Early Risk Stratification for Dengue Triage Using First-24-Hour Admission Data: A Multicentric Study

Dengue outbreaks place substantial pressure on healthcare systems because many clinically stable patients are admitted for precautionary monitoring. Early triage remains challenging because mortality is rare and early clinical features often overlap between survivors and non-survivors.

We conducted a retrospective multicentric study of 1,148 laboratory-confirmed dengue patients, including 1,127 survivors and 21 non-survivors, admitted in India between 2012 and 2017. Only variables recorded within the first 24 hours of admission were analyzed. After nested cross-validated LASSO-based feature selection, ten machine learning classifiers were evaluated under three scenarios: default-threshold classification, SMOTE-based resampling, and cost-sensitive threshold optimization designed to reduce unsafe low-risk classification while preserving real-world outcome prevalence.

Default-threshold models showed high overall accuracy but limited specificity for non-survivor identification. SMOTE improved minority-class recognition but altered classification behavior. Cost-sensitive threshold optimization improved balanced performance without changing the original class distribution. The threshold-optimized SVM model achieved specificity of 0.857, sensitivity of 0.914, balanced accuracy of 0.886, and correctly identified 18 of 21 non-survivors. Calibration analysis showed low Brier scores among top-performing models, and McNemar testing confirmed significant classification changes after threshold optimization for most classifiers.

A survival-centric, cost-sensitive machine learning framework using first-24-hour admission data improved clinically relevant dengue triage performance while preserving real-world outcome prevalence. These findings support threshold-aware machine learning as a practical decision-support approach for risk-sensitive dengue triage during outbreak conditions.