Author(s) :

Gautama Narula.

Volume/Issue :

Abstract :

Background: This study evaluates a supervised classification approach for estimating short-horizon dengue outbreak risk using weekly country-level case-count data. Methods: Weekly country-level dengue case counts were extracted from the OpenDengue national dataset and merged with countrylevel demographic, health-system, and Water, Sanitation and Hygiene (WASH) covariates. A locationweek was labeled positive when the following four weeks met a rule-based elevated-risk definition: forward incidence of at least 100 cases per 100,000 population, or forward cases at least 50% above the prior four-week baseline with at least 20 absolute cases. All predictors were restricted to current or prior information. A temporal split was used: training observations occurred before January 1, 2020; validation observations occurred during 2020; and test observations occurred from January 1, 2021 through the final available observation. RandomForestClassifier and GradientBoostingClassifier models were implemented in scikit-learn and calibrated with Platt scaling on the validation set. Results: After preprocessing, the modeling dataset contained 19,456 country-week observations from 47 countries covering 1924-04-06 through 2022-12-25. The final model was RandomForestClassifier. On the held-out temporal test set (n=3,703; event rate 14.0%), the model achieved AUROC 0.860 (95% CI 0.844-0.875), AUPRC 0.549 (95% CI 0.511-0.590), Brier score 0.090, ECE 0.038, and precision@top-10% 0.589. Conclusions: This study reports a 4-week dengue outbreak-risk classification analysis using temporally separated evaluation and calibrated probabilities. The findings support the feasibility of a leakage-aware, calibrated 4-week probability-estimation pipeline for dengue surveillance, with a 1-100 display scale intended only for prioritization. The study does not claim validation for COVID-19 or influenza without appropriate disease-specific outcome data.