EEG-Based Depression Detection Using Welch Power Spectral Density Analysis and Transformer Learning

Abstract

Depression diagnosis remains largely dependent on subjective clinical assessments, creating a strong need for objective and reproducible biomarkers. Electroencephalography (EEG) offers a non-invasive and cost-effective solution for capturing neural activity associated with depressive states. This study proposes a deep learning-based framework for automated depression detection using EEG Power Spectral Density (PSD) features estimated through the Welch method. After preprocessing with bandpass filtering, notch filtering, and Independent Component Analysis (ICA)-based artifact removal, PSD features were extracted from the delta, theta, alpha, beta, and gamma bands. These spectral representations were then classified using Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Transformer-based architectures. Experimental results on EEG recordings from 90 subjects (60 depressed and 30 healthy controls) demonstrated that the Transformer model achieved a classification accuracy of 83.78%, with favorable convergence, generalization, and computational efficiency compared with the CNN and RNN models. The findings confirm that Welch-based spectral analysis combined with Transformer learning provides an effective and scalable framework for objective depression detection, with strong potential for clinical decision support and portable mental health monitoring systems.