A Graph-Discriminative Low-Rank Embedding Approach for Fusion of EEG and Eye Tracking in Depression Recognition

Sun, Wu; Zhang, Xin; Fan, Dongming; Wang, Xiuzhen; Yang, Linping; Gou, Jingfeng; Peng, Hong; Zhang, Hongqiang · 2026 · Crossref

DOI: 10.21203/rs.3.rs-9005389/v1

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of accurately diagnosing Major Depressive Disorder (MDD), which currently relies heavily on subjective clinical assessments that lack consistency and early detection capabilities. To overcome these limitations, the authors propose a multimodal approach fusing electroencephalography (EEG) and eye tracking (ET) data. While EEG provides high-temporal-resolution cortical activity and ET offers behavioral indicators of attention and cognitive control, integrating these heterogeneous signals into stable, interpretable features remains difficult. Existing fusion methods, such as simple concatenation or standard Canonical Correlation Analysis (CCA), often fail to capture complex cross-modal relationships or are vulnerable to noise and redundancy. To address this, the authors introduce the Graph Discriminative Low-Rank Correlation Embedding (GDLRCE) framework. This method integrates low-rank representation learning, graph-based discriminative constraints, and correlation analysis within a unified optimization model. The framework projects EEG and ET features into a shared low-rank subspace that preserves local structural information and class discriminability while minimizing redundancy. The study utilized data from 50 older adults (25 MDD patients and 25 healthy controls) who underwent three visual oculomotor paradigms: prosaccade (PS) for perceptual processing, antisaccade (AS) for executive control, and fixation stability (FS) for sustained attention. EEG and ET signals were recorded simultaneously, preprocessed using filtering and Independent Component Analysis, and then fused using the GDLRCE algorithm, which employs an Alternating Direction Method of Multipliers (ADMM) for optimization. Experimental results demonstrate that the GDLRCE framework outperforms unimodal approaches and conventional fusion techniques in depression recognition accuracy. The method effectively captures multistage brain-eye coupling patterns, revealing interpretable neurobehavioral biomarkers associated with MDD. By jointly maximizing cross-modal correlation and enforcing low-rank structure with graph-based supervision, the model enhances the separation between depressed and healthy subjects. The findings highlight the potential of this robust multimodal framework for automatic MDD identification and provide a mechanism-oriented analysis of the interaction between neural processes and observable behavior. This work contributes a novel, interpretable tool for affective computing, moving beyond black-box deep learning models to offer clinically relevant insights into the cognitive deficits associated with depression.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms