Dual Adaptive Disentangled Representation Learning With Multimodal Data for Disease Diagnosis

Area of Science:

• Biomedical data analysis
• Computational biology
• Medical imaging and genetics

Background:

• Multimodal data fusion for disease diagnosis faces challenges due to data heterogeneity.
• Exploring consistency and variability across similar diseases is crucial for improving model performance.

Purpose of the Study:

• To propose a unified framework, dual adaptive disentangled representation learning (DADRL), for simultaneous biomarker detection and disease diagnosis.
• To address challenges in multimodal data fusion and leverage information from similar diseases.

Main Methods:

• Developed a biology information constraints-based modality fusion strategy to explore inter- and intra-modal correlations.
• Integrated modality fusion and disease diagnosis within a unified framework.
• Incorporated disentangled representation learning and adaptive metric constraints to separate disease-specific and shared features.

Main Results:

• The DADRL framework effectively fuses heterogeneous multimodal data (imaging and genetic).
• Achieved simultaneous disease-shared and disease-specific biomarker detection.
• Significantly improved performance in disease diagnosis compared to existing methods.

Conclusions:

• DADRL offers a novel approach to biomarker detection and disease diagnosis by effectively handling multimodal data heterogeneity.
• The method enhances understanding of disease pathogenesis by separating disease-specific and shared features.
• Demonstrated significant improvements in both biomarker detection and disease diagnosis accuracy across multiple datasets.