Brain-inspired offline reinforcement learning in sepsis prognosis

Sepsis remains one of the leading causes of mortality in intensive care units, where its rapid and unpredictable progression often complicates timely clinical intervention. Existing machine learning and deep learning approaches for sepsis prediction often fail to capture the dynamic evolution of patient conditions. We develop a Brain-Inspired Reinforcement Learning algorithm with Stochastic Policy (BIRLSP) for sepsis prognosis from fixed-batch data. The stochastic policy mechanism allows adaptive exploration during offline learning, reflecting principles of brain decision-making and enhancing the balance between exploration and exploitation. In addition, a dynamic–static neural network architecture is introduced to mimic information processing across distinct brain regions. To accurately predict sepsis risk, we model the action as four stratified sepsis risk levels, which is determined based on the SOFA score, low-risk, medium-risk, high-risk and severe-risk. When evaluated against machine learning and RL approaches, BIRLSP achieved 74.18% accuracy (precision 74.09%, recall 74.07%, F1 score 73.96%, and AUC 82.73%). These findings suggest that BIRLSP offers a promising offline deep reinforcement learning (DRL) framework to assist clinicians in sepsis prognosis and support early, informed risk assessment in critical care settings.