Continuous video-based postural tracking for Parkinson's disease: a pose estimation approach with landmark...

[Purpose] We aimed to develop and validate an automated, continuous, video-based postural tracking system using pose estimation in people with Parkinson's disease. [Participants and Methods] This cross-sectional validation study included 14 people with Parkinson's disease with postural abnormalities. Participants performed a 2 min standing endurance test while being recorded using iPads. We developed Continuous Video AutoPosture for Parkinson's disease, an automated system with optional operator-guided initial-frame landmark calibration integrating MediaPipe pose estimation and DeepLab semantic segmentation to track postural angles at 1 s intervals. We measured the lumbar anterior trunk flexion, thoracic anterior trunk flexion, and lateral trunk flexion and manually calculated ground-truth postural angles from 140 sampled frames using NeuroPostureApp for validation. An optional initial-frame landmark calibration feature was implemented. [Results] Following calibration, the system demonstrated excellent reliability for lumbar (intraclass correlation coefficient=0.98) and thoracic (intraclass correlation coefficient=0.96) anterior trunk flexion and good reliability for lateral trunk flexion (intraclass correlation coefficient=0.80). Continuous tracking captured gradual postural deterioration during the standing task. [Conclusion] The system demonstrated excellent-to-good reliability and enables continuous postural tracking with superior temporal resolution compared with conventional assessments, advancing clinical evaluation and research on postural abnormalities in people with Parkinson's disease.