Robot-aided rehabilitation effectively supports treatment of upper-limb disorders and enhances outcomes when combined with traditional therapy. Artificial intelligence enables behavioral cloning of physiotherapists' expertise to autonomously modulate robot assistance from real-time multimodal patient data. Therefore, this paper aims to propose and validate a behavioral cloning strategy, namely Physiotherapist-Supervised Parameter Adaptation (PSPA), for online tuning the robot assistance level replicating the physiotherapists' decision-making. The experimental validation was conducted in a clinical setting involving ten post-surgical orthopedic patients who participated in a robot-aided rehabilitation session using the KUKA LWR 4+ robot. The sessions were supervised by physiotherapists who could adjust the level of robotic assistance as needed, thus labelling the collected patient multimodal data. The validation aimed at i) identifying the best-performing input modality, feature set, and classifier, and ii) comparing the capability of the approach in tailoring the assistance level with respect to the established performance-based (PB) one. Combining biomechanical and physiological features significantly improved the classification performance across all classifiers, with the highest performance observed for the Multi-layer Perceptron on the present dataset. Moreover, using the optimized feature set, the proposed PSPA methodology achieved an even greater alignment with the physiotherapists' decisions with respect to the PB approach (ΔF1-score = 15.40 ± 30.33%, ρ = 0.56 ± 0.21 for PSPA, ρ = -0.12 ± 0.43 for PB).
Molle et al. (Thu,) studied this question.