서강대학교

초록 (요약문)

Many people experience abnormal walking, falls, and cognitive decline due to brain lesions, such as stroke. Modern medical institutions have introduced wearable rehabilitation robots to reduce the physical burden on therapists and provide patients with repetitive and active rehabilitation treatments. Therapists are expected to implement procedures such as robot parameter setting, patient motivation, and functional ability measurement for effective robot-assisted therapy. Because the outcomes of robot-based rehabilitation can vary depending on the proficiency of the therapist, a system is needed to implement the treatments of professional therapists consistently. This study aims to develop artificial intelligence (AI) that implements the expert’s therapeutic methods by processing time-series robot data and clinical data collected in robot-assisted gait training for stroke survivors. The AI’s output variables involved three therapeutic methods as follows: (1) realization of six verbal cues for correcting the patient’s posture and gait pattern, (2) estimation of the patient’s Borg scale every minute; and (3) estimation of the Berg balance scale (BBS), which represents the patient’s balancing ability. 107 stroke patients participated in 1,367 sessions of robot-assisted gait training (RAGT) using the Sogang University biomedical assistive robot (SUBAR). Prospective datasets were established, including collected clinical and robot data. These datasets included retrospective clinical data for 500 stroke patients who did not participate in the SUBAR RAGT to reflect more comprehensive patient information. The major input variables, known as feature sets, were selected using statistical tools to be maximally relevant to the output variables and minimally redundant for other input variables. The machine-learning algorithms were trained using prospective datasets and retrospective clinical data. The trained AI was inserted into the SUBAR and evaluated using the RAGT of stroke patients not included in the training dataset. The AI-equipped SUBAR implemented six verbal cues from a professional therapist at 93.7% accuracy, estimated the patient’s Borg scale at 90.9% accuracy, and estimated the patient’s BBS with 93.3% accuracy. The SUBAR robot equipped with the trained AI can contribute to improving the effectiveness of rehabilitation treatment by assisting therapists and patients during RAGT in hospitals.