Explainable Machine Learning to Identify Clinical, Environmental, and Social Factors Associated with Diagnosed Pediatric Asthma.

BACKGROUND

Pediatric asthma reflects clinical, physiologic, environmental, and social determinants, yet most classification models use clinical variables alone and are difficult to interpret.

OBJECTIVE

To use explainable machine learning on nationally representative data to identify and rank factors associated with a reported physician diagnosis of asthma in U.S. children, and to test whether fewer higher-ranked features retain full performance.

METHODS

We analyzed 6,567 children aged 6 to 17 years in NHANES 2007-2012 (weighted asthma prevalence 18.8%). Of 477 harmonized variables, 22 entered the primary CatBoost model after pre-specified eligibility screening, feature engineering, and correlation pruning. Data were split into training, validation, and held-out internal test sets (60/20/20), with the operating threshold locked on validation. SHAP ranked features; 95% confidence intervals used stratified bootstrap.

RESULTS

The primary model reached a test-set AUC of 0.78 (95% CI 0.74-0.81), sensitivity 0.75, specificity 0.63, and negative predictive value 0.92. The highest-ranked features included recent wheezing, family history of asthma, general health status, serum cotinine, household size, FEV₁/FVC ratio, race/Hispanic origin, insurance coverage, and U.S. birth. A reduced model of these ten plus two spirometry-availability indicators had more favorable point estimates on every metric (AUC 0.80; paired difference 0.023, 95% CI 0.007-0.041; sensitivity 0.79; specificity 0.67; negative predictive value 0.93).

CONCLUSION

Explainable machine learning recovers established clinical signals and ranks social and contextual factors associated with a reported pediatric asthma diagnosis; a parsimonious feature set retains this performance. Given the cross-sectional design and internal evaluation, external and prospective validation is required before clinical use.