Developing a prediction model for persistent airflow limitation in asthmatic children.

Background: A small proportion of asthmatic children will develop persistent airflow limitation (PAL). The purpose of this study was to develop predictive models using classical and machine learning methods to identify asthmatic children at risk of PAL. Methods: A total of 1,671 asthmatic children were enrolled between January 1, 2019, and December 31, 2020, to serve as training and internal validation sets. Temporal validation included 401 patients from January 1, 2021, to December 31, 2021. PAL was determined in the third year after enrollment, defined as a fixed forced expiratory volume in 1 second (FEV1)/forced vital capacity (FVC) ratio below 0.75. Predictors included demographic and clinical data. Machine learning algorithms, including random forest (RF) and extreme gradient boost (XGBoost), along with the classical logistic regression (LR) methods, were utilized to develop prediction models. Discrimination ability evaluation was conducted using the area under the curve (AUC), accuracy, sensitivity, and specificity, while fitness estimation utilized calibration curves and Brier scores. Additionally, decision curve analysis was employed for clinical value evaluation. Results: In the internal validation, the RF model achieved an AUC of 0.857 (95% CI: 0.791-0.924), followed by LR with an AUC of 0.849 (95% CI: 0.780-0.908) and XGBoost with an AUC of 0.835 (95% CI: 0.761-0.909). In the temporal validation, the three prediction models exhibited similar performance. Specifically, RF attained an AUC of 0.853 (95% CI: 0.771-0.935), LR achieved an AUC of 0.836 (95% CI: 0.742-0.938), and XGBoost reached an AUC of 0.848 (95% CI: 0.757-0.940). The calibration curve and low Brier score indicated good fitness of all prediction models, and decision curve analysis revealed desirable net benefits for all prediction models in both internal and temporal validation. Conclusions: PAL in asthmatic children can be predicted with clinically meaningful accuracy using routinely available clinical data, and three prediction models (LR, RF, and XGBoost) demonstrated comparable performance in identifying high-risk patients.