Machine Learning for Respiratory Health and Pediatric Asthma: A Dual Framework Combining Environmental Prediction of Respiratory Hospitalizations with Digital Biomarkers of Adherence to Diaphragmatic Breathing

Background: Asthma is a chronic respiratory disease shaped by environmental, meteorological, and behavioral factors. Few approaches combine population-level surveillance with individual-level monitoring within a single analytical framework. Methods: This work developed a dual machine learning framework. Study 1 modeled the daily count of respiratory admissions (chapter X of the ICD-10) in São Paulo, Brazil, from 2017 to 2022 as a nowcasting task, combining ElasticNet, residual CatBoost, direct CatBoost, and adaptive blending, validated by walk-forward over 30 bimonthly windows. Study 2 applied an XGBoost and Random Forest pipeline to 913 diaphragmatic-breathing sessions from 17 patients aged 9 to 16 years in the Respire Bem system, with Asthma Control Test and salivary cortisol represented using evidence-based synthetic simulation. Results: Study 1 achieved a mean MAE of 18.22, RMSE of 23.99, and R2 of 0.675, exceeding the seasonal baseline by 41.5%, with a significant advantage over all three baselines (Wilcoxon and Diebold–Mariano, p ≤ 0.038). Ablation showed each single-component configuration to be significantly worse than the full hybrid, but removing the environmental block cost only 0.27 admissions per day, an effect indistinguishable from zero. SHAP rankings were stable across windows (Kendall W = 0.640), led by NO2, PM2.5, and temperature. In Study 2 the pipeline ran end-to-end on real behavioral data, but because the outcomes were simulated, no predictive-accuracy metric is reported. Conclusions: Study 1 delivers a validated population-level nowcasting model whose accuracy rests mainly on the temporal structure of the series. Study 2 contributes a real behavioral dataset and a reproducible pipeline; the clinical validity of the digital biomarkers remains open and requires prospective work with directly measured outcomes.