AttenStateNet: Depression Classification Method for EEG Signals Based on State Space and Attention Mechanisms

Electroencephalography (EEG) is a powerful tool for capturing neural signals within the brain, and has been increasingly applied in the detection of neurological disorders such as depression and so on. Recent advances in deep learning have led to significant progress in EEG-based depression detection models. However, these models still face some challenges, particularly in addressing the varying importance of different EEG channels and the features within the same channel. To overcome these challenges, this work proposes AttenStateNet, a novel model that integrates the state space models and attention mechanisms. AttenStateNet effectively captures long-term dependencies in EEG signals through the structured state-space model, while utilizing the SENet module to dynamically adjust the attention on different channels and features. This model aims to refine the extraction of spatiotemporal features, making it highly effective for depression classification tasks. Experimental results based on two public depression datasets demonstrate that AttenStateNet has the good performance achieving the accuracies of 99.87% on the MODMA dataset and 99.67% on the EDRA dataset. The results show its robust performance in handling complex EEG data for depression prediction.