Artificial intelligence predictions of knee kinematics, kinetics, and internal biomechanics during walking in people with knee osteoarthritis: A systematic review and meta-analysis.

This review synthesizes current literature on the use of artificial intelligence (AI) to predict knee biomechanics during walking in people with knee osteoarthritis (OA). Four databases were searched from inception to 22/01/2025. Risk of bias was assessed using a modified Newcastle-Ottawa Scale. Study quality was assessed using Grading of Recommendations, Assessment, Development, and Evaluations. Gait spatiotemporal parameters, knee kinematics, knee kinetics, and knee internal biomechanics calculated with both AI and physics-based methods were compared using root mean squared error (RMSE), normalized RMSE (NRMSE), mean absolute error (MAE) with standard deviation (SD), or correlation coefficients (R2), and pooled for reporting. Of 883 studies screened, 8 were included for review, and four provided appropriate data for meta-analysis. Studies ranged from very low to high quality. Limited data were available for spatiotemporal parameters, with few studies including direct physics-based comparators. AI-predicted knee flexion time-series had RMSE ranging from 8.39 ± 4.13° to 8.81 ± 4.25° across the gait cycle. Meta-analysis found AI-predicted peak knee adduction moment was highly correlated with its physics-based counterpart (R2: 0.86 and 0.60) with moderate errors (MAE: 0.37 and 0.45) and mean differences 0.03%BW*Ht [95% CI: -0.08 to 0.14]). AI-predicted peak knee contact forces (medial, lateral, and total) had correlations ranging from R2 = 0.17 to 0.92, and NRMSE varied between 0.21 (0.01) and 0.70 (0.05) relative to physics-based values. Overall, AI approaches have potential to predict specific knee biomechanics, but refinement and validation are needed to improve prediction accuracy across all knee biomechanical variables.