Identifying EEG Biomarkers of Depression with Novel Explainable Deep Learning Architectures

If deep learning models trained on raw electroencephalogram (EEG) data are to be used in clinical or research contexts, methods to explain them must be developed. Moreover, in research contexts, methods for combining explanations across large numbers of models must be developed to counteract the randomness of existing training approaches. EEG model visualization-based explainability methods involve structuring a model architecture such that its extracted features and activations can be characterized. Nevertheless, model visualization-based explainability methods are underexplored in multichannel EEG. In this study, we present two convolutional neural network architectures and train them for automated major depressive disorder (MDD) diagnosis across 50 folds. We then perform two model-agnostic explainability analyses comparing the frequency bands and channels learned by each model before applying two model-specific explainability analyses that identify differences in extracted frequency bands and in the correlation of activations for each channel. Our models have mean sample-level accuracies around 83%, which are lower than the mean baseline architecture sample-level accuracy of 89%. Importantly, model-specific analyses find that individuals with MDD have higher β power, potentially higher δ power, and higher brain-wide correlation most strongly represented in the right hemisphere. This study provides multiple key insights into MDD and represents a significant step forward for the domain of explainable deep learning applied to raw EEG. We hope that it will inspire future efforts that will eventually enable the development of explainable EEG deep learning models that can contribute both to clinical care and novel medical research discoveries.