Abstract: PO0006
Development and Validation of a Model to Predict In-Hospital AKI Among Hospitalized US Veterans
Session Information
- AKI Epidemiology, Risk Factors, and Prevention: Clinical Research
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Singh, Karandeep, University of Michigan, Ann Arbor, Michigan, United States
- Zhang, Xiaosong, University of Michigan, Ann Arbor, Michigan, United States
- Yin, Maggie, University of Michigan, Ann Arbor, Michigan, United States
- Shahinian, Vahakn, University of Michigan, Ann Arbor, Michigan, United States
- Weitzel, William, University of Michigan, Ann Arbor, Michigan, United States
- Saran, Rajiv, University of Michigan, Ann Arbor, Michigan, United States
- Crowley, Susan T., VA Connecticut Health System, West Haven, Connecticut, United States
- Heung, Michael, University of Michigan, Ann Arbor, Michigan, United States
Background
Acute kidney injury (AKI) occurs commonly in hospitalized patients, affects millions of Americans annually, and is recognized as one of the most significant contributors to chronic kidney disease (CKD). As the primary treatment for most AKI is limited to supportive care, improving the health consequences of AKI (most notably CKD) relies on identifying high-risk patients in a timely manner and targeting them for preventive interventions.
Methods
Using data from a random sample of 100,000 hospitalized veterans, we developed and internally validated a gradient-boosted decision tree (GBDT) model for the prediction of in-hospital acute kidney injury. We excluded patients without an available creatinine during their hospitalization (n = 21,165), and randomly divided the remaining patients into a training cohort (n = 39,418), tuning cohort (n = 13,139), and a test cohort (n = 26,278). We divided each patient’s hospitalization into 6-hour steps and removed all steps where a patient actively had known AKI. We then used the prior 48 hours of clinical data to predict the onset of AKI during the next 48 hours. After fitting the GBDT model on all steps from the training cohort, we iteratively checked model performance on the tuning cohort to determine optimal parameters for early stopping to prevent overfitting. We then evaluated the model’s performance on the test cohort using the maximum predicted score for each patient prior to the onset of AKI.
Results
Our cohort had a mean age of 72, was 96% male, 34% with baseline diabetes mellitus, and 24% with baseline CKD. 11% of patients experienced in-hospital AKI by creatinine-based KDIGO criteria (3.8% in any 6-hour window), and 1.6% of patients experienced AKI stage 2+. The area under the curve (AUC) was 0.76 (95% bootstrap CI 0.75-0.77) for predicting any AKI in the next 48 hours and 0.70 (95% CI: 0.67-0.73) for predicting AKI stage 2+ in the next 48 hours. Patients who met or exceeded a risk threshold of 14.5% had a 33% chance of developing AKI during their hospitalization with a 42% sensitivity.
Conclusion
In a retrospective validation in a large national cohort of veterans, a predictive model for acute kidney injury identifies nearly half of patients who will experience AKI prior to its onset.
Funding
- Veterans Affairs Support