Robotic Assistance for Training Finger Movement Using a Hebbian Model: A Randomized Controlled Trial

Abstract

Background: Robots that physically assist movement are increasingly used in rehabilitation therapy after stroke, yet some studies suggest robotic assistance discourages effort and reduces motor learning.

Objective: To determine the therapeutic effects of high and low levels of robotic assistance during finger training.

Methods: We designed a protocol that varied the amount of robotic assistance while controlling the number, amplitude, and exerted effort of training movements. Participants (n = 30) with a chronic stroke and moderate hemiparesis (average Box and Blocks Test 32 ± 18 and upper extremity Fugl-Meyer score 46 ± 12) actively moved their index and middle fingers to targets to play a musical game similar to GuitarHero 3 h/wk for 3 weeks. The participants were randomized to receive high assistance (causing 82% success at hitting targets) or low assistance (55% success). Participants performed ~8000 movements during 9 training sessions.

Results: Both groups improved significantly at the 1-month follow-up on functional and impairment-based motor outcomes, on depression scores, and on self-efficacy of hand function, with no difference between groups in the primary endpoint (change in Box and Blocks). High assistance boosted motivation, as well as secondary motor outcomes (Fugl-Meyer and Lateral Pinch Strength)-particularly for individuals with more severe finger motor deficits. Individuals with impaired finger proprioception at baseline benefited less from the training.

Conclusions: Robot-assisted training can promote key psychological outcomes known to modulate motor learning and retention. Furthermore, the therapeutic effectiveness of robotic assistance appears to derive at least in part from proprioceptive stimulation, consistent with a Hebbian plasticity model.

Keywords: hand; movement; proprioception; rehabilitation; robotics; stroke.

Figure 1

Experiment overview. (A) CONSORT flow diagram. (B) Annotated screenshot of the computer game used for this experiment. As popular music played, colored circles denoting “notes” flowed from left to right at the top (green), middle (blue), or bottom (yellow) of the screen toward like-colored targets. Targets were placed at 80% of full flexion. Smaller grey circles showed the position of the index and middle fingers. Participants had to start behind the green line and flex the correct finger(s) so as to stop inside of the targets just as the notes reached them. A score for each finger was displayed at the top of the screen, based on percentage of notes hit, calculated as moving average. Screenshot is for training the left hand; the screen and note directions were flipped when training the right hand (C) The FINGER robot used to provide participants with assistance as they played the computer game. Mechanisms attached to the index and middle fingers assisted in a naturalistic finger curling motion.

Figure 2

Quantified aspects of the robot-assisted training. (A) Mean number of practice movements across three weeks of training for participants in the high and low assistance groups. The dashed line shows the number of possible practice movements, calculated as the number of notes that the computer game presented to the participants. (B) Mean success rate in hitting notes for a song presented in the first training session with the robot adaptively assisting movement. The algorithm quickly adapted the amount of assistance during the song to achieve the target levels of success for the high and low assistance groups. (C) Mean success rates for a different song, part of the once/week assessment during which the robot did not assist in movement. In each panel, error bars and shaded regions denote +/− 1 SD. (D) Evolution of unassisted success rate across six evaluation sessions (baseline, week 1, week 2, week3, end of therapy (EOT), and one month follow-up (1MoFU), demonstrating motor learning of the computer game task. Participants played one song without assistance during evaluations. Participants improved performance of the song as they played it, then retained some of that improvement the next training session, associated with a net improvement in performance. Shaded regions show +/− 1 SD.

Figure 3

Clinical outcomes and covariates. (A) Change in Box and Blocks score measured from baseline to end of therapy evaluation and to the one month follow-up evaluation. * denotes significant difference, p < 0.05. (B) Change in the Upper Extremity Fugl-Meyer score. Error bars in A and B show +/− 1 SD. (C) Change in Box and Blocks score at the one month follow-up as function of proprioception error measured at baseline. (D) Change in Fugl-Meyer score, measured from baseline to the end of therapy, versus baseline Fugl-Meyer score. (E) Change in Fugl-Meyer score, measured from baseline to the one month follow-up, versus baseline Fugl-Meyer score.

Figure 4

Motivation and self-efficacy across training. (A) The weighting of the first principal component of the Intrinsic Motivation Inventory, across the nine training sessions. Individuals in the high assist group reported significantly higher motivation (p < 0.001) (B) The effort subscale of the IMI across the nine training sessions; the high assist group reported significantly higher effort (p < 0.002). (C) Change in self-efficacy across the three weeks of training. Self-efficacy was quantified by asking participants to estimate the probability of various scores in the Box and Blocks tests. The value on the y-axis is the number of additional blocks the participants estimated they could lift and place, with 50% confidence. Self-efficacy improved across training (p < 0.04), but the improvement was not different between groups. Error bars show +/− 1 SD.