Automatic focal EEG identification based on deep reinforcement learning. (May 2023)

• Record Type: Journal Article
• Authors: Liu, Xinyu; Ding, Xin; Liu, Jianping; Nie, Weiwei; Yuan, Qi
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: EEG -- Focal and non-focal -- AFM-DQN -- Deep learning -- Reinforcement learning -- Semi-JMI algorithm; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104693
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel AFM-DQN model is proposed for focal EEG recognition. Unify the decision making capability of Q-learning and the perceptual capability of CNN. Semi-JMI algorithm is introduced for feature selection to reduce computational burden. Comparisons with other methods indicate the better performance of this model. Abstract: Electroencephalogram (EEG) signals convey information about the electrical activity of neurons and are commonly used in clinical practice to evaluate the epileptic activity of patients. There is a part of the human brain that is closely linked to epileptic activity, namely the epileptogenic zone. Successful resection of the epileptogenic zone requires high precision classification of focal (F) and non-focal (NF) EEG signals. In this paper, we improve the Deep Q-Network (DQN) by adding Additional Functional Modules (AFM) to propose a novel focal EEG recognition method which is named AFM-DQN. Compared to the traditional Reinforcement Learning (RL), this model incorporates a deep convolutional neural network (CNN), and unifies the decision making capability of Q-learning and the perceptual capability of CNN, which greatly improves the learning of this network. The AFM, including pre-training, the high performance classifier (HPC), reward control mechanism and three RL related techniques, imparts this model a stronger generalization ability. To reduce computational burden of the network while preserving the correlation and interdependence of EEG signals, we introduce different kinds of EEG features and a semi-supervised feature selection algorithm based on joint mutual information (Semi-JMI algorithm). Experiments on the two publicly available EEG databases have achieved classification accuracy of 95.87% and 97.5%, respectively, which demonstrates that the proposed method has the potential for clinical application of accurate localization of epileptogenic zone. Compared with other published methods, this method has better performance and generalization ability.