# Attention-based multi-scale convolution and conformer for EEG-based depression detection

## TL;DR

This paper introduces a new EEG-based model for detecting depression using attention mechanisms and deep learning, achieving high accuracy with potential for practical use.

## Contribution

The novel AMCCBDep model combines AMPC, Conformer, and BiGRU for improved depression detection from EEG signals.

## Key findings

- The AMCCBDep model achieved 98.68% ± 0.45% accuracy on the MODMA dataset.
- Reducing electrodes from 128 to 16 had minimal impact on model performance.
- The model shows strong potential for scalable depression detection in clinical settings.

## Abstract

Depression is a common mental health issue, and early detection is crucial for timely intervention. This study proposes an end-to-end EEG-based depression recognition model, AMCCBDep, which combines Attention-based Multi-scale Parallel Convolution (AMPC), Conformer, and Bidirectional Gated Recurrent Unit (BiGRU). The AMPC module captures temporal features through multiscale convolutions and extracts spatial features using depthwise separable convolutions, while applying the ECA attention mechanism to weigh key channels, enhancing the model’s focus on crucial electrode channels. The Conformer module further captures both global and local temporal dependencies in EEG signals to ensure the capture of long-range dependencies and local patterns. The BiGRU module improves the model’s ability to recognize depressive states by utilizing bidirectional modeling. We used the 128-channel resting-state EEG signals from the MODMA dataset, which includes data from 24 depression patients (13 males, 11 females, aged 16 to 56) and 29 healthy individuals (20 males, 9 females, aged 18 to 55). Experimental results show that the AMCCBDep model achieved an accuracy of 98.68% ± 0.45% on the MODMA dataset. The model evaluation results for both 128-channel and 16-channel configurations demonstrate that reducing the number of electrodes has a minimal impact on performance, suggesting that electrode reduction could be considered in practical applications. This model showcases strong potential in advancing depression detection in neuroscience, providing an efficient and scalable solution for clinical and practical applications. Future research will further optimize model performance and explore the impact of reducing the number of electrodes on clinical practice.