AI Therapist Realizing Expert Verbal Cues for Effective Robot-Assisted Gait Training

Repetitive and specific verbal cues by a therapist are essential in aiding a patient's motivation and improving the motor learning process. The verbal cues comprise various expressions, sentences, volumes, and timings, depending on the therapist's proficiency. This paper proposes an AI therapist (AI-T) that implements the verbal cues of professional therapists having extensive experience with robot-assisted gait training using the SUBAR for stroke patients. The AI-T was developed using a neuro-fuzzy system, a machine learning technique leveraging the benefits of fuzzy logic and artificial neural networks. The AI-T was trained with the professional therapist's verbal cue data, as well as clinical and robotic data collected from robot-assisted gait training with real stroke patients. Ten clinical data and 16 robotic data are input variables, and six verbal cues are output variables. Fifty-eight stroke patients wore the SUBAR, a gait training robot, and participated in the robot-assisted gait training. A total of 9059 verbal cue data, 580 clinical data of stroke patients, and 144 944 robotic data were collected from 693 training sessions. Test results show that the trained AI-T can implement six types of verbal cues with 93.7% accuracy for the 1812 verbal cue data of the professional therapist. Currently, the trained AI-T is deployed in the SUBAR and provides six verbal cues to stroke patients in robot-assisted gait training.