcare robotics, including a heart rate (HR)-adaptive wheelchair velocity modulation, an autonomous treadmill system integrating physiological feedback, and a nocturnal monitoring system based on a robotic rocking bed. These case studies demonstrate that the framework enables modular component reuse, standardized communication, and interoperability for better human-robot integration. Beyond healthcare, we highlight ROS4HC’s generalizability for critical applications such as industrial safety, human-robot collaboration, and performance monitoring, establishing a standardized infrastructure for safer, more adaptive, and context-aware robotic systems across diverse domains.

[This corrects the article DOI: 10.3389/frobt.2026.1745197.].

Autonomic Dysreflexia (AD) is a potentially life-threatening condition characterized by sudden, severe blood pressure (BP) spikes in individuals with spinal cord injury (SCI). Early, accurate detection is essential to prevent cardiovascular complications, yet current monitoring methods are either invasive or rely on subjective symptom reporting, limiting applicability in daily file. This study presents a non-invasive, explainable machine learning framework for detecting AD using multimodal wearable sensors. Data were collected from 27 individuals with chronic SCI during urodynamic studies, including electrocardiography (ECG), photoplethysmography (PPG), bioimpedance (BioZ), temperature, respiratory rate (RR), and heart rate (HR), across three commercial devices. Objective AD labels were derived from synchronized cuff-based BP measurements. Following signal preprocessing and feature extraction, BorutaSHAP was used for robust feature selection, and SHAP values for explainability. We trained modality- and device-specific weak learners and aggregated them using a stacked ensemble meta-model. Cross-validation was stratified by participants to ensure generalizability. HR- and ECG-derived features were identified as the most informative, particularly those capturing rhythm morphology and variability. The Nearest Centroid ensemble yielded the highest performance (Macro F1 = 0.77+/-0.03), significantly outperforming baseline models. Among modalities, HR achieved the highest area under the curve (AUC = 0.93), followed by ECG (0.88) and PPG (0.86). RR and temperature features contributed less to overall accuracy, consistent with missing data and low specificity. The model proved robust to sensor dropout and aligned well with clinical AD events. These results represent an important step toward personalized, real-time monitoring for individuals with SCI.

The growing prevalence of chronic health conditions in aging populations highlights the need for innovative solutions in rehabilitation and long-term care. We propose a multimodal system designed to automatically classify Activities of Daily Living (ADLs) and, in the future, support the prevention of secondary health conditions in institutionalized elderly individuals. This system continuously integrates six commercially available wearable and nearable sensors to monitor ADLs over two weeks, ensuring data completeness and maintaining high data quality throughout the trial. In this study, we present pilot data from two residents of a Japanese elderly care facility, demonstrating the proposed system's feasibility and usability. The collected data was comprehensive and robust, with residents showing strong acceptance of the wearable and nearable technologies for long-term use. These findings underscore the potential of multimodal sensory systems to enhance rehabilitation strategies by enabling continuous health monitoring. Integrating ADLs monitoring into rehabilitation programs may facilitate early detection of health changes and support personalized interventions, building on the success of digital health tracking in other clinical domains.

Human Activity Recognition (HAR) is a valuable tool for healthcare and rehabilitation, enabling applications like remote patient monitoring and rehabilitation progress assessment. This paper introduces TIFEX-Py, a comprehensive Python toolbox designed for time series feature extraction in HAR. TIFEX-Py offers a rich set of 195 feature extraction methods across statistical, amplitude, spectral, and time-frequency domains. To evaluate its effectiveness, TIFEX-Py was applied to 11 publicly available HAR datasets: DSADS, HHAR, MHEALTH, MotionSense, PAMAP2, REALDISP, RealWorld, UniMiBSHAR, USC-HAD, WARD, and WISDM. Machine learning pipelines utilizing TIFEX-Py features, evaluated under both random and subject-stratified cross-validation settings, consistently achieved performance that is competitive with or superior to state-of-theart (SOTA) benchmark performances available for the datasets. In 11 out of 11 random split cross-validation scenarios, our pipeline surpassed or matched SOTA performance. For stratified by subject cross-validation, this was the case for more than half of the datasets. These results highlight the power of TIFEX-Py's feature space in representing time series data. TIFEX-Py is opensource and publicly available for researchers in rehabilitation and movement analysis fields.

In the ageing society, there is an increasing presence of chronic health conditions that can benefit from long-term monitoring technologies. Our proposed long-term care system aims to classify activities of daily living automatically and, in the future, to prevent secondary health conditions in institutionalised elderly individuals. We implemented a multimodal system comprising six commercial wearable and nearable sensors for tracking daily living functioning. In this work, we present pilot data from two residents of a Japanese elderly care facility. The results show the feasibility and acceptance of our multimodal sensory system. Data was successfully captured and stored while the wearables were well-accepted by the users for long-term monitoring. In conclusion, we argue that capturing activities of daily living from elderly individuals would enable future tracking of health changes, as has been demonstrated in specific digital health tracking in other conditions.

Current blood pressure (BP) estimation methods have not achieved an accurate and adaptable approach for application in populations at risk of cardiovascular disease, with generally limited sample sizes. Here, we introduce an algorithm for BP estimation solely reliant on photoplethysmography (PPG) signals and demographic features. Our approach automatically obtains signal features and employs the Markov Blanket (MB) feature selection to discern informative and transmissible features, achieving a robust space adaptable to the population shift. We validated our approach with the Aurora-BP database, compromising ambulatory wearable cuffless BP measurements for over 500 individuals. By evaluating several machine-learning regression methods, Gradient Boosting emerged as the most effective. The comparative assessment encompassed both a generic model (trained on unclassified BP data) and specialized models (tailored to each distinct BP population), with the former demonstrating consistent superiority with MAE of 10.2 mmHg (0.28) for systolic BP and 6.7 mmHg (0.18) for diastolic BP on the whole dataset. Moreover, a comparison of in-clinic and ambulatory model performance showed a significant decrease in accuracy for the latter of 2.85 mmHg in systolic (p < 0.0001, F-value = 32764.76) and 2.82 mmHg for diastolic (p < 0.0001, F-value = 65675.36) estimation errors. Our work contributes to a resilient BP estimation algorithm from PPG signals, underscoring the advantages of causal feature selection and quantifying the disparities between ambulatory and in-clinic measurements.

Monitoring activities of daily living (ADLs) for wheelchair users, particularly spinal cord injury individuals is important for understanding the rehabilitation progress, customizing treatment plans, and observing the onset of secondary health conditions. This work proposes an innovative sensory system for measuring and classifying ADLs relevant to secondary health conditions. We systematically evaluated multiple wearable sensors such as pressure distribution mats on the wheelchair seat, accelerometer data from the ear and wrists, and IMU data from the wheelchair wheels to achieve the best unobtrusive combination of sensors that successfully distinguished ADLs. Our work resulted in an XGBoost classifier with a 20-second window size and extracted features in statistical, time, frequency, and wavelet domains, with an average class-wise F1 score of 82% (with only 3 out of 12 classes being mislabeled). Our study results demonstrate that the newly investigated modality of the bottom pressure mat emerges as the most relevant information source for recognizing ADLs, while heart and respiratory rates did not provide added value for the selected set of ADLs. The proposed sensory system and methodology proved high quality in most classes and easily extendable for long-term monitoring in outpatient rehabilitation, with the need for an extended database of activities.

In recent developments of Human Activity Recognition systems (HAR), It has been found that deep learning models are being studied by researchers, especially convolutional neural networks integrated with long shortterm memory cells such as convolutional LSTM (ConvLSTM) networks. The deep structures require large datasets which demand extensive data collection. Therefore, various data augmentation methods are under focus nowadays. Furthermore, the challenge of time-series data augmentation is to choose the method that preserves the correct labels. In this paper, we evaluated and compared six data augmentation methods (i.e., autoencoder, time warping, amplitude warping, scaling, jittering, and linear combination) utilizing ConvLSTM networks for classification. Consequently, the WISDM dataset (tri-axial accelerometer signals of six activities) was augmented to the final size of 1.5 times the original dataset. Further, the proposed ConvLSTM model was trained seven times (once with the raw dataset and six times with the augmented dataset). The results indicated classification accuracy improvements for the test data from 92% to 93%, 97% and 98%, when training the models using augmented datasets, augmented using linear combination, scaling, and jittering methods respectively. Activity-wise analysis suggested the stairing activities to be the most challenging for the model to classify when the dataset was augmented by time warping, amplitude warping as well as autoencoder.

Human Activity Recognition (HAR) using wearable systems in telerehabilitation and clinical applications has caught the attention of many researchers, especially for Parkinson’s disease (PD) movement therapy. However, the distinction between simple activities and complex ones and how to handle them have not been thoroughly investigated. We propose and compare two variants of a multi-task network with shared parameters to recognize simple activities (SAs) and complex activities (CAs) simultaneously. We do so by introducing a branched deep neural network that uses a shared feature space for both SAs and CAs, and further enriches the features for CAs using a deep recurrent neural network. The variants are CNN-LSTM and CNN-BiLSTM. We trained and evaluated the models with 65 activities; 51 SAs and 14 CAs composed of Lee Silverman Voice Treatment-BIG (LSVTBIG) and functional activities. Our dataset consisted of 43 healthy subjects, seven women and 36 men. The data were recorded using four smart bands with embedded IMUs, placed on both wrists and both thighs. Our results show that the CNN-BiLSTM model with an average accuracy of 84.17% and 78.78% for SAs and CAs, correspondingly, outperforms the CNN-LSTM model with average accuracies of 71.83% and 66.46%.

Wearable human activity recognition systems (HAR) using inertial measurement units (IMU) play a key role in the development of smart rehabilitation systems. Training of a HAR system with patient data is costly, time-consuming, and difficult for the patients. This study proposes a new scheme for the optimal design of HARs with minimal involvement of the patients. It uses healthy subject data for optimal design for a set of activities used in the rehabilitation of PD1 patients. It maintains its performance for individual PD subjects using a single session data collection and an adaptation procedure. In the optimal design, several classifiers (i.e. NM, k-NN, MLP with RBF as a hidden layer, and multistage RBF SVM) were investigated. Features were signal-based in the time, frequency, and time-frequency domains. Double-stage feature extraction by PCA and fisher technique was used. The optimal design reached a recall of 95% on healthy subjects using only two sensors on the left thigh and forearm. Implementing the adaptation procedure on two PD subjects, the performance was maintained above 80%. Post analysis on the performance of the adapted HAR showed a slight drop in precision (above 87% to above 81%) for activities that was performed in sitting condition.

Inertial Measurement Units have become one of the most widely used instruments in Human Activity Recognition and clinical applications. Their performance needs to be validated against gold standard systems to be reliably used for any particular application. In this study, we validated MPU9250 in TTGO T-Wristband smart band (Shenzhen Xinyuan Electronic Technology© Ltd.) against MetaMotionR (MBIENTLABO, San Fransisco, USA) for the Lee Silverman Voice Treatment-BIG (LSVT-BIG) and functional activities. The validation metrics in this study are the Pearson’s correlation and Root Mean Square Error (RMSE) between acceleration and gyroscope readings, the Bland-Altman plots for the error in acceleration magnitudes, and linear regression on the error in acceleration magnitudes. Our results show that sensor readings between the IMUs were highly correlated, with a few exceptions. RMSE was mostly less than 0. 05g and 15°/s. Proportional biases between the error in acceleration magnitudes and the mean acceleration magnitudes were mostly insignificant, with the significant ones having a slope of no greater than 0.1 over a range of at most 2g. Our findings show that the TTGO T-Wristband is a valid choice for kinematic measurements for LSVT-BIG and functional activities.

This study proposes a new method for the detection of a weak scatterer among strong scatterers using prior-information ultrasound (US) imaging. A perfect application of this approach is in vivo cell detection in the bloodstream, where red blood cells (RBCs) serve as identifiable strong scatterers. In vivo cell detection can help diagnose cancer at its earliest stages, increasing the chances of survival for patients. This work combines time-domain US with frequency-domain compressive US imaging to detect a 20-μ MCF-7 circulating tumor cell (CTC) among a number of RBCs within a simulated venule inside the mouth. The 2D image reconstructed from the time-domain US is employed to simulate the reflected and scattered pressure field from the RBCs, which is then measured at the location of the receivers. The RBCs are tagged one time by a human operator and another time, automatically, by template-based computer vision. Next, the resulting signal from the RBCs is subtracted from the measured total signal in frequency domain to generate the scattered-field data, coming from the CTC alone. Feeding that signal and the background pressure field into a norm-one-based compressive sensing code enables detecting the CTC at various locations. As errors could arise in determining the location of the RBCs and their acoustic properties in the real world, small errors (up to 10% in the former and 5% in the latter) are purposefully introduced to the model, to which the proposed method is shown to be resilient. Localization errors are smaller than 12 μ when a human tags the RBCs and smaller than 25 μ when computer vision is applied. Despite its limitations, this study, for the first time, reports the results of combining two US modalities aimed at cell detection and introduces a unique and useful application for ultrahigh-frequency US imaging. It should be noted that this method can be used in detecting weak scatterers with ultrasound waves in other applications as well.

Background: The advent of Inertial measurement unit (IMU) sensors has significantly extended the application domain of Human Activity Recognition (HAR) systems to healthcare, tele-rehabilitation & daily life monitoring. IMU’s are categorized as body-worn sensors and therefore their output signals and the HAR performance naturally depends on their exact location on the body segments. Objectives: This research aims to introduce a methodology to investigate the effects of misplacing the sensors on the performance of the HAR systems. Methods: The properly placed sensors and their misplaced variations were modeled on a human body kinematic model. The model was then actuated using measured motions from human subjects. The model was then used to run a sensitivity analysis. Results: The results indicated that the transverse misplacement of the sensors on the left arm and right thigh and the rotation of the left thigh sensor significantly decrease the rate of activity recognition. It was also shown that the longitudinal displacements of the sensors (along the body segments) have minor impacts on the HAR performance. A Monte Carlo simulation indicated that if the sensitive sensors are mounted with extra care, the performance can be maintained at a higher than 95% level. Conclusions: Accurate mounting of the IMU’s on the body impacts the performance of the HAR. Particularly, the transverse position and rotation of the IMU’s are more sensitive. The users of such systems need to be informed about the more sensitive sensors and directions to maintain an acceptable performance for the HAR.

Human activity recognition (HAR) systems are used to monitor Parkinson's disease (PD) patients' mobility progress. Machine learning methods are commonly used for the development of the HAR. These methods, however, require large amount of data collected from the human subjects. Data augmentation is an affordable alternative for facilitating the development of such systems by producing similar data from actual data collected from the human subjects. In this work, three methods of data augmentation: time warping, amplitude warping and linear combination were carried out on acceleration and gyroscope signals of 6 Inertial Measurement Unit (IMU) sensors. Data was collected from a subject performing 21 therapeutic activities. To evaluate and compare the methods, a 3D human body model was utilized to visualize the motions. Then, a group of 18 individuals familiar with human motion simulation, graded the activities produced by the augmented data. The results showed that overall, time warping is the best in maintaining the structure of the activity while creating minor variations. However, it almost fails in walking activities which are more dynamic. Additionally, amplitude warping was seen to be a better choice for walking down the stairs, passing obstacle with left foot, as well as standing posture activities.