A neural network for predicting knee contact forces from clinic-friendly data

Physics-based neuromusculoskeletal models have been widely used to assess biomechanics not readily measured in laboratory or clinical settings (e.g., joint contact forces). However, these physics-based models require extensive and rich input data, as well as expertise to operate. Artificial intelligence (AI) is a promising alternative to physics-based modelling. AI can learn complex input-output relationships directly from laboratory biomechanical data, thus enabling scalable estimation of internal biomechanics (e.g., joint contact loading) using clinic-friendly data (e.g., video, body measures, etc). In this study, we developed a bi-directional long short-term memory (BiLSTM) model to estimate lower limb joint kinematics and moments, as well as knee contact force using clinic-friendly data (e.g., height, weight, sex, sparse positions of anatomical key points, and surface electromyography (EMG)). We also explored the influence of the number and specific combinations of EMG on the BiLSTM prediction accuracy. The BiLSTM accurately predicted hip, knee, and ankle kinematics (coefficient of determination (R²) = 0.94 ± 0.04; root mean squared error (RMSE) = 3.60 ± 0.92°) and moments (R² = 0.91 ± 0.03; RMSE = 0.15 ± 0.02 Nm/kg), as well as knee contact force (R² = 0.81 ± 0.11; RMSE = 0.39 ± 0.17 bodyweights (BW)) using an optimal EMG configuration (n = 3). Across all possible EMG numbers and combinations (n = 255), four-quadrant and pareto analyses confirmed vastus medialis, gastrocnemius medialis, and gastrocnemius lateralis muscles consistently contributed to accurate knee contact force prediction. This 3-EMG configuration offered the best trade-off between BiLSTM accuracy and practicality when predicting knee contact force. These findings suggest minimal-EMG AI models could support future adoption of AI-based biomechanical analysis in clinical settings, by reducing cost and complexity without sacrificing accuracy.