Individualizing mechanical ventilation treatment regimes remains a challenge in the intensive care unit (ICU). Reinforcement Learning (RL) offers the potential to improve patient outcomes and reduce mortality risk, by optimizing ventilation treatment regimes. We focus on the Offl
...
Individualizing mechanical ventilation treatment regimes remains a challenge in the intensive care unit (ICU). Reinforcement Learning (RL) offers the potential to improve patient outcomes and reduce mortality risk, by optimizing ventilation treatment regimes. We focus on the Offline RL setting, using Offline Policy Evaluation (OPE), specifically importance sampling (IS), to evaluate policies learned from observational data. Using a running example, we illustrate how a large difference between the learned policy and actual clinical behavior (behavior policy) limits the reliability of IS-based OPE. To assess this reliability, we use the Effective Sample Size (ESS) as a diagnostic. To achieve reliable evaluation, we apply policy shaping, by incorporating a divergence constraint in the policy learning objective, aiming to reduce the difference between the evaluation and behavior policy. We consider both a Kullback-Leibler (KL) divergence constraint and introduce a new constraint, the ESS divergence. Since effective OPE relies on an accurate estimate of the true behavior policy, we address how such an estimate is acquired. Various classifiers for estimating the behavior policy are systematically evaluated, focusing on both discrimination and calibration performance. Empirical results show the difficulty of learning policies that outperform existing clinical practices and generalize well to unseen patients. Although policy shaping improves the reliability of policy evaluations, no policies that consistently outperform clinician practice were found. The KL divergence constraint generalized better to unseen patients than the ESS divergence, which achieved large ESS without actually reducing the difference between the evaluation and behavior policy. We underscore the necessity of a cautious approach to applying RL in healthcare, and advocate that assessing OPE reliability and behavior policy calibration becomes standard practice, to ensure that only effective and reliable RL policies are considered for real-world clinical trials.
