COPD-MMDDxNet: A Multimodal Deep Learning Framework for Accurate COPD Diagnosis Using Electronic Medical Records

Yuanyuan Yi¹Lei Shi²Haoran Liu³Mingyu WANG³Min Feng^1*Yanxia Li^1*

• ¹Department of Respiratory Medicine, The First Affiliated Hospital of Dalian Medical University, Dalian, China
• ²Department of Emergency Medicine, First Affiliated Hospital, Dalian Medical University, Dalian, Liaoning Province, China
• ³Dalian Medical University, Dalian, Liaoning, China

COPD affects approximately 391 million people globally. While spirometry is recognized as the gold standard for diagnosing COPD according to the GOLD guidelines, its availability is limited in primary healthcare settings, particularly in low-and middle-income countries.Furthermore, spirometry requires patient cooperation, which may be challenging for individuals with physical limitations or comorbidities, potentially impacting its accuracy. As a result, there is a need for alternative diagnostic methods, particularly those suited for resourceconstrained environments.This study proposes a novel multimodal deep learning framework, COPD-MMDDxNet, which integrates structured pulmonary CT reports, blood gas analysis, and hematological analysis from electronic medical records (EMRs) to overcome the limitations of existing diagnostic methods. This framework develops the first multimodal diagnostic tool for COPD that does not rely on spirometry. It innovatively fuses cross-modal data, incorporating four key components: parametric numerical embedding, hierarchical interaction mechanisms, contrastive regularization strategies, and dynamic fusion coefficients. These innovations significantly enhance the model's ability to capture complex cross-modal relationships, thereby improving diagnostic accuracy.The dataset used in this study comprises 800 COPD patients, with a balanced age and sex distribution, and data were collected over a 24-month period.Experimental results demonstrate that COPD-MMDDxNet outperforms traditional single-modality models and other state-of-the-art multimodal models in terms of accuracy (81.76%), precision (78.87%), recall (77.78%), and F1 score (78.32%). Ablation studies confirm the critical importance of each model component, particularly the contrastive learning module and cross-modal attention mechanism, in enhancing model performance.This framework offers a robust solution for more accurate and accessible COPD diagnosis, particularly in resource-constrained environments, without the need for spirometry.