Evaluating the Clinical Validity of Single-Camera Pose Estimation for Joint Kinematics Across Functional Tasks

Motion analysis is critical for diagnosing and managing musculoskeletal disorders, but conventional multi-camera systems are expensive, limited to controlled environments, and require technical experience to operate. Single-camera, deep-learning-based pose estimation has emerged as a low-cost, portable alternative, yet its accuracy in clinical contexts remains underexplored. This study evaluated the accuracy of a single-camera pose estimation model (MeTRAbs) against a validated multi-camera system (THEIA3D) during gait, sit-to-stand, and trunk flexion-extension. 51 participants completed 669 movement trials recorded simultaneously with a smartphone camera (30Hz) and THEIA3D (180Hz). System outputs were aligned using similarity transforms. Root-mean-square errors (RMSEs) assessed positional accuracy for tracked joints and angular accuracy for computed sagittal joint angles (hips, knees, trunk). Intraclass correlation coefficients (ICCs) and Bland-Altman plots characterized reliability and systematic bias. Across all tasks, mean trajectory RMSE was 5.95 cm and joint angle RMSEs ranged 2.10\u02da-10.98\u02da. Stratifying errors per movement type and anatomical plane demonstrated that proximal joints and frontal plane motions showed higher fidelity, with the greatest errors in distal, dynamic joints (e.g., ankles in gait). Despite discrepancies, ROM ICCs exceeded 0.93 for all angles with a mean ICC of 0.967, indicating strong test-retest reliability. Errors were predominantly attributable to systematic bias, not random error, and harmonic correction substantially reduced angular error, confirming their predictable nature. These findings support single-camera pose estimation as a feasible, scalable tool for clinical motion analysis. While not a full substitute for multi-camera capture, its accuracy and reliability suggest value for outpatient, telehealth, and rehabilitation contexts.