Objective gait analysis using wearable sensors is critical for managing neurological and orthopedic conditions. However, optimizing sensor placement for specific clinical tasks and mitigating the impact of hidden dataset biases on artificial intelligence (AI) models remain significant challenges. A multi-stream, attention-based deep learning model was applied to the Voisard et al. (2025) multi-cohort gait dataset, which includes data from 260 participants. The model was trained and evaluated on four distinct binary classification tasks: Parkinson’s Disease (PD) screening, Osteoarthritis (OA) screening, post-stroke asymmetry detection, and a differential diagnosis between PD and Cerebrovascular Accident (CVA). The model’s outputs included both classification performance and a quantitative map of learned sensor importance weights for four body locations: Head, Lower Back, Left Foot, and Right Foot. The model achieved robust performance in Asymmetry Detection (AUC 0.96) and PD Screening (AUC 0.83). Crucially, for OA Screening, rigorous normalization corrected an artifactually high performance (AUC 1.00 → 0.76), exposing a dataset confounder. This framework provides a data-driven method for optimizing sensor protocols and functions as an automated data auditor. We emphasize that the proposed minimal sensor sets are hypotheses requiring prospective external validation.
Hamidreza Sadeghsalehi (Sun,) studied this question.