Abstract: FR-OR005
KidneyPilot: Reinforcement Learning Model for Prevention of Persistent AKI After Cardiac Surgery
Session Information
- AKI Everywhere All at Once: Risks, Trajectories, and Practice Patterns
November 07, 2025 | Location: Room 320A, Convention Center
Abstract Time: 05:10 PM - 05:20 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Sabounchi, Moein, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Desman, Jacob M, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Shenfeld, Idan, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
- Jayaraman, Pushkala, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Campoli, Michelle, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Kohli-Seth, Roopa D., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Charney, Alexander W., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Agrawal, Pulkit, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
- Nadkarni, Girish N., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Sakhuja, Ankit, Icahn School of Medicine at Mount Sinai, New York, New York, United States
Background
Acute Kidney Injury (AKI) occurs in over one-third of patients after cardiac surgery. When lasting 48 hours or more (persistent AKI), it more than doubles the risk of death and dialysis compared to transient AKI. AKI management aims to prevent persistence by optimizing fluid status and blood pressure using IV fluids, vasopressors, and inotropes. However, these decisions are not personalized. We developed and validated a reinforcement learning model, KidneyPilot, to optimize fluid, vasopressor, and inotrope dosing in post-cardiac surgery patients to prevent persistent AKI.
Methods
We included adult patients (≥18 years) admitted to the ICU after cardiac surgery from MIMIC-IV (70% training, 15% validation, 15% internal test) and SICdb (external validation). Routinely available features (demographics, vitals, labs, fluid balance, medications) as well as previous fluid, vasopressors, and inotropes, were aggregated into hourly bins. We applied conservative Q-learning, a state-of-the-art offline RL model, to personalize fluid, vasopressor, and inotrope dosing. Model performance was assessed using fitted Q evaluation, an established off-policy method to compare total rewards of KidneyPilot versus clinicians in preventing persistent AKI.
Results
There were 6,623 patients in MIMIC-IV and 2,230 in SICdb. The mean age in MIMIC-IV was 68.1 years with 71.87% males, compared to 67.1 years and 72.78% males in SICdb. KidneyPilot generally recommended lower IV fluid volumes and vasopressors, and higher inotropes than clinicians in both internal test and external validation datasets. On OPE, KidneyPilot achieved a mean reward of 54.38 ± 1.33 vs. 18.69 ± 1.10 for clinicians in the internal test set, and 76.48 ± 0.78 vs. 21.33 ± 1.05 in the external validation set.
Conclusion
The KidneyPilot, provides a novel tool to personalize fluids, vasopressors, and inotropes dosing after cardiac surgery to prevent the development of persistent AKI. Future prospective testing is essential to establish its real-world applicability.
Funding
- Other NIH Support