Deep Learning Models to Predict Hip Joint Kinetics From Kinematics During Gait

Hip joint reaction forces serve as crucial indicators of underlying hip disorders such as osteoarthritis, avascular necrosis, and femoro-acetabular impingements. Quantifying hip kinetics traditionally relies on telemeterized implants or non-invasive conventional gait labs, which measure kinematics and external kinetics in controlled laboratory settings. These methods, although non-invasive, require specialized expertise, particularly in the workflows of musculoskeletal modeling. This study aims to develop a sequence-to-sequence (seq2seq) model capable of predicting hip joint reaction forces in three planes using positional data obtained from either marker-based or marker-less motion capture systems. The model was trained and validated using a publicly available database of eighty healthy subjects, employing leave-one-subject-out cross-validation. Through optimization of hyperparameters and early stopping techniques, the model achieved RMSE values below 0.5% BW for antero-posterior shear force, 1% BW for compressive forces, and 0.31% BW for medio-lateral shear forces. Normalized RMSE values were less than 0.25% across all components of hip joint reaction forces. In conclusion, this study proposes a deep learning seq2seq model as a viable method for accurately predicting hip joint reaction forces in all three planes. These seq2seq models present a cost-effective alternative to traditional and cumbersome gait lab equipment and current musculoskeletal modeling approaches used in hip kinetics assessment.