Deep Q-Learning for Channel Optimization in MRCP BMI Systems: A Teleoperated Robot Implementation

Abstract

Brain–machine interface (BMI) systems utilize brain signals to control external devices. Such systems can assist brain injury survivors who are partly or entirely unable to move the affected parts of the body. Electroencephalogram (EEG), a non-invasive method for recording brain signals in different locations of the subject’s scalp, is commonly used in such applications due to its cost-effectiveness and portability. Although EEG signals provide high temporal features, the signals are not robust due to both internal and external noises. Using all available EEG channels causes the system’s performance to deteriorate. Therefore, it is important to select the most informative channels to improve the system performance while reducing computation complexity. In this work, we propose a new Deep Q-Network (DQN)-based method to identify the best EEG channel combination for motor-related cortical potential tasks. The deep learning model is used to evaluate the DQN’s selected channels and provide feedback in terms of recognition rate. To evaluate the DQN’s performance, we compared the results with Genetic Algorithm and Backward Elimination channel optimization methods. Confusion matrix and recognition rates shows that the proposed DQN-based EEG channel optimization outperforms other methods. In addition, the results demonstrated that the DQN approach significantly reduced the number of channels while improving the BMI recognition rates. Furthermore, the EEG signals of optimized channels are used to control a teleoperated robotic hand in real time. The results of this work demonstrate the effectiveness of EEG channel optimization for the Internet of Things implementation of BMI systems.

Keywords

• Brain–machine interface (BMI)
• electroencephalogram (EEG)
• brain signals
• channel optimization
• deep reinforcement learning
• robotic hand