An Interpretable Hybrid Neural Network Integrating Sinc-Convolution and Transformer for EEG-Based Depression...

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Electroencephalogram (EEG) recordings are valuable for depression detection.
• Convolutional Neural Networks (CNNs) show promise but have limitations in capturing global dependencies and interpretability.

Purpose of the Study:

• To propose an interpretable hybrid neural network, SINCFORMER-SHAP, for improved depression detection using EEG.
• To address limitations of CNNs regarding receptive field size and parameter count.

Main Methods:

• Developed SINCFORMER-SHAP, a hybrid model with spatial-frequency and temporal feature extraction modules.
• Employed sinc-based convolution for spatial-spectral patterns and Transformer for global time-domain dependencies.
• Utilized kernel visualization and SHAP for enhanced interpretability.

Main Results:

• SINCFORMER-SHAP achieved superior performance compared to state-of-the-art methods on public datasets (MODMA, EDRA, Mumtaz).
• Identified potential biomarkers in alpha and gamma rhythms across various brain regions.
• Demonstrated enhanced model interpretability through visualization and SHAP analysis.

Conclusions:

• SINCFORMER-SHAP offers an effective and interpretable approach for depression diagnosis using EEG.
• The findings advance the development of computational diagnostic techniques for practical psychiatric applications.