Transformer-Based Multi-Modal Data Fusion Method for COPD Classification and Physiological and Biochemical Indicators...

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Data Fusion Techniques

Background:

• Increasing biomedical data modalities necessitate advanced multi-modal fusion for complex biological relationship capture.
• Current fusion methods inadequately exploit intra- and inter-modal interactions in biomedical data.
• Powerful fusion techniques are underutilized in biomedical data analysis.

Purpose of the Study:

• To propose a novel multi-modal data fusion method addressing limitations in current approaches.
• To enhance disease classification by effectively integrating diverse biomedical data sources.
• To improve the exploitation of intra- and inter-modal interactions.

Main Methods:

• Utilizing Graph Neural Networks (GNNs) and 3D Convolutional Networks (3D CNNs) for intra-modal relationship identification.
• Employing Low-rank Multi-modal Fusion for effective multi-modal integration, noise reduction, and redundancy minimization.
• Incorporating Cross-modal Transformers for automated learning of inter-modal relationships and enhanced information exchange.

Main Results:

• The proposed method demonstrated superior performance in disease classification accuracy.
• Validation using lung CT imaging and physiological/biochemical data for Chronic Obstructive Pulmonary Disease (COPD).
• Outperformed various existing fusion methods and their variants in classification tasks.

Conclusions:

• The novel fusion method effectively integrates multi-modal biomedical data.
• Achieved superior disease classification accuracy, particularly for COPD.
• Highlights the potential of advanced fusion techniques in biomedical data analysis.