Abstract: Cardiogenic shock is a deadly and complicated illness. Despite extensive research into treating cardiogenic shock, mortality remains high and has not decreased over time. Patients suffering from cardiogenic shock are highly heterogeneous, and developing an understanding of phenotypes among these patients is crucial for understanding this disease and the appropriate treatments for individual patients. In this work, we develop a deep mixture of experts approach to jointly find phenotypes among patients with cardiogenic shock while simultaneously estimating their risk of in-hospital mortality. Although trained with information regarding treatment and outcomes, after training, the proposed model is decomposable into a network that clusters patients into phenotypes from information available prior to treatment. This model is validated on a synthetic dataset and then applied to a cohort of 28,304 patients with cardiogenic shock. The full model predicts in-hospital mortality on this cohort with an AUROC of 0.85 ± 0.01. The model discovers five phenotypes among the population, finding statistically different mortality rates among them and among treatment choices within those groups. This approach allows for grouping patients in clinical clusters with different rates of device utilization and different risk of mortality. This approach is suitable for jointly finding phenotypes within a clinical population and in modeling risk among that population.