COPD-MMDDxNet: a multimodal deep learning framework for accurate COPD diagnosis using electronic medical records

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Pulmonology

Background:

• Chronic Obstructive Pulmonary Disease (COPD) affects over 391 million globally.
• Spirometry, the GOLD standard for COPD diagnosis, has limited availability and requires patient cooperation, posing challenges in primary care and resource-constrained settings.
• There is a critical need for alternative, accessible diagnostic methods for COPD.

Purpose of the Study:

• To develop and validate a novel multimodal deep learning framework, COPD-MMDDxNet, for diagnosing COPD without spirometry.
• To integrate structured pulmonary CT reports, blood gas analysis, and hematological data from electronic medical records (EMRs).
• To establish the first spirometry-independent multimodal diagnostic tool for COPD.

Main Methods:

• A multimodal deep learning framework (COPD-MMDDxNet) was developed, integrating parametric numerical embedding, hierarchical interaction mechanisms, contrastive regularization, and dynamic fusion coefficients.
• The framework was trained and evaluated on a dataset of 800 COPD patients with balanced demographics, collected over 24 months.
• Performance was compared against single-modality models and other state-of-the-art multimodal approaches.

Main Results:

• COPD-MMDDxNet achieved high diagnostic performance: accuracy (81.76%), precision (78.87%), recall (77.78%), and F1 score (78.32%).
• Ablation studies confirmed the significance of individual components, especially contrastive learning and cross-modal attention, in improving diagnostic accuracy.
• The model demonstrated superior performance compared to traditional single-modality and existing multimodal diagnostic methods.

Conclusions:

• COPD-MMDDxNet provides a robust, accurate, and accessible solution for COPD diagnosis, particularly in resource-limited environments.
• This spirometry-independent approach overcomes limitations of traditional diagnostic methods.
• The multimodal deep learning framework shows promise for improving COPD management globally.