Explainable machine learning and multivariable logistic regression for identifying prevalent self-reported physician-diagnosed chronic lung disease in postmenopausal women: development and multi-cohort external validation

Background Chronic lung disease (CLD) is an important cause of morbidity among postmenopausal women, yet simple non-invasive community screening tools are lacking. We aimed to develop and externally validate an interpretable multivariable prediction model for identifying prevalent self-reported physician-diagnosed CLD in this population. Methods We analyzed postmenopausal women from the US Health and Retirement Study (HRS; development n = 4,578, internal validation n = 1,962), with external validation in the English Longitudinal Study of Ageing (ELSA; n = 4,574) and a Chinese clinical cohort ( n = 650). The primary outcome was prevalent self-reported or clinically recorded physician-diagnosed CLD, including chronic bronchitis, emphysema, or COPD where available. Predictors were selected via nested LASSO penalization. Twelve candidate statistical and machine-learning algorithms were benchmarked; the primary multivariable logistic regression model was evaluated on native uncalibrated probabilities at a locked screening threshold (9.81%) across discrimination, calibration, and clinical net benefit. Results The current 9-predictor model retained smoking status, self-rated health, arthritis, heart disease, osteoporosis, basic activities of daily living difficulty, depressive symptoms, physical activity, and employment. The multivariable logistic regression model achieved AUROCs of 0.771 (95% CI, 0.748–0.793) in HRS development, 0.769 (0.734–0.803) in internal validation, 0.762 (0.735–0.790) in ELSA external validation, and 0.693 (0.634–0.750) in the Chinese cohort. Current smoking and poor self-rated health showed the strongest associations with prevalent CLD. In the Chinese clinical cohort, the model exhibited severe calibration underestimation (observed-to-expected ratio 2.763), demonstrating negative external validation driven by spectrum bias between community and hospital-based settings. Conclusion A parsimonious 9-predictor logistic regression model achieves acceptable discrimination and calibration across community cohorts for identifying prevalent self-reported physician-diagnosed CLD among community-dwelling postmenopausal women. However, direct transport to hospital-based settings is not supported without local recalibration because of pronounced spectrum bias.