Gait Through a Lens: Contextual Fall Risk Assessment in Parkinson’s

Introduction
Free-living gait assessment for fall risk in Parkinson’s disease (PD) often uses inertial measurement units (IMUs) to extract gait characteristics. However, IMU-only approaches lack environmental and behavioural context, which can lead to inflated estimates of fall risk when gait changes are driven by extrinsic factors e.g., obstacles. We combine IMUs with wearable eye-tracking video glasses and AI-based computer vision (CV) to enrich free-living gait assessment with environmental and gaze information, enabling a robust interpretation of fall risk.

Methods
Seven individuals with PD completed a 3-hour free-living assessment while wearing an IMU (100 Hz, ±8 g) and eye-tracking glasses. IMU signals were analysed using validated algorithms to extract gait characteristics (step time variability). Video recordings were processed using a YOLOv8 model trained on ≈3k annotated home-environment images. The model detected environmental features and obstacles, anonymised sensitive content, and supported inference of gaze behaviour during walking. Temporal alignment of gait metrics with video-derived context allowed differentiation between intrinsic gait instability and adaptive responses to environmental demands.

Results
IMU-only analysis identified multiple gait bouts with elevated step time variability (e.g., 0.20 s versus 0.01 s), suggesting increased fall risk. When contextualised with CV outputs, many of these bouts reflected appropriate adaptations to environmental challenges such as stepping around clutter. Gaze data further supported this interpretation, showing attention directed toward hazards or obstacles. In a preliminary analysis of 10 gait bouts, four potential high-risk events were identified using IMU data alone; three were subsequently reclassified as low risk once environmental and gaze context was considered, reducing false positives.

Conclusion(s)
Combining IMUs with wearable eye tracking and AI-driven CV improves interpretability of free-living gait by accounting for environmental adaptations. This multimodal approach enhances fall risk accuracy, supports privacy, and may enable personalised interventions. Ongoing work focuses on larger PD cohorts to generalise this framework.