Title: Moving Past Augmentation: A Score-Driven Pathological Prior for Identifying Depression via EEG

Abstract:

The application of deep learning to detect Major Depressive Disorder (MDD) through Electroencephalography (EEG) is primarily hindered by the "small-sample dilemma." Traditional generative data augmentation techniques are often problematic; they demand significant computational resources and may inject synthetic noise that obscures the distinctions necessary for accurate classification. Challenging the conventional emphasis on maximizing data volume, we introduce a new framework titled "Beyond Augmentation," specifically Score-Guided Classification (SGC). Rather than generating pseudo-samples, SGC employs an unsupervised generative architecture to quantify the structural and statistical anomaly levels within samples. This metric functions as a crucial "Pathological Prior." Following robust normalization, this prior is integrated directly with deep feature representations to sharpen the classifier’s decision boundaries. Additionally, we present a Cross-Channel Spatial Adaptation module designed to handle fluctuating channel configurations. This module leverages a spatial mapping mechanism to address hardware inconsistencies and mismatched channels found in multi-center datasets. Our extensive evaluations on the Mumtaz2016 and high-density MODMA datasets confirm the method’s efficacy and strong generalizability. Notably, these results are achieved under a stringent "zero data augmentation" condition and at zero cost for sample synthesis.

Keywords: Electroencephalography (EEG), Depression Detection, Anomaly Score, Diffusion Models, Few-Shot Learning

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC