Abstract
Understanding the dynamics of human gait, including both motions and forces, is vital to promote human health and performance. Conventional gait analysis requires laboratory-based experiments and physics-based simulations to quantify gait dynamics and analyze how dynamics change with treatment, training, injury, and disease. However, the high costs associated with experiments and simulations has confined the use of gait dynamics to small-scale research studies. While deep learning models offer low-cost prediction, and can be highly expressive in fitting large-scale data, existing models have primarily been trained on small datasets with homogenous demographics and focused on predicting a single output. To overcome these limitations, we developed GaitDynamics, a generative foundation model for human gait that is trained on a large dataset with diverse participant demographics and gait patterns. GaitDynamics can be used for diverse tasks with different inputs, outputs, and clinical applications, which we illustrate in three examples: i) estimating ground reaction forces from kinematics with high accuracy and robustness even with missing kinematic data and for populations not included in the training dataset, ii) predicting the influence of gait modifications on knee loading without the need for resource-intensive experiments, and iii) predicting kinematic and force changes that occur with increasing running speeds. These representative tasks demonstrate that GaitDynamics makes accurate and rapid predictions in seconds based on flexible inputs, showing its potential to assess and optimize gait for injury prevention, disease treatment, and performance coaching. All data, code, and trained models are publicly shared.