Stacking ensemble framework for early warning of acute respiratory infections: application to the Arequipa region, Perú (2000–2024)

Abstract

Early detection of acute respiratory infections is critical for timely public health interventions, especially in vulnerable populations. This study presents a hierarchical ensemble framework with stacking applied to Peru's National Epidemiological Surveillance System (RENACE), covering data from 2000 to 2024 for the Arequipa region (130,970 records). The framework integrates six diverse base models—XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, and Ridge Regression—combined using RidgeCV meta-learning to predict six simultaneous targets: pneumonia cases, hospitalizations, and deaths for children under 5 and adults over 60. Using comprehensive spatiotemporal feature engineering (more than 80 features including lags, moving statistics, seasonal patterns, and geographic aggregations), the stacking ensemble achieved exceptional performance with R2=0.9957, MAE=0.0005, and RMSE=0.0082, outperforming all individual models. Notably, Ridge regression achieved R2=0.9999, indicating an almost perfect fit to the aggregated departmental data. The proposed system demonstrates strong potential as an operational early warning tool for resource allocation and epidemic preparedness in developing countries with limited surveillance infrastructure.