Estimating measles reintroduction risk in São Paulo using machine learning.

OBJECTIVES

This study aimed to develop a measles risk estimation tool to predict municipal-level occurrences in Brazil's most populous state, São Paulo.

STUDY

DESIGN

Measles risk for the 645 municipalities of the State of São Paulo was estimated by using public data spanning 2007-2023, integrated via the use and selection of machine learning models.

METHODS

Five machine-learning prediction models were trained via K-fold stratified cross validation using publicly available demographic, mobility, socioeconomic, and vaccination data to predict annual measles occurrence per municipality. Model-specific decision thresholds to classify municipalities into high/low risk were obtained from the Receiver Operating Characteristic (ROC) curve evaluated on validation data. Shapley Additive Explanations (SHAP) on test data were used to identify the most influential features.

RESULTS

All models showed similar performances across metrics like AUC-ROC, F1-Score and Brier Score on training and test data. The highest measles risk was consistently identified in the São Paulo municipality metropolitan region (state capital) and in Santos (largest South American port). Using the decision threshold for the Random Forest model, a total of 2 municipalities were highlighted as high risk. Passenger arrivals at international airports, WHO-reported cases along with municipal population size and population density were the most influential features, consistent with the current non-endemic measles scenario in São Paulo State.

CONCLUSIONS

A measles risk-estimation tool is proposed for estimating measles risk in São Paulo State. Because it relies on publicly available data, the proposed framework may serve as a basis for similar initiatives in other Brazilian states or regions.