Prediction of chronic obstructive pulmonary disease based on multimodal data and deep learning

Area of Science:

• Medical Imaging
• Pulmonology
• Artificial Intelligence

Background:

• Early prediction of chronic obstructive pulmonary disease (COPD) is challenged by multimodal data integration, heterogeneity, and data gaps.
• Existing models often suffer from poor generalization due to these limitations.

Purpose of the Study:

• To propose a deep learning-based multimodal dynamic fusion network (MMDF-Net) for enhanced early prediction of COPD.
• To address challenges in utilizing multimodal information, modal heterogeneity, and data gaps.

Main Methods:

• MMDF-Net integrates chest CT images, pulmonary function indicators, and environmental exposure data.
• A dual-tower cross-modal contrastive learning module aligns image and non-image features.
• A conditional generative adversarial network generates environmental data, and a dynamic gating fusion mechanism adaptively weights modalities.

Main Results:

• MMDF-Net achieved an AUC of 0.92, sensitivity of 92.3%, and specificity of 88.7% on the COPD Gene dataset.
• The model significantly outperformed single-modal approaches.
• Dynamic weight adjustment based on patient attributes improved noise suppression and model performance.

Conclusions:

• The multimodal dynamic fusion strategy effectively addresses data heterogeneity and individual differences in COPD prediction.
• MMDF-Net provides a robust framework for precise early intervention in COPD.
• This approach demonstrates the potential of deep learning in complex disease prediction.