Abstract
Processing electroencephalogram (EEG) data using neural networks is becoming increasingly important in modern medicine. This study introduces the development of a neural network method using a combination of psychological questionnaire data and spectral characteristics of resting-state EEG. The data were collected from 71 individuals: 42 healthy and 29 with major depressive disorder (MDD). We evaluated four classes of algorithms-traditional machine learning, deep learning (LSTM), ablation analysis, and feature importance analysis-for two primary tasks: binary classification (healthy vs. MDD) and regression for predicting Beck Depression Inventory scores (BDI). Our results demonstrate that the superiority of a given method is task-dependent. For regression, an LSTM network applied to delta-rhythm EEG data achieved a breakthrough performance of R(2) = 0.742 (MAE = 6.114), representing an 86% improvement over traditional Ridge regression. Ablation studies identified delta and alpha rhythms as the most informative neurophysiological biomarkers. Furthermore, feature importance analysis revealed a triad of dominant psychometric predictors: ruminative thinking (31.2%), age (27.9%), and hostility (18.5%), which collectively accounted for 75.2% of the feature importance in predicting severity. LSTM on spectral EEG data provides a superior quantitative assessment of depression severity, while Logistic Regression on psychometric or EEG data offers a highly reliable tool for screening and confirmatory diagnosis. This methodology provides a robust, objective framework for MDD diagnosis that is independent of a patient's subjective self-assessment, thus facilitating enhanced clinical decision-making and personalized treatment monitoring.