Novel EEG-based diagnostic framework for Major Depressive Disorder using microstate and entropy features

This study aimed to develop a novel, non-invasive diagnostic framework for Major Depressive Disorder (MDD) by integrating EEG-based entropy and microstate features, transformed into image representations for deep learning analysis. EEG data were obtained from 63 MDD patients and 36 healthy controls, collected from two publicly available datasets. To address class imbalance, data augmentation techniques including Gaussian noise addition, time warping, and frequency perturbation were applied to the healthy control group, resulting in a balanced dataset of 126 subjects. EEG signals were decomposed into five canonical frequency bands: delta (0.5-4 Hz), theta (5-8 Hz), alpha (9-13 Hz), beta (14-30 Hz), and gamma (31-50 Hz). Microstate segmentation was performed for each frequency band, yielding five microstates (A to E) that reflect distinct spatiotemporal brain dynamics. From these microstates, three key features (duration, occurrence, and coverage) were extracted alongside entropy measures (Shannon, sample, and approximate). All features were transformed into 2D image representations to capture spatial and temporal patterns and were subsequently classified using Convolutional Neural Networks (CNN), Deep Neural Networks, and Random Forest. Model performance was evaluated using fivefold cross-validation with accuracy and ROC-AUC metrics. The proposed framework enabled CNN models to achieve outstanding classification performance. Entropy-based images yielded an accuracy of 99.60 ± 0.22% and a ROC-AUC of 99.96 ± 0.02%, while microstate-based images achieved an accuracy of 96.96 ± 4.79% and ROC-AUC of 98.94 ± 2.27%. Notable group differences were observed in microstate dynamics, with microstate E exhibiting consistent reductions in both duration and occurrence among MDD patients, particularly within the alpha band at an 80-s window size (21.12% reduction in occurrence, p < 0.01). Transition probability analysis revealed altered state-switching patterns in the beta and gamma bands. Discriminative activity was most prominent in the occipital and frontal regions across delta and gamma frequencies. Integrating entropy and microstate-derived features with deep learning presents a high-accuracy, non-invasive solution for diagnosing MDD. The study highlights significant disruptions in functional brain dynamics, particularly in the Default Mode Network and prefrontal areas, reflected through microstate alterations and entropy variability. These results provide deeper insight into the neurophysiological signatures of depression and support the development of objective EEG-based biomarkers to complement clinical psychiatric evaluations and guide personalized interventions.