Injury, hospitalization, and even death are common consequences of falling for elderly people. Therefore, early and robust identification of people at risk of recurrent falling is crucial from a preventive point of view. This study aims to evaluate the effectiveness of an interpretable semi-supervised approach in identifying individuals at risk of falls by using the data provided by ankle-mounted IMU sensors. Our method benefits from the cause–effect link between a fall event and balance ability to pinpoint the moments with the highest fall probability. This framework also has the advantage of training on unlabeled data, and one can exploit its interpretation capacities to detect the target while only using patient metadata, especially those in relation to balance characteristics. This study shows that a visual-based self-attention model is able to infer the relationship between a fall event and loss of balance by attributing high values of weight to moments where the vertical acceleration component of the IMU sensors exceeds 5 m/s² during an especially short period. This semi-supervised approach uses interpretable features to highlight the moments of the recording that may explain the score of balance, thus revealing the moments with the highest risk of falling. Our model allows for the detection of 71% of the possible falling risk events in a window of 1 s (500 ms before and after the target) when compared with threshold-based approaches. This type of framework plays a paramount role in reducing the costs of annotation in the case of fall prevention when using wearable devices. Overall, this adaptive tool can provide valuable data to healthcare professionals, and it can assist them in enhancing fall prevention efforts on a larger scale with lower costs.

Identifying the foot stance and foot swing phases, also known as foot-to-ground (FTG) detection, is a branch of Human Activity Recognition (HAR). Our study aims to detect two main phases of the gait (i.e., foot-off and foot-contact) corresponding to the moments when each foot is in contact with the ground or not. This will allow the medical professionals to characterize and identify the different phases of the human gait and their respective patterns. This detection process is paramount for extracting gait features (e.g., step width, stride width, gait speed, cadence, etc.) used by medical experts to highlight gait anomalies, stance issues, or any other walking irregularities. It will be used to assist health practitioners with patient monitoring, in addition to developing a full pipeline for FTG detection that would help compute gait indicators. In this paper, a comparison of different training configurations, including model architectures, data formatting, and pre-processing, was conducted to select the parameters leading to the highest detection accuracy. This binary classification provides a label for each timestamp informing whether the foot is in contact with the ground or not. Models such as CNN, LSTM, and ConvLSTM were the best fits for this study. Yet, we did not exclude DNNs and Machine Learning models, such as Random Forest and XGBoost from our work in order to have a wide range of possible comparisons. As a result of our experiments, which included 27 senior participants who had a stroke in the past wearing IMU sensors on their ankles, the ConvLSTM model achieved a high accuracy of 97.01% for raw windowed data with a size of 3 frames per window, and each window was formatted to have two superimposed channels (accelerometer and gyroscope channels). The model was trained to have the best detection without any knowledge of the participants’ personal information including age, gender, health condition, the type of activity, or the used foot. In other words, the model’s input data only originated from IMU sensors. Overall, in terms of FTG detection, the combination of the ConvLSTM model and the data representation had an important impact in outperforming other start-of-the-art configurations; in addition, the compromise between the model’s complexity and its accuracy is a major asset for deploying this model and developing real-time solutions.

This study addresses the critical need for effective rehabilitation monitoring for individuals with gait issues through the integration of interpretable artificial intelligence (AI) models. The study meticulously details the methodology, emphasizing the systematic data collection process, participant categorization based on gait severity, and the deployment of a novel physical activity monitoring device. Through the lens of the Berg balance test, the analysis scrutinizes functional mobility and balance across seven rehabilitation activities. Data augmentation techniques and data preprocessing methods are employed, and the proposed hybrid model combines both time and frequency-domain data using a multibranch CNN-BiLSTM architecture embedded with a dual-stage attention mechanism. The study compares our approach with machine learning and deep learning models, showcasing the superior performance of the proposed interpretable architecture. Interpretability is highlighted through attention mechanisms, revealing the model’s prioritization of fast Fourier transform-transformed data and its ability to identify influential branches. Our model provides transparent insights to healthcare professionals for personalized rehabilitation strategies. These results show high potential in developing real-time applications, including the integration of additional sensor modalities, interdisciplinary collaborations, and the expansion of demographic inclusivity for more robust and applicable rehabilitation monitoring solutions.

To explore the feasibility of enhancing independence among the aging population through wearable sensing technologies, this preliminary study presents and evaluates a novel system for fine-grained human activity recognition in smart home environments. Conducted at the Université du Québec à Chicoutimi with volunteer students and personnel aged between 20 and 45 years, the study investigated the use of photoplethysmography (PPG) and accelerometer sensors to identify six cooking gestures. The proposed model employed Spatial Dilated Convolution (SDC) encoders to capture local temporal patterns, achieving up to 95% accuracy in gesture detection. By comparing the attention weights of a Multivariate Multi-Head Attention (MMA) mechanism with the gate strengths of a Focal Modulation (FM) mechanism, we further assessed their interpretability in identifying hand-specific contributions without explicit annotations, obtaining an 83% accuracy in dominant hand identification. Validation through a Leave-One-Group-Out evaluation across participants resulted in mean accuracy scores of 82% for gesture detection and 75% for dominant hand identification. These findings demonstrate the potential of repurposing PPG sensors to detect subtle, movement-induced signals as a proof of concept for a non-invasive and privacy-conscious monitoring approach. While preliminary, this work lays the groundwork for future studies targeting real-time, personalized assistance systems to promote elderly well-being in smart home contexts.

Rehabilitation training plays a vital role in helping individuals regain motor function after injury or neurological conditions. This article presents two novel multimodal deep learning models, CLAP-Net and its interpretable extension FoCLAP-Net, for the detection and classification of abnormal postures during rehabilitation. These models are evaluated on the MovePort dataset, which contains synchronized signals from surface electromyography (EMG) sensors, inertial measurement units (IMUs), insole pressure sensors (IPSs), and an optical motion capture system (MoCap). The dataset includes 25 sessions recorded from 25 participants, covering a range of postures such as normal standing, forward leaning, backward leaning, and half squats. CLAP-Net employs convolutional and recurrent layers combined with an attention mechanism to capture relevant spatiotemporal features. FoCLAP-Net further incorporates a focal modulation (FM) mechanism to enhance model interpretability by identifying key sensor modalities influencing classification decisions. FoCLAP-Net achieves a classification accuracy of 98.66% and supports real-time detection, enabling dynamic adaptation of rehabilitation strategies such as functional or epidural electrical stimulation (FES/EES) and gravity-assist systems. The insights offered by FoCLAP-Net enhance clinical trust by making model outputs more transparent and understandable to practitioners. These findings highlight the promise of multimodal deep learning in building intelligent, personalized rehabilitation platforms capable of optimizing movement therapy based on data-driven posture analysis.