MS²-GNN: Exploring GNN-Based Multimodal Fusion Network for Depression Detection.

Area of Science:

• Computational psychiatry
• Machine learning for mental health
• Multimodal data fusion

Background:

• Major depressive disorder (MDD) is a prevalent and severe mental illness with significant societal impact.
• Existing multimodal methods for MDD detection show promise but overlook inter-modal heterogeneity/homogeneity and feature separability.
• There is a need for advanced fusion strategies that capture complex relationships within and between different data modalities and subjects.

Purpose of the Study:

• To propose a novel graph neural network (GNN)-based multimodal fusion strategy for enhanced major depressive disorder (MDD) detection.
• To address limitations of previous methods by investigating heterogeneity/homogeneity among psychophysiological modalities and inter-subject relationships.
• To develop a robust model that ensures data fidelity and achieves a compact multimodal representation for accurate MDD classification.

Main Methods:

• Developed a modal-shared and modal-specific graph neural network (GNN) architecture to extract inter- and intra-modal characteristics.
• Incorporated a reconstruction network to maintain fidelity within individual data modalities.
• Utilized an attention mechanism to generate a compact multimodal representation for MDD detection.

Main Results:

• The proposed GNN-based multimodal fusion strategy demonstrated superior performance in MDD detection.
• The method effectively captured heterogeneity and homogeneity across different psychophysiological modalities.
• Experiments on two public depression datasets confirmed the algorithm's effectiveness and robustness.

Conclusions:

• The modal-shared, modal-specific GNN approach offers a significant advancement in multimodal fusion for major depressive disorder detection.
• This strategy successfully addresses the limitations of earlier methods by considering complex data relationships and ensuring feature compactness.
• The findings highlight the potential of advanced GNN architectures in improving the accuracy and reliability of computational psychiatry tools.