level: research
researchers created a deep learning benchmark called gait2hip-60 to predict hip muscle forces and joint moments from lower-limb joint angles during walking. they collected gait data from 60 healthy adults walking at three different speeds set by a metronome. ten joint angles from both legs served as inputs, while reference outputs came from opensim musculoskeletal simulations. three sequence models—lstm, transformer, and mamba—were trained and compared using the same data splits, preprocessing, and evaluation metrics.
the best-performing model was then tested on an external group of 9 patients with osteonecrosis of the femoral head, a condition that affects hip joint health. this test was done without any retraining, showing how well the model generalizes to clinical data. the study provides a standardized protocol for future comparisons, aiming to make hip force estimation faster and more accessible than traditional simulation methods.
by using only kinematic data from motion capture, the approach could simplify biomechanical analysis in clinics where full musculoskeletal modeling is impractical. the benchmark includes all code and data, encouraging reproducible research in applied machine learning for human movement. the comparison of lstm, transformer, and mamba architectures offers insights into which sequence models best capture the temporal patterns of gait.
why it matters: this method could speed up clinical gait analysis by replacing slow simulations with fast neural network predictions, aiding treatment planning for hip disorders.