Many patients in critical care are at high risk of acute kidney injury. Renal replacement therapy can be life-saving, but also puts patients at risk, besides imposing a high financial and logistical burden to critical care units. Motivated by this, we studied the question of when to start renal replacement therapy, in a close collaboration with intensive care clinicians and nephrologists. In this talk, I will reflect on our experience with regard to learning an optimal treatment policy based on the Ghent University Intensive Care database. This has motivated methodological work on variable importance, which I will mention briefly, and on causal prediction, which I will discuss in detail. Orthogonal meta-learners, such as DR-learner, R-learner and IF-learner, are increasingly used to predict treatment effects (conditional on patient characteristics). These improve convergence rates relative to naive meta-learners through de-biasing procedures that involve applying standard learners to specifically transformed outcome data. However, these transformations lead them to disregard the possibly constrained outcome space, which can be particularly problematic for dichotomous outcomes: these typically get transformed to values that are no longer constrained to the unit interval, making it difficult for standard learners to guarantee predictions within the unit interval. To address this, I will show how to construct Neyman-orthogonal meta-learners for the prediction of counterfactual outcomes which respect the outcome space. As such, the obtained i-learner or imputation-learner is more generally expected to outperform existing learners, even when the outcome is unconstrained, as we confirm empirically in simulation studies and illustrate in an analysis of critical care data.