Abstract: SA-PO164
External Validation of an Electronic Health Record (EHR)-Based Machine Learning Risk Score for Hospital-Based AKI
Session Information
- AKI: Epidemiology, Risk Factors, Prevention - III
November 09, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Koyner, Jay L., University of Chicago, Chicago, Illinois, United States
- Afshar, Majid, Loyola University Chicago, Chicago, Illinois, United States
- Gilbert, Emily R., Loyola University Medical Center, Maywood, Illinois, United States
- Carey, Kyle, University of Chicago, Chicago, Illinois, United States
- Churpek, Matthew M., University of Chicago, Chicago, Illinois, United States
Background
We seek to externally validate our previously published EHR-based machine learning AKI risk score in data from a new hospital system.
Methods
All hospitalized patients who had SCr measured at Loyola University Medical Center (LUMC) from 2008 to 2016 were eligible. Patients with a first serum creatinine (SCr)>3.0mg/dl, those who had an ICD codes for CKD Stage 4 or higher, or received renal replacement therapy(RRT) within 48 hours(hrs) of admission were excluded. Demographics, vital signs, laboratory results, and nursing scores were utilized in the previously published gradient boosted machine learning algorithm based on data from the University of Chicago (UofC) to predict SCr-based KDIGO AKI. Areas under the curve (AUC) were calculated in the LUMC cohort, and subgroup analyses were conducted across admission SCr, AKI severity, and hospital location.
Results
Among the 194,930 included LUMC patients, 27,374 (14.0%) developed KDIGO AKI with 7,364 (3.8%) developing Stage 2 and 3,393 (1.7%) requiring RRT. These rates were similar compared to the UofC cohort (14.4% AKI, 3.5% Stage 2). The AUC (95%CI) of the model in the LUMC cohort was 0.80(0.80-0.80) for predicting Stage 2 AKI within 48 hrs compared to 0.86(0.86-0.86) in the UofC cohort. The AUC was 0.80(0.80-0.80) for Stage 3 in 48 hrs in the LUMC cohort. AUCs for subgroups (patient location and admitting SCr) at LUMC (24 and 48 hr predictions) can be found in the table.
Conclusion
We report the first externally validated machine learning EHR-based AKI risk algorithm. EHR data can be used to predict impending AKI prior to significant changes in SCr across different patient locations and baseline SCr values. We are using this validated EHR model in real-time in an active clinical trial seeking to improve AKI outcomes.
Validation of EHR Risk Score in Subgroups Based on Patient Location and Admission SCr
AUC (95%CI) for Predicting Stage 2 AKI within 24 hrs in LUMC cohort | AUC (95%CI) for Predicting Stage 2 AKI within 48 hrs in LUMC | |
Patient Location | ||
Ward | 0.80 (0.80 - 0.80) | 0.75 (0.75 - 0.75) |
ICU | 0.77 (0.77 - 0.77) | 0.74 (0.74 - 0.74) |
Admission SCr (mg/dL) | ||
<1.0 | 0.81 (0.81 to 0.81) | 0.78 (0.77 to 0.78) |
1.0 to 1.9 | 0.84 (0.84 to 0.84) | 0.80 (0.80 to 0.80) |
2.0 to 2.9 | 0.87 (0.86 to 0.87) | 0.83 (0.83 to 0.84) |
LUMC- Loyola University Medical Center Validation Cohort
Funding
- NIDDK Support