Dynamic Patient-Robot Assignment in a Simulated Stochastic Robotic Rehabilitation Gym

A robotic rehabilitation gym is a venue where multiple patients perform motor rehabilitation as a group using multiple robots. Such gyms produce positive outcomes, but it can be hard to create patient-robot assignment schedules that maximize patient skill improvement during exercise sessions. We thus present a neural-network-informed scheduler methodology that monitors patient exercise performance and dynamically assigns patients to robots based on their performance. These schedulers were trained and tested using synthetic datasets from a simulated rehabilitation gym. They were evaluated in 36 scenarios that differed in several ways (e.g., number of robots/patients, degree of stochasticity). Results showed that our neural-network-informed schedulers led to greater mean skill improvement compared to several baseline schedulers (e.g., only switching robots once). Additionally, the outcome difference between neural-network-informed schedulers and baseline schedulers increased as skill improvement became more stochastic. We also performed secondary studies where we showed that our schedulers still outperform baselines when patients can arrive or leave midsession and analyzed how the training dataset size affects scheduler performance. While several limitations need to be addressed before the scheduler is ready for use in real-world gyms, our study represents a step toward the use of artificial intelligence for patient monitoring in group settings.