Reinforcement Learning for Optimizing Laboratory Test Utilization in Intensive Care Units: A Framework for Reducing Unnecessary Testing

Abstract

Unnecessary laboratory testing in intensive care units (ICUs) has long been identified as a major contributor to low efficiency, increased costs, and heightened medical risks. Laboratory tests are very critical in monitoring and managing patient health; however, a high percentage of these tests have very limited clinical value. This underscores the urgent need for optimization strategies for lab test utilization without compromising clinical outcomes. Most current decision-making processes regarding laboratory test orders still require dynamic adaptation to patient conditions changes. This study proposes a novel framework leveraging reinforcement learning (RL) that makes high-quality decisions about ordering a laboratory test. By learning from historical patient data, our RL model aims to recommend whether a lab test is necessary at any given time, balancing patient needs with resource efficiency. Using the MIMIC-IV dataset, the model captures temporal patterns in patient health trajectories, considering test timing and clinical event irregularities. Herein, our approach tackles two significant challenges in irregularities in the ICU data and dynamic decision-making by incorporating reward systems prioritizing safety and cost reduction. By implementing our RL framework, we could reduce the number of unnecessary tests without affecting clinical outcomes.