Classifying Physician-Diagnosed Childhood Asthma Status Using Explainable Machine Learning: NHANES 2007-2016

To develop and internally evaluate explainable machine-learning classifiers for physician-diagnosed childhood asthma status and assess the incremental value of five complete-blood-count (CBC)-derived inflammatory indices beyond gender, race, age, family history of asthma, poverty income ratio (PIR), and body mass index (BMI).

NHANES 2007-2016 data from 6,437 children aged 2-14 years were split into training (n = 4,505) and held-out test (n = 1,932) sets. Six machine-learning algorithms were compared using PR-AUC as the primary selection metric. Calibration and the classification threshold were derived from training out-of-fold predictions. The final model was evaluated in the held-out test set. KernelSHAP assessed feature contributions, and repeated nested cross-validation evaluated the incremental value of the five CBC-derived indices.

Among 6,437 children, 784 (12.2%) had physician-diagnosed asthma. SVM achieved the highest mean PR-AUC (0.294) during training-set repeated cross-validation and ranked first in all 10 repeats. In the held-out test set, ROC-AUC was 0.743 (95% CI, 0.708-0.778) and PR-AUC was 0.312 (95% CI, 0.268-0.370). At the training-derived threshold of 0.081, sensitivity was 0.672, specificity 0.740, and Brier score 0.097. Family history of asthma had the largest mean absolute SHAP value (0.089). Adding the five CBC-derived indices changed mean PR-AUC from 0.2944 to 0.2956 (paired difference +0.0012; SD 0.0049).

The SVM showed moderate discrimination for physician-diagnosed childhood asthma status. CBC-derived inflammatory indices provided no clear incremental improvement beyond gender, race, age, family history, PIR, and BMI. External validation is required before clinical use.