Personalized Hemodynamic Management Using Reinforcement Learning to Prevent Persistent Acute Kidney Injury After...

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Nephrology

Background:

• Acute kidney injury (AKI) is a significant complication post-cardiac surgery, with persistent AKI (pAKI) leading to worse outcomes.
• Current hemodynamic management relies on bundled interventions with inconsistent implementation, highlighting the need for personalized approaches.

Purpose of the Study:

• To develop and validate a reinforcement learning (RL) model for individualized dosing of IV fluids, vasopressors, and inotropes.
• The goal is to prevent persistent AKI (pAKI) in patients undergoing cardiac surgery.

Main Methods:

• Retrospective cohort study involving model development and validation in MIMIC-IV, SICdb, and Mount Sinai Health System databases.
• Utilized RL to optimize treatment policies based on reward-based learning.
• Assessed model performance against clinician management using Fitted Q Evaluation and logistic regression.

Main Results:

• The RL model achieved higher cumulative rewards than clinician policies across all cohorts.
• Alignment between clinician actions and RL recommendations was linked to reduced odds of pAKI.
• The RL model suggested smaller IV fluid volumes and adjusted vasopressor/inotrope dosing compared to standard care.

Conclusions:

• Personalized hemodynamic management using an RL model demonstrated a decreased risk of pAKI post-cardiac surgery.
• AI-guided hemodynamic strategies hold potential for enhancing postoperative care.
• Further testing in randomized clinical trials is warranted to confirm these findings.