Abstract: TH-OR011
Multimodal Deep Learning AKI Model Identifies Patients at Risk for Severe AKI and Adverse 90-Day Outcomes
Session Information
- AKI Across the Care Continuum: From Detection to Recovery
October 22, 2026 | Location: Room 203, Convention Center
Abstract Time: 04:40 PM - 04:50 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Koyner, Jay L., The University of Chicago Division of the Biological Sciences, Chicago, Illinois, United States
- Fatima, Aiman, The University of Chicago Division of the Biological Sciences, Chicago, Illinois, United States
- Anjorin, Ola, The University of Chicago Division of the Biological Sciences, Chicago, Illinois, United States
- Oguss, Madeline K., University of Wisconsin-Madison, Madison, Wisconsin, United States
- Weber, Michael J., University of Wisconsin-Madison, Madison, Wisconsin, United States
- Singh, Tripti, University of Wisconsin-Madison, Madison, Wisconsin, United States
- Carey, Kyle, The University of Chicago Division of the Biological Sciences, Chicago, Illinois, United States
- Churpek, Matthew M., The University of Chicago Division of the Biological Sciences, Chicago, Illinois, United States
Group or Team Name
- ESTOP Investigators
Background
Timely identification of patients at risk of severe AKI is critical in improving outcomes. Artificial intelligence-based risk scores may help identify patients earlier, but few scores have been prospectively validated.
Methods
Across 2 academic medical centers, we identified inpatients at high risk for severe AKI (≥KDIGO Stage 2) based on a previously published multimodal deep learning model (ESTOP2.0) value ≥0.091. Patients were enrolled within 8 hours of elevated risk, and demographics, clinical course, hospital-based outcomes (including KDIGO AKI), and 90-day outcomes were collected. We excluded patients who had already developed ≥Stage 2 AKI before the time of the high-risk score, as well as those with ESKD and kidney transplants.
Results
We enrolled 148 patients, the median (IQR) age was 64(52-73) years. Patients were 55% male, 50% Caucasian and 47% African American with 47(31%) of them ventilated and 114(76%) in the ICU at the time of enrollment. Patients had a median pre-hospital baseline serum creatinine (SCr) mg/dL of 1.09(0.82--1.35) which was 1.30 (1.0-1.66) at the time of enrollment, with 68(45%) patients having Stage 1 AKI at the time of consent. The table summarizes the AKI, hospital-based and 90-day outcomes for those with no AKI and Stage 1 AKI at enrollment. An ESTOP 2.0 score ≥0.091 identified a cohort at high risk for adverse events (severe AKI (38%), inpatient mortality (19%) and 90-day readmission (63%). Outcomes were similar across those enrolled with no AKI and Stage 1 AKI.
Conclusion
In this prospective, multicenter study, our previously developed multimodal deep learning AKI model identified patients at high risk of severe AKI, death, and readmission. Future interventional studies are needed to determine if it could be harnessed for early identification to improve patient outcomes.
Outcomes Stratified By AKI Status At Enrollment
| n(%) | No AKI at Enrollment (n=80) | Stage 1 AKI at Enrollment (n=68) |
| Maximum KDIGO AKI Stage within 7 days | ||
| No AKI Stage 1 Stage 2 Stage 3 | 29 (36) 22(28) 20 (25) 9 (11) | n/a 41(60) 15(25) 10 (15) |
| Receipt of Dialysis | 6(8) | 3(4) |
| Inpatient Mortality | 17(21) | 12(18) |
| 90-Day Outcomes | ||
| Hospital Readmission | 51(64) | 42(62) |
| Nephrology Issue on Readmission | 14(27) | 12(29) |
| New Receipt of Dialysis | 1(1) | 1(1) |
| 90-Day Mortality | 24(30) | 24(35) |
p=NS for all comparisons
Funding
- NIDDK Support