Electronic Phenotype for Detection, Staging, and Subtyping of Acute Kidney Injury

Distinguishing between transient and sustained subtypes of acute kidney injury (AKI) among hospitalized patients is valuable for clinical management and risk stratification. This study developed and validated a pragmatic electronic phenotype (e-phenotype) for the diagnosis, staging, and subtyping of AKI using electronic health record (EHR) data.

Development of a computable rule-based algorithm to diagnose, stage, and subtype AKI using longitudinal changes of serum creatinine values recorded in an EHR. Assessment of the test characteristics of these algorithm-based diagnoses was implemented using a set of patients admitted to the emergency department (ED) with or without clinically diagnosed AKI. Validation of AKI diagnoses was implemented in 2 ways: using a set of patients hospitalized for COVID-19 and using a dataset of patients hospitalized for any cause following an ED visit. These analyses examined the associations of AKI stage and subtype with mortality.

Assessment of the test characteristics of the algorithm-based diagnoses: 90 ED visits for AKI and 376 visits without clinical evidence of AKI at Columbia University. Validation in the setting of COVID-19: EHR data from 117,514 instances of COVID-19 infection diagnosed at Columbia University Medical Center throughout the pandemic. Validation in the setting of general hospital admissions: 405,467 general hospital admissions at Beth Israel Medical Center in Boston, Massachusetts.

The algorithm-based diagnosis, stage, and subtype of AKI were compared with the clinically adjudicated AKI diagnosis, stage, and subtype.

AKI detected, staged, and subtyped by an electronic algorithm, and 30-day mortality.

The AKI e-phenotype had a positive predictive value of 95.4%, sensitivity of 69.0%, specificity of 99.2%, and overall accuracy of 93.4%. In COVID-19 patients, pre-existing CKD was an independent predictor of AKI. COVID-19-related AKI was associated with mortality in a stage-dependent manner. The subtype of sustained AKI was associated with higher mortality compared with transient AKI within and across all pandemic waves. These results were reproducible in the dataset of patients hospitalized for any condition.

Reliance on serum creatinine patterns alone and the inability to incorporate urine output or molecular markers to diagnose and subtype AKI.

The AKI e-phenotype is accurate, scalable, and generalizable to diverse EHR datasets with reproducible associations with mortality.

Acute kidney injury (AKI) is a common and serious complication in hospitalized patients, but identifying and tracking it accurately in electronic health records is challenging. We developed an algorithm that uses patterns in kidney function test data to detect, classify, and distinguish AKI subtypes. After testing and validating the method, we applied it to 2 large hospital databases: one including patients hospitalized with COVID-19 and another including hospital admissions for any cause. Our approach reliably identified AKI and its association with poor outcomes. Patients with more severe or sustained AKI were more likely to die. This tool may help researchers study kidney injury more consistently across different health care settings.