Comparative Analysis of Depression Detection Using EEG Signals

This paper presents a comprehensive analysis of Multilayer Perceptron (MLP) models for the classification of EEG signals in the context of depression state detection. Experiments were conducted using two separate databases: the Depression Rest Database and the MDD vs. Control Database. For the Depression Rest Database, the MLP model reached an accuracy of 84.65% on the training set but faced challenges with validation, plateauing at 68.79%. Conversely, the MLP model excelled in the MDD vs. Control Database, achieving an accuracy of 89.99% on the training data and 88.97% on the validation data. It displayed high precision and recall values for both healthy and depressed classes, indicating a balanced performance. Additionally, feature selection was explored on a combined database, yielding promising results but with room for further optimizations. The novelty of this study consists in unifying two EEG datasets that are using different hardware acquisition devices and protocols and investigating if the method demonstrates a more balanced performance between the two classes and a superior performance compared to the individual database analyses.