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

To address the issues of insufficient utilization of multimodal information, modal heterogeneity, and data gaps leading to poor model generalization in early prediction of chronic obstructive pulmonary disease (COPD), a deep learning-based multimodal dynamic fusion network (MMDF-Net) is proposed. This model integrates chest CT images, pulmonary function indicators, and environmental exposure data. It aligns image and non-image features via a dual-tower cross-modal contrastive learning module, mitigating semantic differences across modalities. A conditional generative adversarial network is used to generate high-fidelity environmental exposure data, reducing reliance on data completeness. A dynamic gating fusion mechanism adaptively adjusts multimodal weights based on patient smoking history, age, and other attributes to suppress noise. On the COPD Gene dataset, MMDF-Net achieves an area under the curve (AUC) of 0.92, a sensitivity of 92.3%, and a specificity of 88.7%, significantly outperforming single-modal models and dynamically adjusting weights according to disease stage. These results demonstrate that this multimodal dynamic fusion strategy can effectively address data heterogeneity and individual differences, providing technical support for precise early intervention in COPD.