Background: Measurement of finger joint range of motion is key for assessment of disability and rehabilitation in numerous hand pathologies. Standard manual goniometry is widely used in clinical practice, but it is time-consuming, requires trained assessors, and may be subject to inter- and intra-observer variability. Image-based assessment, including AI-assisted estimation of joint angles, may provide a more scalable alternative, particularly for remote assessment where there can be accessibility issues. However, before clinical or research adoption, its agreement with conventional goniometry must be evaluated.
Aim: This study is aimed to assess the accuracy and reliability of AI in assessment of measurements of finger joint angles compared with goniometer-derived measurements, which are currently treated as the reference standard and is what is most commonly used for angular assessments in everyday practice. There is little evidence so far in AI’s accuracy and whether it is a reliable tool.
Methods: Thirty-seven paired finger posture assessments of a non pathological index and middle finger were analysed for a total of 111 measurements. For each assessment, goniometer and AI-derived measurements were recorded for the metacarpophalangeal joint (MCPJ), proximal interphalangeal joint (PIPJ), and distal interphalangeal joint (DIPJ). 37 low profile thermoplastic splint moulds were made to maintain a fixed angle through PIPJ and DIPJ and reduce any measurement error when transitioning from goniometer to photo. A Single hand therapist was responsible for moulding of the thermoplastic splints and of the goniometry measurements. Photos taken were uploaded to Chat GPT angle assessments made.
Error was calculated as AI measurement minus goniometer measurement, such that positive values represented overestimation by AI and negative values represented underestimation. Agreement was assessed separately for each joint and in a pooled analysis across all joints. Statistical measures included mean bias, mean absolute error (MAE), root mean square error (RMSE), Bland–Altman 95% limits of agreement, percentage of measurements within ±5° and ±10° of goniometry, and intraclass correlation coefficient (ICC).
Results: Joint-specific performance varied substantially. PIPJ measurements demonstrated the strongest agreement, 68% of AI measurements were within 10 degrees of the goniometer, 49% within 5 degrees with a mean bias −3.3°, MAE 8.6°, RMSE 11.0°, limits of agreement from −24.2° to +17.6°, and ICC 0.91, indicating good reliability. DIPJ measurements showed moderate agreement, with mean bias −4.7°, MAE 10.0°, RMSE 12.7°, limits of agreement from −28.2° to +18.8°, and ICC 0.67. MCPJ measurements performed poorly, only 35% were within 10 degrees with mean bias +11.6°, MAE 22.2°, RMSE 27.5°, very wide limits of agreement from −38.0° to +61.1°.
Conclusion: AI image-based measurements demonstrated good reliability for PIPJ angles and moderate reliability for DIPJ angles when compared with goniometry. However, MCPJ measurements showed poor agreement, likely reflecting difficulty estimating the metacarpal axis from two-dimensional photographs. Further refinement of image acquisition, landmark placement, and MCPJ axis estimation would be next steps in an attempt to improve accuracy for more reliable clinical application.