Machine learning in population health: Opportunities and threats

Machine learning (ML) has succeeded in complex tasks by trading experts and programmers for data and nonparametric statistical models. However, the applications for which ML has been successfully deployed in health and biomedicine remain limited These limits also apply in population health, in which we are concerned with the health outcomes of a group of individuals and the distribution of outcomes within the group. In our metrics, we deal with messy global health data, and a large effort goes into piecing together sparse, noisy information to understand what causes how much health loss, where it occurs, and how it is changing. In our interventions, we often face stringent constraints on resources and need to develop appropriate and acceptable solutions under these constraints. How might ML-based approaches change population health? Here, we discuss opportunities and threats from ML, with our views on further development needed within ML to create the best possible outcomes.