ReTa-Diffusion: Exploring Task-State from Resting-State in EEG Signals via Bidirectional Decoupling and Latent Guiding for Early Detection of Subclinical Depression

WanYing Yang, KaiWen Guo, Shangen Han, Cheng He, Jinhong Ding, Huihui Fang, Miaomiao Yu, Chenyu Wang, Likun Xia

Keywords: Conditional Diffusion, Resting-state EEG, Subclinical Depression, Cross-subject, Riemannian Feature Extraction

TL;DR: By guiding the mining of stress features in resting-state electroencephalogram (EEG) signals through task-state EEG signals, subclinical depression can be diagnosed.

Abstract: Depression is a persistent and difficult-to-treat condition that poses a serious threat to public health. Early detection of subclinical or subthreshold depression (SD) is critical for preventing its progression into major depressive disorder. Compared to other modalities like wrist-worn wearables, electroencephalography (EEG) offers greater clinical value due to its direct link to brain function. Task-state EEG can reflect the characteristics of task related brain regions, its requirement for subjects to perform time-consuming tasks renders it impractical for rapid, large-scale screening. Whereas, resting-state EEG (rs-EEG) has advantages of task-free nature and ease of acquisition. Crucially, it contains spontaneous neural activity and can reflect the dynamic information across the entire brain, which is theorized to contain the neural substrates of all potential task-states. This makes it a highly valuable, yet underutilized, tool for SD prediction and diagnosis. However, effectively extracting task-specific information from rs-EEG remains a major challenge, hindered by two primary issues: (1) the limited understanding of relationships between resting-state and task-state EEG, and (2) the absence of effective guidance for feature extraction of generated task related EEG. To address the issues, we propose ReTa-Diffusion, a novel framework designed to mine task-related features from rs-EEG for the early detection of SD. It comprises two core modules: a Bidirectional Decoupled Conditional Diffusion (BDCD) module and a Wavelet-Riemannian Feature Extraction (WRFE) module. Inspired by the principle that task-state EEG can guide the interpretation of resting-state data to improve diagnostic accuracy, the BDCD addresses issue (1) by disentangling and aligning features from resting and task states through a bidirectional diffusion mechanism, thereby generating task-state-informed EEG signals from resting-state data. Subsequently, the WRFE tackles issue (2) by capturing rich dynamic temporal patterns and functional connectivity from these generated signals, leveraging Riemannian manifolds and a cross-attention mechanism. Together, these modules enable effective feature learning from rs-EEG, significantly enhancing the generalizability and accuracy of early SD detection. The proposed ReTa-Diffusion is evaluated using five-fold cross-validation on three datasets, achieving a classification accuracy of 54.52% on multiclass classification tasks, outperforming existing state-of-the-art methods by 18.95%. Further validation through leave-one-subject-out and visual analysis confirms its robust cross-subject capability and the effectiveness of using task-state features to guide stress-level prediction from rs-EEG. These results underscore ReTa-Diffusion’s potential as a powerful tool for early depression screening and prevention.

Primary Area: generative models

Submission Number: 12278