Enhanced Depression Detection from EEG Using Deep Learning and Channel Selection on Synthetic-Balanced Data

EEG, or electroencephalogram, has emerged as a valuable and non-invasive method for evaluating depression. This study introducing a deep learning-based channel identification model for detecting depression using EEG signals. The dataset, sourced from AHEPA General Hospital, consists of EEG recordings from 293 patients, including 95 healthy individuals and 198 diagnosed with depression. To address the data imbalance, we applied Generative Adversarial Networks (GANs), increasing the healthy patient count to 184 and expanding the dataset to 382 subjects. EEG signals were recorded using 19 electrodes, following the 10–20 system across various frequency bands (Alpha, Beta, Delta, Gamma, High Beta, and High Gamma). ANOVA and Tukey’s test was used to filter the Significant Frequency Band (Alpha) for detailed research. Feature selection using Recursive Feature Elimination (RFE) reduced the model's input to 11 key channels: Fp1, Fp2, F7, F4, F8, T3, C3, Cz, T4, T5, and P3. A Multi- Layer Perceptron (MLP) model was employed, achieving an accuracy of 98.70%, with a Precision Score of 1.00, an F1-Score of 0.983, and a Recall Score of 0.967. These results highlight the efficacy of using deep learning in identifying key brain channels for the detection of depression.