Validation of markerless human pose estimation methods for clinical assessment of elbow range of motion

Abstract

Clinicians rely on the assessment of joint range of motion to determine the necessity of clinical interventions and to track rehabilitation progress. They often perform visual assessments of these joint angles due to limited time for longer but more accurate methods. Markerless monocular human pose estimation methods could help reduce the workload of physicians in clinical practice, allow patients to conduct regular assessments remotely, and be used to provide reliable assessments in the absence of expert raters, but they have not been validated extensively on clinical populations. We measured the elbow flexion and extension ranges of motion of 46 patients with shoulder and elbow pathologies. These values were assessed by RTMW and HSMR, two different monocular pose estimation methods, and an experienced physician who performed visual assessments. A state-of-the-art passive-marker optical motion capture system was used as the reference system. For elbow flexion, HSMR outperformed visual assessment. HSMR had a higher concordance correlation coefficient of 0.92 and a lower minimal detectable change of 6.74° as determined by a concordance analysis and a linear mixed-effects model. Conversely, for elbow extension, visual assessment outperformed both markerless methods. RTMW was the best-performing markerless method with a concordance correlation coefficient of 0.83 and a minimal detectable change of 7.93°. Thus, HSMR represented a viable alternative to visual assessment for measuring elbow flexion range of motion in a clinical population, but neither RTMW nor HSMR offered an improvement over visual assessment for measuring elbow extension range of motion.