Abstract
People with stroke (PwS) face increased fall risk on uneven surfaces; however, gait stability under such conditions remains unexplored. This study used machine learning (ML) to identify acceleration features distinguishing PwS from healthy controls (HC) during uneven-surface walking and to predict them from even-surface gait parameters. Trunk acceleration data from 71 PwS and 39 HC were analyzed using classification and regression models. The ML classifiers achieved an accuracy of over 95%. The key discriminative features included the vertical root mean square (RMS_VT), anterior-posterior sample entropy (SampEn_AP), and harmonic ratio (HR_AP). In PwS, even-surface gait speed < 0.8 m/s predicted reduced speed and higher RMS_VT on uneven surfaces. SampEn_AP and HR_AP were influenced by ankle kinematics and their even-surface values, respectively, showing nonlinear associations. These findings support the use of wearable sensor data and interpretable ML to assess gait stability and adaptability, facilitating development of digital biomarkers for personalized stroke rehabilitation aimed at improving outdoor mobility.