A Functional Connectivity-Based Model With a Lightweight Attention Mechanism for Depression Recognition Using EEG Signals

Numerous studies on depression recognition utilize attention mechanisms as tools for feature extraction. Applying the standard multi-head self-attention mechanism to the spatial domain of EEG data is a feasible approach for extracting spatial features. However, there are challenges in the practical implementation. This algorithm generates a large number of model parameters and involves complex computations. Therefore, it heavily relies on computational resources with high computing power and incurs significant time costs. Furthermore, the randomness in the initialization process of these parameters potentially contributes to the instability of the model performance. In this study, we design a lightweight attention mechanism based on the standard multi-head self-attention mechanism, which generates fewer model parameters and incurs lower computational costs. In addition, we construct a deep learning model named Functional Connectivity Attention Network (FCAN) using this lightweight attention mechanism. FCAN can achieve effective depression recognition through EEG data and its coherence matrix. FCAN has two key components: the spatial attention module, which extracts deep spatial features of EEG data, and the feature integration module, which consolidates the extracted features. We evaluate the classification performance of FCAN and baseline models using a public EEG dataset. Our model achieves an accuracy of 95.20% (+/- 3.99%) and outperforms the baseline models in classification performance.