LAUSR

Objective. Sepsis seriously threatens human life. Early identification of a patient’s risk status and appropriate treatment can reduce septic shock risk and mortality. Our purpose is to design and validate an adjunctive therapy system based on deep reinforcement learning (DRL), which can provide treatment recommendations with providence and assess the patient’s risk status and treatment options in the early stages. Approach. Data is from the Beth Israel Deaconess Medical Center. The raw data included 53 423 patients from MIMIC-III. Of these, 19 620 eligible samples were screened to form the final cohort. First, the patient’s physiological parameters were fed into the DRL therapy strategy recommendation module (TSRM), which provides a forward-looking recommendation for treatment strategy. The recommended strategies were then fed into the reinforcement learning risk assessment module (RAM), which predicts the patient’s risk status and treatment strategy from a long-term perspective. The DRL model designed in this paper assists in formulating treatment plans and evaluating treatment risks and patient status through continuous interaction with patient trajectory; this model therefore has the foresight that a supervising deep learning model does not. Main results. The experiment shows that, in the test set for the TSRM, mortality is the lowest when the treatment strategy that is actually implemented is the same as the AI-recommended strategy. Regarding the RAM, it can accurately grasp a patient’s deterioration trend, and can reasonably assess a patient’s risk status and treatment plans at an early stage. The assessment results of the model were matched with the actual clinical records. Significance. A DRL-based sepsis adjunctive therapy model is proposed. It can prospectively assist physicians in proposing treatment strategies, assess the patient’s risk status and treatment methods early on, and detect deterioration trends in advance.