Prediction of bronchial hyperresponsiveness from routine pulmonary function and inflammatory markers: an interpretable model and scorecard

Bronchial provocation testing (BPT) is a valuable method for assessing bronchial hyperresponsiveness (BHR), but its use may be limited in busy outpatient settings. We aimed to develop and internally validate an interpretable model for predicting BHR from routine pulmonary function and blood inflammatory markers, and to derive a simplified scorecard for pretest risk stratification. We retrospectively included 3,993 outpatients with asthma-suggestive symptoms who underwent BPT, all with a post-bronchodilator FEV \({}_1\) /FVC in the normal range, no substantial smoking exposure and HRCT findings that were normal or consistent with asthma, split into training and test sets (8:2). Candidate predictors included demographics, blood count markers, and spirometric indices. Logistic regression, LASSO, random forest, XGBoost, LightGBM, a spline logistic regression, and two explainable boosting machine (EBM) models were developed, with feature selection, hyperparameter tuning, and threshold selection performed in the training set. Discrimination, calibration, and decision curve analysis were assessed in a randomly held-out internal test set, and a simplified scorecard was derived from the main-effect EBM model. Overall, 718 (18.0%) participants had BHR. In the test set, the two EBM models showed discrimination similar to gradient-boosting and random-forest models (AUC \(≈\) 0.81), with numerically more favorable calibration (EBM_main expected calibration error 0.0233). Differences in calibration among the leading models were small. Higher eosinophil count and lower MEF50%, FEV \({}_1\) /FVC, and MMEF% contributed most to predicted risk. The simplified scorecard retained most of the predictive performance (AUC 0.7964, AUPRC 0.4797). In this selected pretest outpatient population, routine clinical variables predicted BHR with good discrimination and with calibration comparable to that of the best alternative models. The derived scorecard may support risk stratification before BPT. External validation is required before clinical deployment. Not applicable.