Comparison of Markerless and Conventional Marker-Based Shoulder Kinematics Models During Activities of Daily Living in Patients With Glenohumeral Osteoarthritis.

Background: Markerless motion capture utilizes deep learning models to evaluate standard video from multiple cameras and is significantly more time-efficient than traditional marker-based systems in both setup and analysis. There has been increasing interest in validating markerless motion analysis in the clinical orthopaedic patient population.

Purpose: To evaluate the concurrent validity of markerless shoulder analysis compared to traditional marker-based shoulder analysis during activities of daily living (ADLs) in patients with glenohumeral osteoarthritis. We hypothesize that the markerless system will accurately and reliably capture shoulder kinematics in patients with glenohumeral osteoarthritis compared to a marker-based system.

Methods: One hundred subjects, eighty-five patients with glenohumeral osteoarthritis scheduled for shoulder arthroplasty and 15 healthy controls were enrolled in this study. Each patient underwent clinical upper extremity assessment with data being captured concurrently by a traditional marker-based motion capture system and a commercially available markerless system. This study assessed ADLs including four tasks: overhead reaching, drinking, hair brushing, and personal hygiene tasks. Marker-based motion was evaluated with University of Southampton Upper Limb Kinematic Model flexion-based (SF1, SF2) and abduction based (SA1, SA2) models. For each combination of task and laterality, the consistency in response between the markerless system with the SF1, SF2, SA1 and SA2 variations of the marker-based system were investigated by determining the interclass correlation coefficient of the peak angle and range of motion in the three planes of motion: flexion/extension, abduction/adduction, and internal rotation.

Results: There was a strong positive relationship between markerless and SF1 and SF2 marker-based models in peak angle (ICC: 0.81-0.95; p-value < 0.001), range of motion (ICC: 0.81-0.97; p-value < 0.001), and shoulder motion pattern (ICC: 0.88-0.99; p-value < 0.001) in flexion/extension and abduction/adduction throughout all tasks. There was a weaker positive relationship between markerless and SA1 and SA2 marker-based models in flexion/extension and abduction/adduction throughout all tasks (ICC: 0.35-0.97; p-value < 0.001). As forward flexion and abduction angles approached the maximum functional range of the shoulder, there was a weaker but consistent relationship between the two systems.

Conclusion: Markerless motion analysis of the shoulder joint is accurate and has the potential to expand the utility of motion analysis in the upper extremity. Markerless systems were within 10 degrees of both the marker-based and markerless models for flexion/extension; however, it underestimated rotation movement across all tasks.

Clinical Significance: Because markerless motion analysis is cheaper, faster, and easier to implement, it can greatly increase the availability of motion analysis within laboratories and clinical practice and has the potential to become a core component of clinical management of shoulder pathologies.