Clinical knowledge-guided deep reinforcement learning for sepsis antibiotic dosing recommendations

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Pharmacology

Background:

• Sepsis is a leading global cause of death, with antibiotic resistance posing a significant treatment challenge.
• Current sepsis medication prediction models, often based on Markov decision processes, lack integration with clinical knowledge, leading to suboptimal treatment decisions.
• Developing effective antibiotic strategies is crucial for improving sepsis patient prognosis.

Purpose of the Study:

• To develop and evaluate a Deep Q-Network (DQN) based model, Sepsis Anti-infection DQN (SAI-DQN), for optimizing antibiotic selection and duration in sepsis treatment.
• To integrate clinical knowledge into the decision-making process of AI models for sepsis medication.
• To provide personalized antibiotic treatment recommendations that align with medical guidelines.

Main Methods:

• Utilized a Deep Q-Network (DQN) algorithm to create the Sepsis Anti-infection DQN (SAI-DQN) model.
• Incorporated sepsis clinical knowledge as reward functions to guide the DQN's decision-making process, ensuring adherence to medical guidelines.
• Trained and evaluated the model on patient data to assess its decision-making performance and treatment recommendation accuracy.

Main Results:

• The SAI-DQN model demonstrated a higher average decision-making value compared to existing clinical decisions.
• The model predicted a favorable prognosis for 79.07% of patients in the test set when using its recommended antibiotic combinations.
• Analysis of decision trajectories confirmed that the model's recommendations align with clinical practices and medical knowledge.

Conclusions:

• The SAI-DQN model offers a promising approach to personalized antibiotic treatment for sepsis, improving decision-making accuracy.
• By integrating clinical knowledge, the model provides medication recommendations that enhance patient outcomes and adhere to established medical guidelines.
• This AI-driven strategy has the potential to significantly improve the management of sepsis and combat antibiotic resistance.