Abstract: TH-OR06
Development of a Machine Learning Algorithm to Predict Major Adverse Kidney Events (MAKE) After Hospitalization
Session Information
- AKI Research: Prediction, Diagnosis, Recovery
November 03, 2022 | Location: W230, Orange County Convention Center‚ West Building
Abstract Time: 05:15 PM - 05:24 PM
Category: Acute Kidney Injury
- 102 AKI: Clinical‚ Outcomes‚ and Trials
Authors
- Koyner, Jay L., University of Chicago Pritzker School of Medicine, Chicago, Illinois, United States
- Carey, Kyle, University of Chicago Pritzker School of Medicine, Chicago, Illinois, United States
- Churpek, Matthew M., University of Wisconsin-Madison, Madison, Wisconsin, United States
Background
MAKE are common after inpatient stays and are associated with increased morbidity, mortality, and costs. We seek to develop a machine learning (ML) risk score to identify patients at high risk for MAKE following hospitalization.
Methods
All patients discharged alive following admission to the University of Chicago between November 2008 and June 2020 were eligible for inclusion, and patients with a history of dialysis and/or chronic kidney disease ((CKD) stage 4 or 5 based on ICD codes or admission creatinine (SCr) ≥3) prior to admission were excluded. An ML-gradient boosted algorithm was developed using demographics, inpatient vital signs, and laboratory results to identify patients at risk for MAKE within 90 of discharge (primary outcome). MAKE was defined as a composite of new CKD (defined by ICD codes for CKD4 or 5 or an SCr > 3.0 mg/dl), recurrent AKI (defined by KDIGO SCr criteria or ICD codes), ESRD / need for dialysis, or mortality. The algorithm was developed in 70% of the admissions and areas under the receiver operating characteristic curve (AUCs) were calculated for the MAKE composite endpoint and the individual components at day 90 and 365 in the held-out 30% test data.
Results
Of the 50,448 included patients, 9,931 (19.7%) developed a MAKE outcome within 90 days of discharge. The ML model provided an AUC(95%CI) of 0.74(0.73,0.75) for the detection of MAKE90 in the test set. The model performed best at identifying those patients who developed post-hospitalization CKD at both 90 and 365 days (Table).
Conclusion
We developed and validated a post-hospital discharge ML risk algorithm to predict the future development of MAKE90. Our model can be used to identify the discharged patients most in need of follow-up with nephrology.
AUC for MAKE Outcomes
| Outcome | Timing | |
| Major Adverse Kidney Event | 90 Days | 0.74 (0.73, 0.75) |
| 365 Days | 0.73 (0.72, 0.73) | |
| CKD Progression | 90 Days | 0.94 (0.93, 0.95) |
| 365 Days | 0.92 (0.91, 0.93) | |
| Recurrent AKI | 90 Days | 0.74 (0.73, 0.74) |
| 365 Days | 0.72 (0.72, 0.73) | |
| Need for New Dialysis | 90 Days | 0.87 (0.83, 0.91) |
| 365 Days | 0.85 (0.82, 0.87) | |
| Death | 90 Days | 0.76 (0.74, 0.78) |
| 365 Days | 0.75 (0.74, 0.76) |
Funding
- NIDDK Support