Prediction of knee adduction moment using foot acceleration: Toward personalized gait retraining for knee osteoarthritis management

Knee osteoarthritis (KOA) progression is strongly influenced by elevated knee adduction moments (KAM) during gait. While gait modification strategies have shown promise in reducing joint loading, real-time personalized feedback remains challenging. This proof-of-concept study proposes a novel machine learning framework for estimating KAM using only tri-axial foot acceleration data, enabling adaptive gait retraining outside laboratory settings. An LSTM-based neural network was trained on continuous treadmill walking data from 10 healthy adults across 33 gait conditions involving variations in step length, width, and speed. Unlike prior approaches that apply generic targets, our method compares predicted KAM against subject-specific baselines to tailor feedback. To ensure robust evaluation and avoid temporal leakage, performance was assessed using Leave-One-Subject-Out (LOSO) and Leave-One-Condition-Out (LOCO) validation schemes. Knee adduction moments were predicted using the model with MAE of 0.101–0.112 Nm/kg and RMSE of 0.151–0.168 Nm/kg under Leave-One-Subject-Out (LOSO) validation, and MAE of 0.092–0.105 Nm/kg and RMSE of 0.144–0.166 Nm/kg under Leave-One-Condition-Out (LOCO) validation. Corresponding waveform agreement metrics were r = 0.68–0.70 and CCC = 0.61–0.65. Within-trial evaluation produced slightly lower MAE and RMSE (0.090–0.096 Nm/kg, 0.139–0.147 Nm/kg), indicating the model captures both inter-subject and within-trial KAM variations. In addition to prediction accuracy, the model was evaluated for its ability to detect increases or decreases in knee adduction moment peaks. The predicted KAM peaks closely followed the trends of the real data across all gait modification trials, with slightly lower magnitudes observed in the predictions. Overall, the model effectively captures inter-condition variations, indicating its potential for tracking biomechanical changes during gait modification. However, the current framework was developed using controlled treadmill data from healthy participants and therefore r epresents a feasibility demonstration rather than a fully validated clinical solution . Future work will focus on validating the approach in clinical populations to ensure generalizability and practical applicability for personalized gait retraining.