Reinforcement Learning-Based Integrated Risk Aware Dynamic Treatment Strategy for Consumer-Centric Next-Gen Healthcare

Reinforcement learning (RL) has gained prominence in healthcare due to its ability to optimize treatment strategies without relying on predefined mathematical models. However, existing approaches face critical challenges: (1) the optimality of learned strategies is assessed without considering treatment risks, potentially leading to unsafe recommendations; (2) distribution shift issues cause learned strategies to diverge from physician decisions; and (3) past observational data and treatment history are often overlooked, leading to suboptimal state representations. To address these limitations, we propose a Dynamic Treatment Strategy Generation Model that integrates Dead Ends with an Offline Supervised Actor-Critic approach (DOSAC-DTR). Our model incorporates Dead Ends into the Actor-Critic framework to evaluate risks associated with recommended treatments. Additionally, physician oversight is embedded to mitigate distribution shift and align the learned strategy with expert decisions while maximizing expected outcomes. To enhance state representation, we employ an LSTM-based encoder-decoder model to capture essential patient history, ensuring robust decision-making. Experimental results on real-world datasets (MIMIC-III) demonstrate that DOSAC-DTR significantly reduces mortality rates (Sepsis: 3.51%, Ventilation: 13.74%) and improves treatment alignment with physicians (Jaccard similarity: 0.362, 0.126) compared to baseline models. These findings underscore the potential of reinforcement learning in personalized healthcare, improving both treatment efficacy and patient safety.