Combining machine learning and mathematical models of disease dynamics to guide development of novel disease interventions

Abstract

The development of novel interventions against a disease entails optimising their specifications to achieve desired health goals such as disease reduction. As testing is limited early in development, it is difficult to predefine these optimal specifications, prioritize or continue investment in candidate interventions. Mathematical models of disease can provide quantitative evidence as they can simulate deployment and predict impact of a new intervention considering deployment, health-system, population and disease characteristics. However, due to large uncertainty early in development, as well as model complexity, testing all possible combinations of interventions and deployments becomes infeasible. As a result, mathematical models have been only marginally used during intervention development to date. Here, we present a new approach where machine learning enables the use of detailed disease models to identify optimal properties of candidate interventions to reach a desired health goal and guide development. We demonstrate the power of our approach by application to five novel malaria interventions under development. For various targeted reductions of malaria prevalence, we quantify and rank intervention characteristics which are key determinants of health impact. Furthermore, we identify minimal requirements and tradeoffs between operational factors, intervention efficacy and duration to achieve different levels of impact and show how these vary across disease transmission settings. When single interventions cannot achieve significant impact, our method allows finding optimal combinations of interventions fulfilling the desired health goals. By enabling efficient use of disease models, our approach supports decision-making and resource investment in the development of new interventions for infectious diseases.