A Machine Learning Approach to Defining and Predicting the Scale of Typhoid Fever Outbreaks

Abstract Despite improvements in access to clean water and sanitation, typhoid fever outbreaks continue to cause substantial morbidity and mortality worldwide. The World Health Organization (WHO) recommends the use of typhoid conjugate vaccines (TCVs) for outbreak response; however, guidance on when to trigger such a response remains unclear due to the absence of a standardized definition of a typhoid fever outbreak. We analyzed a dataset of 34 typhoid fever outbreaks with detailed time series data between 2000 and 2022 and found that all documented outbreaks lasted at least 7 days and were associated with at least 6 total cases or at least 2 laboratory-confirmed cases. Applying unsupervised machine learning methods to this dataset revealed two distinct clusters: one containing relatively large outbreaks (≥288 total cases), and the other containing relatively small outbreaks (≤191 total cases). We then labeled 215 typhoid outbreaks from the same period as either large or small using a simple threshold of 250 total cases and trained supervised machine learning models to predict the outbreak size classification on the basis of country-level features in the year before the outbreak. We found that large outbreaks are most likely to occur in countries where access to safely managed sanitation services, the proportion of urban dwellers, and basic drinking water services are low. We then used these models to project the magnitude of typhoid outbreaks for 192 countries in 2023. We identified several high-risk areas in Africa and South Asia that were consistently predicted to face large-scale outbreaks across most of the models. These findings provide evidence-based criteria for defining the onset of an outbreak, enabling the triggering of timely responses to typhoid outbreaks and preparation for potential large-scale epidemics.