Abstract: TH-OR098

Development of an AKI Prediction Model Using Machine Learning

Session Information

  • Predicting AKI
    November 02, 2017 | Location: Room 282, Morial Convention Center
    Abstract Time: 04:30 PM - 04:42 PM

Category: Acute Kidney Injury

  • 003 AKI: Clinical and Translational

Authors

  • Koyner, Jay L., University of Chicago, Chicago, Illinois, United States
  • Carey, Kyle, University of Chicago, Chicago, Illinois, United States
  • Churpek, Matthew M., University of Chicago, Chicago, Illinois, United States
Background

Early identification of hospitalized patients at risk for the development of AKI prior to changes in serum creatinine(SCr) may improve patient outcomes. We aimed to develop an AKI risk prediction algorithm using electronic health record(EHR) data across ward and ICU patients

Methods

All hospitalized patients at the University of Chicago who had SCr measured from 11/2008 to 1/2016 were eligible. Patients with a first SCr>3.0mg/dl, those who had an ICD9 code for CKD Stage 4 or higher, or received renal replacement therapy(RRT) within 48 hours (hrs) of admission were excluded. Demographics, vital signs, lab results, interventions, medications, blood transfusion & diagnostic testing were utilized in a gradient boosted machine learning algorithm to predict SCr-based KDIGO stage 2 AKI, with 60% of the data used for derivation and 40% for validation. Area under the curve (AUC) was calculated in the validation cohort, and subgroup analyses were conducted across admission SCr, AKI severity, and hospital location

Results

Among the 121,158 included patients, 17,481(14.4%) developed KDIGO AKI, with 4,251(3.5%) developing Stage 2 and 1,997(1.6%) Stage 3. The AUC(95%CI) of the model in the validation cohort was 0.90(0.90-0.90) for predicting Stage 2 AKI within 24 hrs and 0.87(0.87-0.87) within 48 hours. The AUC was 0.95(0.95-0.95) for Stage 3 in 24 hrs and 0.92(0.92-0.93) for 48 hrs. Accuracy was excellent for predicting RRT(n=821) in the next 72 hrs 0.94(0.94-0.94). AUCs for the subgroups can be found in the table. At a threshold with a sensitivity of 85% and a specificity of 85%, the median time from first reaching the threshold to stage 2 AKI was 42(IQR 12,140) hrs

Conclusion

Readily available EHR data can be used to predict impending AKI prior to changes in SCr with excellent accuracy across different patient locations and admission SCr. Real-time use of this model would allow early diagnostic and therapeutic interventions for those at high risk of AKI and may improve cost and outcomes

 AUC(95%CI) for Predicting Stage 2 AKI within 24 hrs
Patient Location 
Ward0.88(0.88-0.88)
ICU0.88(0.88-0.89)
Admission SCr (mg/dL) 
<1.00.88(0.88-0.89)
1.0 to 1.90.92(0.92-0.92)
2.0 to 2.90.92(0.92-0.92)