Predição da necessidade de hospitalização em pacientes com Chikungunya utilizando técnicas de aprendizado de máquina no estado de Sergipe

Abstract

Chikungunya fever is a major public health concern in Brazil, characterized by high incidence rates and the potential for severe clinical outcomes, particularly in the Northeast region. This study aimed to develop predictive models based on machine learning techniques to estimate the need for hospitalization among patients with confirmed chikungunya infection in the state of Sergipe. Secondary data were obtained from the Notifiable Diseases Information System (SINAN), covering the period from 2020 to 2025. The study adopted an observational, retrospective, and quantitative design, analyzing sociodemographic characteristics, clinical manifestations, and pre-existing comorbidities. Given the pronounced class imbalance in the outcome variable, with only 5.1% of confirmed cases resulting in hospitalization, an oversampling technique was applied to mitigate bias during model training. Logistic Regression, Naive Bayes, Decision Tree, Bagging, and Random Forest models were evaluated. The results showed that data balancing substantially improved the identification of the minority class. Among the evaluated models, Logistic Regression and Naive Bayes achieved the best performance, reaching a sensitivity of 0.769 and areas under the receiver operating characteristic curve (AUC) of 0.790 and 0.808, respectively. Additionally, the interpretation of the odds ratios from Logistic Regression allowed the identification of the main factors associated with hospitalization, highlighting leukopenia (OR = 7.47), white race/color (OR = 4.91), black, brown, or indigenous race/color (OR = 3.49), autoimmune diseases (OR = 3.08), hematological diseases (OR = 2.28), as well as the clinical manifestations of vomiting and arthritis. The findings suggest that Logistic Regression is the most suitable approach for the analyzed context, as it combines satisfactory predictive performance with high interpretability. Consequently, it may support the early identification of patients at greater risk of hospitalization and contribute to strengthening epidemiological surveillance and public health decision-making.