Abstract

In healthcare, the dynamic nature of clinical processes, coupled with the prevalence of complex diseases and evolving patient conditions, necessitates adaptable and personalized treatment approaches. While existing treatment recommendation systems rely heavily on rule-based protocols derived from clinical guidelines, these may overlook the nuances of individual patient cases, particularly in intensive care units (ICUs) where deviations from standard protocols could be beneficial. However, accessing reliable evidence, such as randomized controlled trials (RCTs), for ICU conditions can be challenging due to various factors including patient eligibility and limited positive findings from RCTs. In such contexts, leveraging large observational datasets and applying artificial intelligence (AI) and machine learning techniques, particularly reinforcement learning (RL), presents a promising avenue for aiding clinical decisions. RL algorithms aim to train agents to maximize cumulative rewards by learning optimal actions based on patient states and trajectories. Unlike traditional clinical protocols, RL policies offer more personalized approaches, capturing individual patient details. Multi-objective reinforcement learning further enhances decision-making by considering multiple objectives, such as cost and optimal path, simultaneously. By mapping state-action pairs to vector rewards, RL algorithms can effectively handle complex decision spaces and facilitate the selection of optimal actions.