An Explainable and Scalable EEG Framework for Major Depressive Disorder Detection and Severity Staging Through Signal Reconstruction, Channel Optimization, and Neuro-Fuzzy Intelligence

Abstract

In the modern world, Major Depressive Disorder (MDD) remains a serious global issue affecting mental health. Therefore, a reliable, non-invasive and scalable method for MDD diagnostics should be found. Analysis of EEG recordings shows that it can become the basis for automatic depression diagnosis and severity rating. Unfortunately, performance of models used for solving this problem is hindered by high dimensionality of data, noise, lack of clinical samples and channel redundancy. In this work, we propose a comprehensive method based on signal reconstruction, channel selection, data driven models, neuro-fuzzy systems and data augmentation techniques for identification and severity rating of MDD with EEG. First of all, several channel selection methods were applied to find the best features. Among them were Asymmetric Variance Ratio (AVR), Amplitude Asymmetry Ratio (AAR), Entropy-based Probability Mass Function (PMF) and Recursive Feature Elimination (RFE). The latter proved to be more efficient and selected 11 important channels minimizing computational costs. With the help of those channels, a Multi-Layer Perceptron (MLP) showed excellent performance achieving 98.7% accuracy, precision of 1.00, recall of 0.966 and F1-score of 0.983. It is significantly better than most common classifiers. Furthermore, to extend our approach to more than two classes, the classification of depression severity (Mild, Moderate, Severe) was conducted using CNN, Adaptive Neuro-Fuzzy Inference Systems (ANFIS) and GAN augmentation technique. The ANFIS proved to be robust and consistent providing 98.34% accuracy with balanced precision-recall values, while GAN allowed us to increase accuracy of severe-stage classification by 15%. In addition, a novel EEG reconstruction framework using Convolutional Autoencoders (CAE), MLP and Variational Autoencoders (VAE) was proposed in order to advance the eminence of the signal and scale up. CAE-MLP combination reached 96% of accuracy for MDD diagnosis, while VAE-based severity rating model showed the accuracy of 97.3% on original dataset and 97.84% on augmented one. Altogether, we managed to show that channel selection, explainable AI, neuro-fuzzy modelling and autoencoder-based signal reconstruction allow developing efficient and scalable approach to automatic diagnosis of depression severity using EEG.