Multimodal machine learning mobility assessment in Parkinson's disease within supervised and unsupervised settings.

backgroundMobility in people with Parkinson’s disease (PwPD) is strongly context-dependent. Clinic-based assessments provide only brief snapshots under controlled conditions and may not reflect everyday functional capacity. Wearable sensors enable continuous free-living monitoring, yet most studies use unimodal signals and simple threshold-based walking detection. Multimodal sensing, combining surface electromyography (sEMG) with inertial measurement units (IMUs), can offer a more complete picture of mobility in Parkinson’s disease (PD). Therefore, developing a data-driven multimodal gait analysis pipeline is essential for accurately identifying walking episodes and deriving detailed gait characteristics of PD in unsupervised free-living settings.

methodsTen PwPD wore four thigh and shank sensors (IMU + sEMG) during supervised laboratory and outdoor walks, followed by unsupervised free-living recording on the same day. Video captured using eye-tracking glasses was used to annotate the supervised data, which were then used to train a machine learning (ML) model based on IMU-only and fused IMU + sEMG features. The trained model was applied to the unsupervised data to detect walking bouts, after which spatiotemporal, kinematic, and muscle-activation measures were computed.

resultsWalking-bout detection using IMU features alone achieved acceptable performance (F1 = 87.53%), with higher performance observed when sEMG features were included (F1 = 94.45%). Exploratory analyses suggested that unsupervised walking may be associated with slower pace, greater variability, and increased asymmetry compared with supervised walking. A reduction in knee flexion-extension (F-E) amplitude and an earlier occurrence of peak F-E were also observed during unsupervised walking. Additionally, higher lower-limb muscle activation during stance was noted, indicating potentially increased neuromuscular demand in free-living conditions. These findings should be interpreted cautiously and are intended to be hypothesis-generating rather than definitive.

conclusionThis pilot study successfully presents a multimodal, data-driven framework integrating neuromuscular and kinematic signals to assess mobility in supervised and unsupervised settings. The analysis of combined changes in muscle activation, joint kinematics, and spatiotemporal gait measures reveals interconnected gait adaptations that are otherwise missed by unimodal systems. These findings suggest the potential for novel. multimodal digital biomarkers that offer an ecologically valid, continuous monitoring of motor function in PwPD.