Prediction of knee loads during activities of daily living using custom instrumented insoles and machine learning.

Medial joint contact force, knee adduction moment, and knee flexion moment are key mechanical risk factors of knee osteoarthritis, yet calculating these knee loads traditionally requires time-intensive procedures with a gait laboratory. Integrating wearable technologies, such as instrumented insoles, with neural networks offers a promising avenue for rapid and accessible estimation of these variables in clinical environments. This study aimed to estimate knee loads during walking, running, sit-to-stand, and stand-to-sit activities using a recurrent neural network model trained on wearable sensor data. We hypothesized that the model would predict the key mechanical risk factors with strong correlation coefficients and low mean errors. Forty-seven participants performed all activities while instrumented insole and motion capture data were recorded. Reference knee adduction and flexion moments were calculated with inverse dynamics, and medial joint contact force was calculated with a reduction modeling approach. Activity-specific recurrent neural network models were trained using insole sensor data to predict knee loading variables. The models demonstrated strong predictive performance across all activities (0.88 < r < 0.98), except for sit-to-stand (r = 0.50) and stand-to-sit (r = 0.48) knee adduction moments, which showed moderate correlation coefficients. Mean absolute errors were moderate to low across variables. These findings highlight the feasibility of using a single instrumented insole combined with neural networks to estimate knee loads associated with osteoarthritis risk. This approach has the potential to enhance the accessibility of biomechanical assessments and support real-time biofeedback applications in both research and clinical settings.