Developing Dynamical Models to Characterize Stroke Gait Impairments

Understanding the dynamics that generate human gait is essential for designing tailored rehabilitative therapies for stroke survivors. Most purely biomechanical models, however, are highly sensitive to parameters, and individuals often express a large degree of cycle-to-cycle variability, preventing robust and accurate inference of the underlying dynamical system that generates an individual’s gait. Here, we present a Recurrent Neural Network (RNN)-based model that produces a robust kinematic gait signature for visualizing and comparing gait kinematics between able-bodied individuals and stroke-survivors. Extracting the model’s internal activations, we identify unique gait signatures for each individual. These signatures reliably distinguish stroke from unaffected individuals, as well as different gait types from each other. Moreover, these metrics are speed invariant - which can be potentially useful for determining the best gait rehabilitation for stroke-survivors without evaluating them at some optimal speed. Building from these dynamical characterizations, it may be possible to build generative models of gait function, allowing us to generate individual-specific rehabilitative therapies.