A medical policy can help personalize treatment recommendations based on patients' characteristics. Classical definitions of such policies are often grounded in a causal framework involving a single clinical outcome. In the common setting where several outcomes must be considered simultaneously, these policies typically neglect the risk of adverse events. I will present PLUC (Policy Learning Under Constraint), a framework for defining and learning policies that explicitly incorporates one or more constraints.

This work is the result of a collaboration with Laura Fuentes-Vicente, Mathieu Even, Gaëlle Dormion, and Julie Josse.