Abstract Background Preserved ratio impaired spirometry (PRISm) and chronic obstructive pulmonary disease (COPD) are progressive respiratory disorders associated with accelerated pulmonary function decline and systemic comorbidities. This multicenter study aimed to develop a three-category classification model that integrates clinical variables with thoracic computed tomography (CT) radiomics to distinguish normal pulmonary function, PRISm, and COPD. Methods A total of 1018 participants from three centers (A, B, C) who underwent chest CT and pulmonary function tests (PFTs) within a 2-week interval were retrospectively analyzed. After applying inclusion and exclusion criteria, 797 individuals were included for analysis (Center A: 667 training/internal test = 534:133; Centers B, C: 130 external test). CT images were preprocessed via resampling and intensity normalization, followed by semi-automated segmentation of the airway tree and whole lung parenchyma using Mimics Research. PyRadiomics extracted 2436 radiomic features (1218 per region). Feature selection combined maximum relevance minimum redundancy with least absolute shrinkage and selection operator regression, employing tenfold cross-validation. Five models were developed using multinomial logistic regression: (1) clinical model, (2) airway model, (3) lung model, (4) airway fusion model, and (5) lung fusion model. Performance metrics included accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and the area under the receiver operating characteristic curve (AUC), with DeLong tests comparing model efficacy. Results 35 airway tree and 48 lung radiomic features were ultimately selected. The best performing model was the lung fusion model, which integrated three clinical predictors (age, gender, and BMI) with selected lung radiomic features. In external test set, it achieved superior performance with AUCs of 0.939 (95% CI 0.898–0.979) for PFT-normal, 0.830 (0.758–0.902) for PRISm, and 0.904 (0.841–0.966) for COPD, with an overall accuracy of 83.59%. DeLong tests indicated that across all three datasets, the lung fusion model outperformed the other four models. Conclusion Combining age, gender, BMI, and lung radiomic features significantly improves detection of PRISm and COPD compared to alternative models. These findings underscore the potential of CT-based radiomics for the early identification and risk stratification of abnormal pulmonary function.
Ma et al. (Thu,) studied this question.