A simple feed forward neural network to predict the 2025 outbreak of measles in the USA.

BackgroundMeasles is a highly contagious viral disease associated with a variety of severe complications. Since 1963, widespread usage of a highly effective vaccine has made measles a largely preventable disease. However, recent rises in vaccine hesitancy in the United States has seen increasing incidence of measles cases, including an ongoing (22nd March 2025) outbreak originating in Gaines County, Texas. Accurate modelling is essential for informing public the health response to measles outbreaks. Current strategies are limited by a variety of challenges primarily associated with obtaining adequate real-time data in conjunction with an active outbreak. MethodsHere, we use a dynamic time warping (DTW) approach, to facilitate the use of historical outbreak data to model the ongoing measles outbreak in the United States. Data from various historical outbreaks were used as features for two neural network architectures: a feedforward neural network (FNN) and a novel, biologically informed neural network (BINN). The latter BINN architecture incorporated dynamics from populations of susceptible (S), infected (I) or recovered (R) individuals, commonly described in association with SIR infectious disease models. FindingsBoth FNN and BINN architectures performed well across 34-weeks of testing data, predicting measles cases with a mean squared error of less than 2 (1.1060 and 1.1451, respectively). Additionally, a 5-week forward prediction of case numbers closely matches CDC estimates, as reported on 22nd of March 2025. Interestingly, no difference in the accuracy of forward predicted case numbers was observed between FNN and BINN architectures. InterpretationOverall, this study highlights the value of historical data in combination with relatively simple FNN architecture for accurately modelling ongoing or emerging measles outbreaks. Such modelling strategies represent an essential tool for future outbreak management and a timely reminder of the importance of consistent public health responses (vaccination) to otherwise preventable diseases. FundingKRS is supported by an NHMRC Investigator Grant (2007919) Research in contextWe searched for the terms "MACHINE LEARNING" AND "MEASLES" AND "OUTBREAK" in Google Scholar. Numerous articles have reported the use of machine learning for the prediction of measles outbreaks. However, no study to date has used dynamic time warping and historical data for feature selection and training data in order to develop a simple feed-forward neural network. Added value of this studyHere, we develop a new, rapid and highly accurate methodology for predicting measles outbreaks in the USA in real time. We use this model to show that number of future measles cases will likely increase over time, highlighting the importance of continued and consistent management (vaccination) of otherwise preventable diseases. ImplicationsThe present study provides a new approach to infectious disease modelling that circumvents the need for diverse and granular datasets for feature selection. We use this novel approach to show that without a dramatic shift in public health policy and vaccine advocacy the number of measles cases in the USA will continue to rise.