Explainable Depression Recognition from EEG Signals via Graph Convolutional Network

Depression is a prevalent mental disorder that poses significant risks to human health and social stability. Current methods for diagnosing depression heavily rely on patient descriptions and psychiatrist observations, which are susceptible to interference from subjective factors and carry the risk of misdiagnosis and missed diagnosis. Therefore, it is crucial to develop an objective method for recognizing depression based on objective criteria. Recently, combining EEG signals with deep learning techniques for depression recognition has become a popular research topic. However, existing EEG-based depression recognition methods are poorly interpretable, making it challenging to explain the neural mechanisms of depression disorders. Consequently, we propose an explainable framework for depression recognition from EEG signals based on a GCN. In this method, a hybrid module of 1DCNN, LSTM and GCN is utilized to extract features from EEG signals, which can effectively capture spatiotemporal correlations between different brain regions. The EEG subgraph construction module explores the differences in crucial connectivity patterns of the brain between different groups, enhancing the interpretability of our model. The experimental results on the MODMA dataset show that our model outperforms the baseline model in all metrics, thus verifying the validity of the proposed model. Additionally, compared with existing explainable algorithms, our method consistently yielded nearly identical experimental results, demonstrating its ability to capture the correlation between depression and neuroscience, and has good interpretability.