Multi-View Deep Clustering for Depression Subtype Identification Across Multi-Band EEG Topographies

Abstract

The significant heterogeneity of Major Depressive Disorder (MDD) necessitates objective subtyping that transcends traditional symptomatology. With the development of machine learning algorithms, using machine learning to identify potential biomarkers has emerged as an efficient approach. However, existing machine learning-based studies on depression are mostly limited to simple binary classification or rely on pre-set features. This study validates a novel unsupervised framework, integrating a CNN-Transformer with the DEMVC model, to identify depression subtypes from resting-state EEG topographies across different frequency bands. The results show that the fusion of multi-band significantly superior to single-band approaches in clustering stability and subtype separability. Specifically, the Alpha-Beta-Theta combination emerged as the best band combination. Ablation studies further revealed the distinct roles of the frequency bands: the Theta band captures core depressive pathology, whereas the Alpha band may be better considered as an indicator of 'near-healthy status' rather than a mere disease marker. This study innovatively takes multi-bands as multi-view inputs for multi-view deep clustering, resolving the challenge of classifying depression subtypes and providing a robust and explicable new paradigm for this task.