Interpretable chronic obstructive pulmonary disease identification using chest X-ray radiomics: a multicenter study

OBJECTIVES: To construct and validate a combined model integrating chest X-ray (CXR)-based radiomic features and clinical characteristics for chronic obstructive pulmonary disease (COPD) identification, while enhancing model interpretability. MATERIALS AND METHODS: Paired CXR images and clinical data were collected from 17 hospitals between January 2017 and December 2023. Data from 11 centers were divided into a training cohort and an internal validation cohort at a 7:3 ratio, with data from the remaining 6 centers serving as an external validation cohort. Three models (radiomic model, clinical model, and combined model) were constructed, and the SHapley Additive exPlanations (SHAP) method was used to interpret model performance. RESULTS: A total of 2433 participants were enrolled, with a mean age of (66.9 ± 11.4) years, including 1564 males and 819 COPD patients. The radiomic model achieved AUCs of 0.760, 0.754, and 0.764 in the training, internal validation, and external validation cohorts, respectively, which were significantly higher than those of the clinical model (AUCs: 0.631, 0.651, and 0.673; all p < 0.001). SHAP analysis revealed that age, radiomic features, smoking history, and sex were crucial for COPD identification. CONCLUSIONS: This study successfully constructed a CXR-based combined radiomic-clinical model for COPD, which demonstrated good performance and high accuracy in identifying COPD in this multicenter study. The SHAP method enhanced the model's interpretability and clinical applicability. CRITICAL RELEVANCE STATEMENT: This study develops a CXR radiomic-clinical COPD identification model with SHAP-enhanced interpretability, advancing interpretable, widely applicable COPD screening in clinical radiology. KEY POINTS: The clinical screening rate for COPD remains severely inadequate. The combined model integrating chest X-ray radiomic features and clinical variables enables accurate differentiation between patients with COPD and non-COPD individuals. Global SHAP analysis reveals that radiomic features are the primary factor influencing COPD identification, followed by age, sex, and smoking status. Local SHAP analysis can intuitively visualize the model's decision-making process at the individual sample level.