Subject-Independent Depression Recognition from EEG Using an Improved Bidirectional LSTM with Dynamic Vector Routing.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Biomedical Engineering

Background:

• Electroencephalography (EEG) is crucial for depression research, offering objective neurophysiological insights.
• Current EEG analysis methods struggle to fully utilize multi-domain signal information, limiting model generalization.
• Depressive disorders are associated with identifiable neurophysiological abnormalities detectable via EEG.

Purpose of the Study:

• To develop an improved bidirectional long short-term memory (BiLSTM) model for enhanced depression diagnosis using EEG.
• To effectively segment and analyze multi-channel temporal EEG sequences for improved classification.
• To investigate the utility of multi-frequency EEG data fusion for depression detection.

Main Methods:

• Continuous EEG signals were segmented into 2-second epochs.
• A BiLSTM encoder processed channel-time matrices (128 channels) after band-pass filtering and resampling.
• A dynamic-routing encapsulated-vector classifier was employed for end-to-end learning.
• Subject-independent five-fold cross-validation was used on the MODMA dataset.

Main Results:

• The proposed BiLSTM model achieved 84.8% accuracy and an AUC of 0.899.
• The method outperformed several representative baseline models, including SVM, EEGNet, InceptionNet, Self-attention-CNN, and CNN-LSTM.
• Joint analysis of multiple frequency bands improved classification performance compared to single-band analysis.

Conclusions:

• The developed BiLSTM model offers an effective approach for automatic depression diagnosis using EEG.
• Multi-domain fusion of EEG signals enhances diagnostic accuracy.
• This study highlights the potential of advanced deep learning models for psychiatric disorder diagnosis.