Modeling the dynamics of the recovery process in robot therapy

Abstract

The mechanisms of action of physical assistance in promoting motor recovery after stroke are poorly understood. To explicitly model this process might help understanding what determines recovery, and how to make it faster and more effective. Linear dynamical models are used to describe the dynamics of sensorimotor adaptation, and could be extended to characterize the process of recovery of motor functions in impaired subjects while they move with the assistance of a therapist, or a robot. To test the feasibility of this approach, here we focus on a robot therapy experiment which involves a hitting task. Nine chronic stroke survivors underwent 8 to 10 rehabilitation sessions. We used a linear dynamical model to describe the trial-by-trial dynamics of the recovery process, with robot-generated assistance as input and subject's motor performance as output. In all subjects, the model correctly reproduced the overall evolution of performance over sessions. A comparison of the estimated model parameters with the clinical scales (Fugl-Meyer arm portion and Ashworth) and their modifications indicated that the time constant of the recovery process is predictive of the retention of the recovery (assessed after three months from completion of the protocol). Moreover, we found that in subjects with little or no spasticity, recovery is mediated by motor error. In contrast, in subjects with high spasticity, recovery is more influenced by performance. Although preliminary, these results suggest that modeling the recovery process with dynamical models is feasible, and could serve as basis to devise ‘optimal’ strategies for regulating assistance with the aim of maximizing recovery1.