GCNs–FSMI: EEG recognition of mental illness based on fine-grained signal features and graph mutual information maximization

Abstract

There is growing evidence that an increasing number of people suffer from mental illness, which seriously affects their quality of life. The study of electroencephalography (EEG) is becoming crucial for understanding the root causes of mental illnesses and preventing them proactively. However, most of the existing approaches are unable to comprehensively investigate the fine-grained EEG data of each channel and the correlations between the signals of each subject. This makes it difficult to establish associations between EEG and mental illnesses. To address this issue, this paper proposes a new method called GCNs–FSMI that combines fine-grained signals and graph mutual information maximization for EEG recognition of mental illness patients. This method integrates fine-grained EEG signals, graph mutual information maximization technology, and pre-trained graph convolutional neural networks (GCN) to explore high-level interactions of multi-channel EEG data between subjects, leading to more discriminative EEG feature representation. This is achieved by filtering the EEG signals on each channel into six frequency bands, calculating the average power of each frequency band, and combining them as fine-grained features of the nodes. Then, we splice the multi-channel EEG signals of a sample into a signal series and construct a network based on the Pearson correlation coefficient to represent the association between channels. After that, we perform pre-training and fine-tuning processes to further improve the EEG feature representations. We pre-train a three-layer GCN model, fix the first two layers of the model, and fine-tune the output layer (third layer) of the model through contrastive learning with graph mutual information maximization. Finally, we conduct several experiments to test the effectiveness of our GCNs–FSMI model. Experimental results on two real-world datasets demonstrate that the features learned by GCNs–FSMI improve the performance of downstream classifiers by 6% and 9%. Furthermore, our method outperformed state-of-the-art methods in classification accuracy. To the best of our knowledge, the GCNs–FSMI is the first EEG recognition method for patients with mental illness that uses fine-grained signals and graph mutual information maximization. Our method provides a new approach to diagnosing mental illness and processing multi-channel biological neural signal processing.