Kidney Week

Abstract: TH-OR010

Discrepancies in Electronic Health Record Labels for AKI: Expert Adjudication of AKI in MIMIC-IV

Session Information

• AKI Across the Care Continuum: From Detection to Recovery
  October 22, 2026 | Location: Room 203, Convention Center
  Abstract Time: 04:30 PM - 04:40 PM

Category: Acute Kidney Injury

• 101 AKI: Epidemiology, Risk Factors, and Prevention

Authors

• Mahzoon, Vahid, Temple University College of Science and Technology, Philadelphia, Pennsylvania, United States

Vahid Mahzoon

• Talebianaraki, Elmira, Temple University Lewis Katz School of Medicine, Philadelphia, Pennsylvania, United States

Elmira Talebianaraki

• Ojeniyi, Solabomi Oyeronke, Temple University Lewis Katz School of Medicine, Philadelphia, Pennsylvania, United States

Solabomi Oyeronke Ojeniyi, MD

• Koul, Sheetal, Temple University Lewis Katz School of Medicine, Philadelphia, Pennsylvania, United States

Sheetal Koul, MD

• Gillespie, Avrum, Temple University Lewis Katz School of Medicine, Philadelphia, Pennsylvania, United States

Avrum Gillespie, MD

• Vucetic, Slobodan, Temple University College of Science and Technology, Philadelphia, Pennsylvania, United States

Slobodan Vucetic

Background

Accurate identification of acute kidney injury (AKI) from electronic health records (EHRs) is essential for AKI research. The Medical Information Mart for Intensive Care IV (MIMIC-IV) is a publicly accessible EHR dataset of patients admitted to the emergency or an intensive care unit at Beth Isreal Deaconess Medical Center. Previous MIMIC-IV AKI studies have relied on International Classification of Diseases 10 (ICD) AKI diagnosis codes or change in creatinine from the structured data (Δcreatinine) per the Kidney Disease: Improving Global Outcomes (KDIGO) AKI creatinine criteria (increase ≥ 0.3 mg/dl or 1.5 x baseline). The accuracy of these methods remains unclear. Therefore, we used expert physician adjudication of AKI diagnoses using structured data as well as unstructured data from the discharge summaries.

Methods

Three nephrologists carefully reviewed and adjudicated a sample of 150 admissions selected from 41,927 ICU admissions in the MIMIC-IV database. AKI adjudication occurred in two stages. First, the experts assessed discharge summaries alone for evidence that treating physicians believed the patients had AKI. Second, they assessed discharge summaries together with the accompanying structured clinical data to determine if the patients had AKI. Experts then evaluated discrepancies until a consensus was reached. Accuracy with adjudicated AKI was measured by Receiver Operator Curve (ROC).

Results

Of the 150 charts, 114 (76%) had AKI as determined by the experts. Compared to the experts, Δcreatinine had the highest accuracy (ROC 0.91) followed by ICD (0.84) and discharge summary (0.42). Discrepancies between Δcreatinine and experts were because of data in the discharge summary but not the structured dataset. Seven percent of expert AKI diagnoses did not have an ICD code and 42% of cases adjudicated as no AKI had an ICD AKI diagnosis. Discharge summaries were the least accurate, missing an AKI diagnosis in 75% of cases, but occasionally providing clinical context and missing data.

Conclusion

This study provides an insight into challenges associated with determining AKI from EHR and limitations of commonly used ICD codes or Δcreatinine from structured data. The MIMIC-IV dataset has the potential for improving AKI outcomes using data science techniques; however, newer methods that use both structured and unstructured data are necessary to improve accuracy of labeling AKI.