An important problem in healthcare is to describe human decision-making based on observational data. A description of the underlying policy can be used to, for example, detect anomalous behavior, compare different strategies, and develop new guidelines. State-of-the-art techniques for uncovering a policy from demonstrated behavior, for example inverse reinforcement learning and imitation learning, often fall short of interpretability and therefore have limited use when the purpose is to describe the policy, not to implement it. In this project, we will study interpretable representations of policies for sequential decision-making, which is common in healthcare. First, we will evaluate baseline methods adapted for single-stage decisions. Next, we will consider sequential models such as recurrent neural networks and recurrent decision tress. Finally, we aim to develop new algorithms for learning interpretable policies.