MMDD: A Multimodal Multitask Dynamic Disentanglement Framework for Robust Major Depressive Disorder Diagnosis Across...

Area of Science:

• Neuroscience
• Computational Psychiatry
• Medical Imaging Analysis

Background:

• Major Depressive Disorder (MDD) diagnosis requires efficient automated methods.
• Current diagnostic approaches face challenges with feature reliance, data heterogeneity, and feature fusion.
• There is a need for advanced computational frameworks to improve MDD diagnosis.

Purpose of the Study:

• To propose the Multimodal Multitask Dynamic Disentanglement (MMDD) Framework for automated MDD diagnosis.
• To address limitations of traditional methods by integrating multimodal data and enhancing generalization.
• To investigate neurobiological patterns associated with MDD using computational modeling.

Main Methods:

• Developed a dual-pathway architecture using 3D ResNet for gray matter volume (GMV) and LSTM-Transformer for time series data.
• Implemented a Bidirectional Cross-Attention Fusion (BCAF) mechanism for dynamic multimodal integration.
• Employed Gradient Reversal Layer-based Multitask Learning (GRL-MTL) to improve domain generalization and mitigate site heterogeneity.

Main Results:

• The MMDD Framework achieved 77.76% classification accuracy on the REST-meta-MDD dataset.
• Ablation studies confirmed the critical roles of BCAF in fusion and GRL-MTL in generalization.
• Interpretability analysis revealed distinct neurobiological patterns in subcortical hubs, cerebellum, and cognitive cortices, with the middle cingulate gyrus showing cross-modal abnormalities.

Conclusions:

• Developed a robust and generalizable computational framework for objective MDD diagnosis using multimodal data.
• Provided data-driven insights into MDD's neuropathological processes, advancing understanding of the disorder.
• Highlighted the potential of advanced computational methods for psychiatric disorder diagnosis.