CastNet: A three-channel EEG-based deep learning model for cross-subject depression detection.

Abstract

Depression is a serious mental health condition affecting millions worldwide. In recent years, deep learning models achieved remarkable performance in EEG-based depression diagnosis. This study proposes CastNet, a depression detection model utilizing three-channel EEG data. CastNet innovatively combines CNN, Transformer, and LSTM to achieve multi-level EEG feature learning, capturing local patterns, global representations, and long-range dependencies. To address the limited information in three-channel data, a feature enhancement module is introduced. Moreover, depthwise separable convolution and a convolutional dynamic sparse attention mechanism are employed to reduce the computational complexity associated with high temporal resolution EEG data. Additionally, a novel coupled bidirectional LSTM is proposed, enabling gate-level interactions to fuse bidirectional information during cell computation instead of conventional fusion at the output layer. Using the Leave-One-Subject-Out cross-validation (LOSO) strategy, experiments on the MPHC and PRED+CT datasets demonstrate that CastNet achieves classification accuracies of 97.2% and 87.5%, respectively, outperforming existing methods. These results not only provide new insights into the relationship between EEG signals and depression but also support the application of deep learning models in depression diagnosis.