Learning optimal treatment strategies for intraoperative hypotension using deep reinforcement learning

Area of Science:

• Anesthesiology and Perioperative Medicine
• Artificial Intelligence in Healthcare
• Nephrology and Critical Care

Background:

• Traditional surgical decision-making relies on variable human experience.
• Suboptimal management of intraoperative hypotension can lead to acute kidney injury (AKI).
• A data-driven system can enhance perioperative decision-making.

Purpose of the Study:

• To develop a Reinforcement Learning (RL) model for optimizing intravenous (IV) fluid and vasopressor dosage during surgery.
• The model aims to prevent intraoperative hypotension and subsequent postoperative AKI.

Main Methods:

• Retrospective analysis of 50,021 surgeries from 42,547 adult patients.
• Development of a Deep Q-Networks based RL model for treatment recommendations.
• Model trained and validated on 34,186 surgeries, tested on 15,835 surgeries.

Main Results:

• The RL model demonstrated agreement with physician decisions for vasopressor dosage in 69% of cases.
• Model recommendations for IV fluids were within 0.05 ml/kg/15 min in 41% of cases.
• The RL policy yielded a higher estimated value than physician treatments, and AKI prevalence was lowest with model-aligned dosages.

Conclusions:

• The developed RL model shows potential for reducing postoperative AKI.
• Implementation of the model's policy may improve outcomes related to intraoperative hypotension.
• Data-driven perioperative management can enhance patient safety and reduce complications.