Towards Context-Aware Clinical Motion Understanding in Daily Living at Home: Freezing of Gait Detection with Egocentric Vision
A study evaluated freezing of gait detection in Parkinson's disease using synchronized egocentric video, wearable IMUs, and expert-annotated labels from 13 participants in their homes. An IMU-based TCN achieved 42.3 F1 and 83.0 AUROC, while V-JEPA2 ego-video features achieved 32.6 F1 and 77.2 AUROC under leave-one-subject-out evaluation.
Researchers investigated context-aware clinical motion understanding for freezing of gait detection in Parkinson's disease using egocentric vision and wearable IMUs. Data from 13 participants in home settings included synchronized video, IMU signals, and expert annotations. An IMU-based temporal convolutional network outperformed frozen features from the V-JEPA2 ego-video foundation model, achieving 42.3 F1 and 83.0 AUROC versus 32.6 F1 and 77.2 AUROC. Ego-video alone showed above-chance discrimination and may capture complementary FOG-relevant information.
The IMU-based TCN trained from scratch outperformed frozen ego-video foundation model features, suggesting that task-specific temporal modeling on inertial data remains more effective for FOG event detection. Qualitative analysis indicates egocentric vision may provide complementary contextual cues not present in IMU signals, motivating future multimodal fusion approaches.
Wearable and egocentric sensing for neurological monitoring is advancing toward home-based clinical applications. The performance gap between IMU and vision models highlights current limitations of general-purpose video foundation models in specialized medical tasks, while the complementary nature of modalities suggests hybrid systems may be needed for robust deployment.
Improved FOG detection could enable continuous remote monitoring and personalized therapy for Parkinson's disease, reducing reliance on episodic clinical assessments. The research informs product development for wearable health tech and egocentric AI platforms targeting neurological care.
Next signals include publication of peer-reviewed results, release of the multimodal dataset, and follow-up work on fusing IMU and egocentric video features to improve FOG detection. Clinical validation in larger cohorts and exploration of real-time on-device inference are likely next steps.