Modelling the recovery of acute stroke survivors during adaptive robot-assisted training

Introduction: In the last years several models of neuromotor recovery after stroke have been proposed [1]. These models aims to identify the mechanisms underlying the recovery process. Recently, a computational model [2] addressed the recovery process of chronic stroke survivors during a robot-assisted training protocol, characterized by distal movements and assistance 'asneeded'. In this study, the analysis of the trial-by-trial performance highlighted three main aspects of the recovery process: learning, retention, and slacking. In the present workweinvestigate whether the same description is suitable for acute stroke survivors. Methods: Ten subjects with an acute stroke participated in 18 robot therapy sessions (45 min/session, 3 sessions/week, 6 weeks). Inclusion criteria were a single episode of stroke, Mini-Mental State Examination (MMSE) > 24, Token Test > 29, no treatment with botulinum toxin in the upper limbs in the last 4 months and no functional surgery in the last 6 months. During treatment, subjects sat in front of a computer screen and grasped the handle of a planar robot manipulandum. Subjects performed sequences of 4 point-topoint reaching movements. The points are arranged as the vertexes of a square. At the end of each movement, subjects receive a score that depends on the time employed to reach the target: subjects who move faster get a higher score. The assistance provided by the robot is adjusted automatically through an adaptive regulator [3], which sets the level of assistance separately for each direction of movement (i.e. for each side of the square) depending on the score obtained in last movement performed in the same direction. Assistive force is position-dependent, with constant magnitude and uniform throughout the whole movement. Its initial value is set at 10 N. We assessed the Fugl-Meyer score just before, just after the therapy protocol and at 3-month follow-up. We used a computational model [2] to describe the dynamics of the recovery process to describe the temporal evolution of performance during exercise. The model is linear and assumes that the subjects performance result from the sum of two components, the first proportional to the voluntary motor command, the second to the assistive force generated by the robot. To model the recovery process, the voluntary control is assumed to be the sum of three terms: a 'memory' term depending on the amount of voluntary control at the previous trial, a 'learning' component which is assumed to be proportional to an additional input - the driving signal - and an 'assistance' component, accounting for the dependence of the magnitude of the assistive force. Model fitting was evaluated by the correlation coefficient between the observed and predicted performance. Model parameters were then correlated with the short-term recovery, assessed in terms of clinical scales. Results and discussion: The study is still at a preliminary stage. In a previous study conducted with chronic patients, we found that neuromotor recovery has different dynamical properties for different motion directions. We expect this with acute patients engaged in a different task and with a novel, Bayesian scheme for the adaptive regulation of the forces provided by the robot.