Development and validation of a Prediction Model for Hospitalization in Dengue Fever: A Machine Learning approach using Brazilian National Surveillance Data

Dengue remains one of the biggest public health challenges in Brazil, with millions of cases annually and high pressure on the hospital system. Although most infections are mild, progression to severe forms can occur suddenly and unclear signs at the onset challenges the identification of patients who need hospitalization. This study aims to predict hospitalization need of dengue-confirmed cases using clinical characteristics to assist decision. Data from 2019-2023 was retrieved from the Brazilian Notifiable Diseases Information System. Four prediction models were developed (Logistic Regression, Decision Trees, CatBoost, XGBoost) and evaluated using area under the ROC (AUC) curve and the calibration plots. A temporal validation was performed on 2024 dataset. Out of the 3.775.896 cases reported, 191,518 (5%) were hospitalized. XGBoost model without a data balancing showed the best performance (AUC[95%CI]: 76.00 [75.67-76.35]) and adequate calibration of risks. Performance was similar in temporal validation, varied among age, dengue types and region.