Validity and robustness of an AI-based tool for the assessment of shoulder joint range of motion in patients with shoulder pathologies

Abstract

Aim

Shoulder dysfunction commonly limits daily activities and independence. Evaluation of shoulder range of motion (ROM) is essential for diagnosis, treatment planning, and rehabilitation. This study evaluated the validity and robustness of Maia, an artificial intelligence-based video analysis tool, compared with optical motion capture (OMC; gold standard) and visual estimation by a shoulder surgeon (standard of care, SOC).

Methods

Thirty-six patients with unilateral shoulder conditions (rotator cuff lesions, instability or ligament rupture) performed shoulder movements. ROM was measured simultaneously using Maia, OMC and SOC. Criterion validity was assessed by comparing Maia and SOC to OMC. Robustness to small camera misalignment was evaluated by testing Maia under optimal and 10°-offset camera positions. Agreement was assessed using Bland–Altman limits of agreement (LOA) and linear mixed models.

Results

Maia showed small mean differences relative to OMC and no significant measurement differences. In contrast, SOC overestimated ROM for flexion and internal rotation with the shoulder abducted to 90° (high internal rotation). Bland–Altman LOA for Maia were approximately ±20° for most movements and exceeded ±30° for internal rotation, indicating substantial variability at the individual level despite small average differences between methods. Maia's performance remained stable under small camera misalignment.

Conclusion

Findings suggest Maia provides clinically useful shoulder ROM estimates under controlled conditions and may be suitable for longitudinal monitoring or telehealth applications, with limitations for internal rotation.