Aortic valve replacement (AVR) involves multiple device types that fall into two broad categories: bioprosthetic valves and mechanical valves. Previous studies for predicting mortality after AVR have only focused on comparing these two broad categories, ignoring variations in multiple device types. Due to the unordered nature of multiple treatments as well as rare outcomes, developing a flexible prediction function to anticipate mortality after AVR is a statistical challenge. This work develops statistical machine learning techniques to account for both of these features in order to build prediction functions for 30-day mortality and 1-year mortality. We consider patients in 2002-2012 who are Massachusetts residents with isolated procedures as well as combined cohorts, and use over 100 additional covariates to build the prediction function, including demographic information, comorbidities, and medications.