Machine learning models using dual-phase CT radiomics for early detection of PRISm

Preserved Ratio Impaired Spirometry (PRISm) is considered an early stage of chronic obstructive pulmonary disease (COPD), which may either revert to normal or progress to COPD. Therefore, early identification is crucial for improving patient prognosis. In this study, we developed multiple machine learning (ML) models based on inspiratory and/or expiratory breath-hold chest computed tomography (CT) images to identify PRISm. A total of 270 subjects were prospectively enrolled, and clinical models, radiomics models, and combined clinical-radiomics models were constructed using inspiratory, expiratory, and dual-phase CT images, respectively. The results demonstrated that combined models outperformed clinical models alone across all three phases. Among them, the logistic regression (LR)-based combined models using expiratory or dual-phase CT achieved the best performance, with comparable area under the receiver operating characteristic curve (AUC) values and superior performance to the inspiratory-phase models. Specifically, the AUCs (95% confidence intervals [CI]) of the clinical model in the training, internal, and external validation sets were 0.825 (0.750-0.900), 0.771 (0.639-0.903), and 0.778 (0.653-0.904), respectively. For the expiratory-phase combined model, the AUCs were 0.901 (0.845-0.956), 0.819 (0.680-0.957), and 0.817 (0.695-0.940), while for the dual-phase combined model, they were 0.901 (0.846-0.955), 0.821 (0.684-0.957), and 0.813 (0.694-0.932), indicating that adding inspiratory data did not significantly improve model performance. Based on these findings, we recommend that single-phase expiratory CT scans, combined with clinical features and analyzed using LR models, be prioritized in clinical practice for efficient PRISm identification, providing support for early diagnosis and timely intervention.