Multi-Loss Disentangled Generative-Discriminative Learning for Multimodal Representation in Schizophrenia

Area of Science:

• Neuroscience
• Genetics
• Microbiome Research

Background:

• Schizophrenia (SCZ) is a complex mental illness requiring multimodal data analysis for comprehensive understanding.
• Existing multimodal approaches often fail to fully leverage inter-modal relationships or modality-specific and common components.
• A unified model for simultaneous classification and regression tasks in SCZ is needed.

Purpose of the Study:

• To develop a novel multi-loss disentangled generative-discriminative learning (MDGDL) model for SCZ research.
• To effectively integrate functional magnetic resonance imaging (fMRI), genetic, and gut microbiome data.
• To address limitations in existing multimodal analysis methods.

Main Methods:

• Developed a multi-loss disentangled generative-discriminative learning (MDGDL) model.
• Utilized disentangled learning to represent gene and gut microbial data into modality-specific and common vectors.
• Employed a generative-discriminative framework to link fMRI features with latent vectors, generating attentive vectors.

Main Results:

• The MDGDL model demonstrated superior performance in SCZ classification and cognitive score regression tasks.
• The model successfully identified crucial multimodal biomarkers associated with SCZ.
• Attentive vectors derived from the model enhanced the utility of fMRI features for downstream analysis.

Conclusions:

• The MDGDL model offers a powerful approach for analyzing multimodal data in SCZ research.
• This method provides complementary insights beyond traditional single-modality analyses.
• The findings pave the way for a more integrated understanding of SCZ pathophysiology.