WEARABLE SENSOR MONITORING OF WALKING ON DIFFERENT SURFACES AS A DIGITAL OUTCOME: DEEP LEARNING MODEL PERFORMANCE WITH SENSOR AND CLASS REDUCTION

The remote monitoring of walking behavior, such as step counts, using wearables is increasingly common in research and clinical settings. However, contextless step counts have limited value for assessing orthopedic patients recovering from trauma or surgery. Instead, step analysis on specific surfaces, like stairs or slopes, is more relevant for evaluating rehabilitation progress and tailoring therapeutic decision making. Algorithms that classify walking surfaces from inertial measurement unit (IMU) signals have been developed, but the lack of standardized and practical methods for IMU data collection and analysis has hindered the creation of robust, generalizable models. This study investigates whether simplifying IMU-based gait monitoring through deep learning (DL) models—via sensor reduction (fewer sensors/signals) or surface class grouping—can maintain or improve classification performance. Data were sourced from an open-science multi-modal Gait Database (Losing et. al, 2022), comprising 20 subjects (5 female, 15 male; 18–69 years) walking on different urban surfaces wearing lycra suits embedded with 17 IMU sensors. Our baseline DL model was limited to lower-body sensors (pelvis, thighs, shanks & feet) which we deemed a feasible, but still encompassing setup. The original surface classes (n=5) included flat walking, stairs up/down, and slopes up/down. Using a previously validated Bi-Directional CNN-LSTM model with batch normalization (Vinco et. al, 2024), we tested variations which included: (a) acceleration signals only, (b) a single pelvis sensor, and (c) surface class reduction to two groups: i) by type (flat, slope, stairs) and ii) by elevation (flat, up, down). Simplifications aimed to maximize usability by enhancing patient compliance (single pelvis sensor) or battery life (acceleration-only signals). Using the baseline sensor arrangement and both the acceleration and gyroscope signals, the model achieved 84.01% accuracy (F1: 0.77–0.98). Using only acceleration, lower-body sensors reached 65.75% accuracy (F1: 0.53–0.96), and the pelvis sensor alone scored 67.08% (F1: 0.63–0.75). With surface type grouping, lower-body sensors achieved 86.86% accuracy (F1: 0.74–0.94); pelvis sensor, 65.80%. Grouping by elevation improved lower-body sensor accuracy to 89.69% (F1: 0.83–0.92) and pelvis sensor to 85.27% (F1: 0.69–0.94). High classification accuracy (>85%) was achieved, even with a single pelvis sensor and grouped classes. While sensor reduction decreased performance, grouping by elevation produced comparable results. These findings suggest practical, patient-compliant solutions for gait monitoring with potential clinical applications. Including gyroscope data appears effective towards model performance especially when many surface types need to be classified.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.