Healthcare utilization and associated costs have increased rapidly in recent years, making the study of healthcare expenditures an important area of public health research. Analysis of healthcare expenditure data is challenging due to heavily skewed distributions and zero-inflation. Myriad methods have been developed for analyzing cost data; however, a priori determination of an appropriate method is often difficult. Super-learning, a technique that considers an ensemble of methods for cost estimation, provides an interesting alternative for modeling healthcare expenditures. The super learner has demonstrated benefits over a single method in recent studies across many disciplines. In this work, we propose a two-stage super learner specifically designed for predicting zero-inflated expenditures. We demonstrate that the two-stage super learner has strong performance in predicting healthcare costs across a variety of cost distributions, in both real and simulated data.