Internally validated machine learning models identify asthma and its phenotypes using multicenter real-world data

Asthma exhibits significant clinical heterogeneity, necessitating precise phenotyping for personalized treatment. This study aims to develop and internally validate machine learning models for cross-sectional identification and phenotyping of asthma using routinely collected clinical data. Data from the MIMIC-IV and OPCRD databases were integrated, encompassing 3,025 eligible cases (2,180 in asthma group and 845 in non-asthma group). Thirty baseline variables were extracted, including demographic characteristics, clinical symptoms, biomarkers, environmental exposures, and genetic history. The dataset was split 7:3 into training (2,118 cases) and test (907 cases) sets. LASSO regression (with L1 regularization) and random forest (RF) models were developed. Performance was evaluated using the area under the ROC curve (AUC), and feature importance was assessed by the mean decrease in the Gini index. In the test set, LASSO models yielded numerically higher AUCs than RF across all tasks. LASSO achieved AUCs of 0.836 for overall asthma, 0.782 for type 2 asthma, 0.751 for allergic asthma, and 0.723 for early‑onset asthma. DeLong's test showed statistically significant AUC differences for type 2 asthma (P < 0.001), allergic asthma (P = 0.045), and early-onset asthma (P = 0.026), whereas the difference for overall asthma was not significant (P = 0.101). At λ.1se, the overall asthma model retained 9 predictors, with history of allergies, family history of asthma, and wheezing frequency as the strongest. Type 2 asthma identification relied on bronchodilator reversibility, family history of asthma, and wheezing frequency after excluding defining biomarkers, while early‑onset asthma was characterized solely by negative associations with allergy history, current age, and FeNO. RF analysis revealed that age, FeNO, IgE, and pulmonary function indices were the dominant features for early‑onset asthma. LASSO regression demonstrates favorable performance and interpretability for asthma phenotyping, with statistically significant advantages over RF for type 2 and early‑onset asthma. Routine clinical variables effectively identify asthma and its subtypes, supporting potential implementation in electronic medical record settings; however, further external validation is required.