Abstract: FR-OR03
Use of a Recurrent Neural Network to Predict Development of Nephrotoxic AKI in Adults
Session Information
- AKI Prediction and Assessment: Traditional and Novel Tools
November 03, 2023 | Location: Room 118, Pennsylvania Convention Center
Abstract Time: 04:48 PM - 04:57 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Griffin, Benjamin R., University of Iowa Hospitals and Clinics, Iowa City, Iowa, United States
- Sarrazin, Mary Vaughan, University of Iowa Hospitals and Clinics, Iowa City, Iowa, United States
- Jalal, Diana I., University of Iowa Hospitals and Clinics, Iowa City, Iowa, United States
- Misurac, Jason, The University of Iowa Stead Family Children's Hospital, Iowa City, Iowa, United States
Background
The Nephrotoxic Injury Negated by Just-in Time Action (NINJA) program identifies children with high exposure to nephrotoxic medications and has been shown to reduce rates of acute kidney injury (AKI) by 25% through preemptive medication adjustments. However, attempts to translate NINJA to the adult population have been hampered by the much larger patient volume at adult hospitals combined with NINJA’s relatively low positive predictive value (PPV), which generates a high number of false alerts. We used a recurrent neural network (RNN) machine learning algorithm to improve nephrotoxic AKI prediction in patients with high nephrotoxin exposure.
Methods
We collected over 100 variables including demographics, laboratory data, vital signs, and medication usage on adult patients admitted to the University of Iowa Hospital from 2017-2022. Patients admitted for >48 hours who met the NINJA definition for high nephrotoxin exposure (≥ 3 nephrotoxins on 1 day or intravenous aminoglycoside or vancomycin for ≥ 3 days) were included in the final dataset. A gated recurrent unit (GRU)-based RNN was trained on 85% of the data, and then tested on the remaining 15%, with the goal of predicting AKI development (defined as a creatinine increase of ≥0.3 mg/dL or ≥1.5x baseline) within 48 hours. We then used an artificial neural network (ANN) to determine feature importance.
Results
There were 37,300 patient-days meeting criteria for NINJA high nephrotoxin exposure. In the testing cohort, 29% of exposures developed AKI within 48 hours (2.4 false alerts per true AKI). The RNN model predicted 48-hour AKI with a PPV of 0.58 (0.7 false alerts per true AKI) and a NPV of 0.87. Hospital day, lowest hemoglobin, lowest platelet count, lowest blood pressure, and highest white blood cell count were the 5 most important variables in the ANN model. Vancomycin, piperacillin-tazobactam, iopamidol, and lisinopril were the most important medication variables.
Conclusion
Our RNN machine learning model was able to dramatically reduce the number of false alerts for nephrotoxic AKI in adults, which may facilitate NINJA translation to adult hospitals by providing more targeted intervention with less resource utilization.