Abstract
Laboratory testing is an integral tool in the management of patient care inhospitals, particularly in intensive care units (ICUs). There exists aninherent trade-off in the selection and timing of lab tests betweenconsiderations of the expected utility in clinical decision-making of a giventest at a specific time, and the associated cost or risk it poses to thepatient. In this work, we introduce a framework that learns policies forordering lab tests which optimizes for this trade-off. Our approach uses batchoff-policy reinforcement learning with a composite reward function based onclinical imperatives, applied to data that include examples of cliniciansordering labs for patients. To this end, we develop and extend principles ofPareto optimality to improve the selection of actions based on multiple rewardfunction components while respecting typical procedural considerations andprioritization of clinical goals in the ICU. Our experiments show that we canestimate a policy that reduces the frequency of lab tests and optimizes timingto minimize information redundancy. We also find that the estimated policiestypically suggest ordering lab tests well ahead of critical onsets--such asmechanical ventilation or dialysis--that depend on the lab results. We evaluateour approach by quantifying how these policies may initiate earlier onset oftreatment.