Early-Life Allergen Sensitization Phenotypes and Exploratory Machine-Learning Interpretation of School-Age Asthma...

Early-life allergen sensitization is associated with childhood asthma, but sensitization patterns are heterogeneous. This study aimed to characterize early-life sensitization phenotypes and evaluate their associations with school-age asthma, with exploratory machine-learning analyses examining model-derived allergen contributions.

Electronic medical records from Renmin Hospital of Wuhan University (2014-2022) were used. Two samples were derived: a phenotype discovery sample (Sample A, n = 4989) and a nested case-control sample (Sample B, n = 435; 84 cases and 351 controls). In Sample A, latent class analysis (LCA) identified population-level sensitization phenotypes. In Sample B, pre-specified multivariable logistic regression models adjusted for age and sex evaluated associations of LCA-derived phenotypes and individual inhalant allergen sensitization with school-age asthma. Random Forest and XGBoost/SHAP were used as exploratory analyses to assess allergen feature importance and model-derived contributions to predicted asthma probability.

LCA identified four sensitization phenotypes: predominantly non-sensitized, food-dominant, inhalant-dominant, and pet dander-predominant. After adjustment for age and sex, The inhalant-dominant phenotype was associated with higher odds of school-age asthma (adjusted OR = 3.14, 95% CI: 1.43-6.91), whereas the pet dander-predominant phenotype showed a positive but non-significant association (adjusted OR = 1.75, 95% CI: 0.94-3.26). In the pre-specified multivariable allergen model, dust mite sensitization remained significantly associated with asthma (adjusted OR = 3.30, 95% CI: 1.94-5.62). Exploratory analyses identified consistent model-derived contribution patterns for dust mite, whereas dog dander sensitization showed more variable patterns. XGBoost AUC decreased from an apparent AUC of 0.859 to a mean internally validated AUC of 0.636.

Early-life sensitization showed distinct population-level phenotypes, with the inhalant-dominant phenotype associated with school-age asthma. Dust mite showed the most consistent signal across association and exploratory prediction analyses. Dog dander findings were model-dependent and should be interpreted as exploratory prediction patterns. SHAP findings require external validation before clinical application.