Infant mortality remains high and uneven in much of sub-Saharan Africa. Given finite resources, reducing premature mortality requires effective tools to identifying left- behind populations at greatest risk. While countries routinely use income- or poverty- based thresholds to target policies, we examine whether models that consider other factors can substantially improve our ability to target policies to higher-risk births. Using machine learning methods, and 25 commonly available variables that can be observed prior to birth, we construct child-level risk scores for births in 22 sub-Saharan African countries. We find that targeting based on poverty, proxied by income, is only slightly better than random targeting, with the poorest 10 percent of the population experiencing approximately 10 percent of total infant mortality burden. By contrast the 10 percent of the population at highest risk according to our model accounts for 15-30% of infants deaths, depending on country. A hypothetical intervention that can be administered to 10% of the population and prevents just 5% of the deaths that would otherwise occur, for example, would save roughly 841,000 lives if targeted to the poorest decile, but over 1.6 million if targeted using our approach.