Smart Neurostimulation: Sensor-Based Wireless Stimulation With Deep On-Policy MPC

Closed-loop deep brain stimulation (cl-DSB) is recognized as one of the most prominent schemes of neurostimulation therapy which aims to reduce the level of hand tremor in Parkinson's disease (PD). Recently, modern neurostimulation has integrated wireless sensors and programming devices with communication infrastructures to monitor and regulate hand speed remotely. However, deploying wireless communication technologies introduces degradation procedures (i.e., latency, time delay, and so on) which adversely affect the neurostimulation treatment. Thus, in this paper, an adaptive fractional-order model predictive control (FOMPC) is designed to address the challenges of wireless neurostimulation in the cl-DSB system. For this purpose, a deep on-policy learning (DOL) algorithm is adopted to design parameters of the FOMPC in such a way that minimizes the fluctuations of hand speed and eliminates side effects of stimulation. Compared with the integer version of the model predictive controller, further flexibility is achieved by the non-integer orders. Deep neural networks (DNNs) of on-policy learning are trained to regulate tremor speed by dynamically adjusting the parameters of FOMPC in an episodic manner. The main novelty of this work is to consider the challenges of unreliable wireless communications, as it allows us to investigate the effects of latency and time delays in cl-DSB systems. This makes the suggested controller a prominent option for real-time applications where the presence of time delay can threaten the system's stability. Extensive simulation scenarios under various levels of random communication delays are considered to assess the feasibility of suggested stimulators for wireless cl-DSB systems. Comparative analysis of cl-DBS's responses reveals the superiority of the suggested FOMPC controller (designed by deep-on-policy learning) to regulate hand tremors over other state-of-the-art strategies.