Feasibility of opportunistic screening for preserved ratio impaired spirometry using chest radiography-based deep learning models.

BACKGROUND

Although deep learning models using chest radiographs can estimate spirometry measurements, further investigation is needed to evaluate their ability to screen for preserved ratio impaired spirometry (PRISm), an important pre-COPD subtype. This study assessed whether a chest radiograph-based deep learning model can accurately detect PRISm and perform efficient opportunistic screening.

METHODS

This retrospective study included 54654 paired chest radiography and spirometry datasets from 38470 health checkup participants at a Japanese institution who underwent chest radiography and spirometry on the same day in 2018 and 2019. The dataset of 80% participants was used for model development, and the remaining 20% was used for performance evaluation. We developed a deep learning model that used frontal chest radiographs and demographic scalar inputs to detect PRISm and other pulmonary dysfunction subtypes, including FEV₁ decline, FVC decline, FVC and FEV₁ decline, and airflow limitation. Model performance was evaluated on the internal testing dataset. Subgroup analyses were performed across five independent factors: age, height, ppFVC, ppFEV₁, and FEV₁/FVC.

RESULTS

The model achieved an AUROC of 0.892 (95% CI, 0.875-0.906), a sensitivity of 71.6%, and a specificity of 87.9% for PRISm detection on the testing dataset (N = 10917). No significant differences in AUROC were observed across subgroups defined by age, height, FEV₁/FVC, or ppFEV₁. The AUROC values for detecting all pulmonary dysfunction subtypes exceeded 0.8.

CONCLUSIONS

The chest radiography-based model can effectively detect PRISm and may be useful for opportunistic screening of PRISm for supporting early risk management and intervention for pre-COPD conditions and obstructive pulmonary diseases.