Robust Brain-Machine Interface Design Using Optimal Feedback Control Modeling and Adaptive Point Process Filtering

摘要

Author Summary Brain-machine interfaces (BMI) decode subjects' intended movements from neural activity. Despite significant advances, performance, robustness, and extendibility remain key challenges. Current BMIs use decoders such as Kalman filters whose parameters are estimated in closed-loop BMI operation. However, current decoders do not model the spikes directly and hence may limit the time-scale of neural processing, control, and parameter estimation. Here we develop a novel BMI training architecture for spike-event-based control and parameter estimation, and show that it enables robust control and extends to various tasks. Moreover, we propose a control-theoretic framework for closed-loop decoder training. The BMI incorporates an optimal feedback-control (OFC) model of brain's control behavior to infer its motor intention. The BMI uses a point process to model the spikes; this enables the subject to control the movement and the decoder to estimate the parameters using spike events. Using closed-loop experiments in a non-human primate, we show that the OFC model improves performance compared with current intention estimation techniques; that spike-event-based adaptation enables faster performance convergence compared with current batch-based methods, and is robust to parameter initialization; and that the architecture extends to various tasks. This architecture has significant implications for future clinically-viable BMIs.