Abstract: FR-PO0035
Epic Cosmos to Describe AKI Events in Acute Heart Failure
Session Information
- Artificial Intelligence and Digital Health at the Bedside
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Roehm, Bethany Angela, The University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Rizvi, Musa Ali, The University of Texas Southwestern Medical Center, Dallas, Texas, United States
- McAdams, Meredith C., The University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Grodin, Justin, The University of Texas Southwestern Medical Center, Dallas, Texas, United States
- Basit, Mujeeb A, The University of Texas Southwestern Medical Center, Dallas, Texas, United States
Background
The prevalence of acute kidney injury (AKI) in acute heart failure (AHF) is highly variable (10-30%). How AKI in AHF is defined in the literature is also highly variable. AKI prevalence and applicability of AKI definitions to real-world data is unknown. We aimed to explore the types of data that could be leveraged to answer questions about AKI in AHF using data from Cosmos, a collaborative Epic health systems database with longitudinal healthcare records for 298 million patients.
Methods
Diagnoses and laboratory values were queried using ICD-10-CM and LOINC codes, respectively. All adult hospitalizations from 2017-2024 with a primary diagnosis of AHF were included. Patients had ≥1 outpatient serum creatinine (sCr) in the 12 months prior to admission (the average of which we called baseline sCr), and ≥1 sCr value during the admission. We defined four sub-cohorts: (1) hospitalizations with AKI by ICD-10-CM code N17.* (AKI-ICD), (2) hospitalizations with Stage 1 AKI per KDIGO criteria, (3) hospitalizations with Stage 2/3 AKI per KDIGO criteria (mutually exclusive), and (4) No-AKI.
Results
There were 606,131 AHF admissions attributable to 446,341 unique patients during the inclusion period. 40% had AKI by ICD-10 code; 43% had AKI by KDIGO criteria (Figure 1A). Only 6% had KDIGO stage 2/3 AKI. The positive predictive value of AKI-ICD was 72%; the negative predictive value was 77% (Figure 1B). sCr trajectories by AKI subgroup are shown in Figure 1C. Data on overall mortality, inpatient nephrology consults, and outpatient nephrology referrals were unreliably stored in Cosmos.
Conclusion
This analysis demonstrated that Cosmos may be a valuable resource of real-world data that can be used to evaluate kidney outcomes in AHF. There is granular data on repeated kidney function measures, medications, and in-hospital outcomes. Data are lacking on out-of-hospital mortality and nephrology consults.
Funding
- Other NIH Support