小児喘息管理の改善に向けたデータマイニングの活用：影響因子に関する研究

To identify the most influential features of asthma control in children and to assess whether feature selection improves model performance.

Records of 890 patients (<18 years) from the Immunology, Asthma, and Allergy Research Institute, Children's Medical Center, Tehran (2013-2018) were analyzed. The binary outcome was asthma control (controlled vs uncontrolled). Eighty-three candidate features (demographics, comorbidities, history, triggers) were considered. Thirteen FS methods (filters, wrappers, embedded methods, plus PCA) were compared. For each method, the top 20 features were used to train an XGBoost classifier; the full 83-feature model served as the baseline. Performance was estimated with repeated holdout cross-validation (10 repeats, 90/10 splits) and summarized using Accuracy, Precision, Recall, F1, ROC-AUC, and PR-AUC.

Relative to the all-features baseline model, the model trained on the SVM-selected subset showed consistent gains in key metrics: Recall increased from 86.12% to 90.32%, F1 from 77.37% to 82.84%, PR-AUC from 77.37% to 82.84%, and ROC-AUC from 52.18% to 64.21%. Features with high consensus across methods were primarily related to modifiable triggers (e.g. smoke exposure and climate-related factors) and medical history (e.g. allergic rhinitis and eczema).

Applying feature selection generally improved performance across multiple metrics, yielding compact, stable subsets that highlight modifiable factors - particularly triggers and allergic history - and support clinical interpretability. These findings are associative; therefore, external validation is recommended.