A Wearable IMU-Based Gait Assessment Framework Utilizing Sequential Multivariate Dynamic Time Warping for Clinical Rehabilitation

Abstract

Impairment resulting from stroke significantly affects patient mobility, making recovering the ability to walk a primary rehabilitation goal. While wearable inertial measurement units (IMUs) offer a cost-effective alternative to subjective clinical observations, conventional gait analysis often struggles with the high pacing inconsistency and severe signal noise typical of poststroke hemiplegic gaits. Building on an established IMU coordinate calibration method from prior research, this study proposes a novel framework that utilizes sequential multivariate dynamic time warping (SMDTW) coupled with multiple linear regression (MLR) for objective gait assessment. Unlike earlier approaches that relied on discrete spatio-temporal parameters and rigid peak detection, the proposed SMDTW algorithm continuously aligns 3-D kinematic trajectories (pitch, roll, and ground clearance). This sequential alignment effectively filters out compensatory movements and nonlinear temporal distortions, yielding a robust average step to extract continuous bilateral symmetry features. Evaluated on a clinical cohort of 26 poststroke patients, the extracted features were utilized in a strictly constrained MLR model. To ensure robustness and prevent overfitting, a leave-one-out cross-validation (LOOCV) protocol was employed. The full-cohort training model achieved a high variance explanation (<inline-formula> <tex-math notation="LaTeX">$\text {R}^{{2}}={0}.{8722}\text {)}$ </tex-math></inline-formula>. More importantly, under the rigorous LOOCV condition, the framework maintained strong predictive generalizability (<inline-formula> <tex-math notation="LaTeX">$\text {R}_{\text {cv}}^{{2}}={0}.{8244}\text {)}$ </tex-math></inline-formula> and strong rank-order agreement (Spearman’s <inline-formula> <tex-math notation="LaTeX">$\rho _{\text {cv}}={0}.{9090}\text {)}$ </tex-math></inline-formula>. In addition, the out-of-sample predictions demonstrated a marked enhancement in linear proportionality, achieving a LOOCV Pearson correlation coefficient of 0.9081, outperforming the baseline literature (<inline-formula> <tex-math notation="LaTeX">$\text {r} = \mathbf {}{0}.{752}\text {)}$ </tex-math></inline-formula>. This indicates that the SMDTW-derived features reliably and linearly mirror clinical severity, validating the framework’s potential for continuous rehabilitation tracking.