Reinforcement Learning-based Kalman Filter for Adaptive Brain Control in Brain-Machine Interface.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-Machine Interfaces (BMIs) translate neural signals into neuro-prosthesis commands for paralyzed individuals.
• Dynamic changes in neural patterns during brain control (BC) pose challenges for decoder adaptation.
• Kalman Filters (KF) require re-calibration for performance with significant neural pattern shifts, while Reinforcement Learning (RL) is less suited for continuous motor state generation.

Purpose of the Study:

• To develop an adaptive algorithm for Brain-Machine Interfaces that can co-adapt with dynamic neural patterns.
• To enable continuous motor state prediction despite neural signal drift over time.
• To improve the robustness and performance of neuro-prosthetic control in clinical applications.

Main Methods:

• Proposed a novel reinforcement learning-based Kalman filter (RLKF) algorithm.
• Maintained KF's state transition model for continuous motor state prediction.
• Utilized RL for adaptive action generation from neural patterns, correcting KF predictions based on reward signals.

Main Results:

• The RLKF algorithm demonstrated online adaptation to neural pattern drift in a simulated rat lever-pressing experiment.
• Maintained high performance by effectively tracking neural pattern changes across days.
• Outperformed standard KF without re-calibration in handling dynamic neural data.

Conclusions:

• The RLKF method successfully bridges online parameter adaptation and continuous neuro-prosthesis control.
• This adaptive approach is promising for real-world clinical applications of brain control.
• The algorithm offers a robust solution for maintaining BMI performance with evolving neural signals.