The study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of Seoul National University Hospital (IRB No. H-2208-015-1347). The requirement for informed consent was waived by the IRB because of the retrospective observational nature of the study.

This was a large-scale retrospective cohort study analyzing electronic health records generated between 2006 and 2021 at the Seoul National University Hospital, Seoul, Korea. A flowchart of the study is shown in Fig. 1. Individuals 20 years of age or older who underwent at least two serum creatinine (SCr) measurements within 7 days from January 1, 2006, to December 31, 2021, were screened for AKI. Among consecutive SCr results, the former was considered the baseline SCr level. In cases where the participants had more than three SCr results, the initial two SCr measurements were used to determine the AKI event. Acute kidney injury events were defined according to the Kidney Disease Improving Global Outcomes (KDIGO) SCr criteria. The exclusion criteria were as follows: (1) individuals with baseline SCr ≥ 4 mg/dL or eGFR < 15 mL/min/1.73 m2, (2) individuals who received kidney replacement therapy (i.e., hemodialysis, peritoneal dialysis, and continuous kidney replacement therapy) before baseline; and (3) individuals whose baseline SCr was measured before the adoption of isotope-dilution mass spectrometry (IDMS), which occurred on January 11, 2011. The remaining participants without missing data were included in the final cohort for construction of the AKI prediction model.

Study flow diagram. SCr, serum creatinine; eGFR, estimated glomerular filtration rate; AKI, acute kidney injury; KRT, kidney replacement therapy

Baseline characteristics (age, sex, body mass index (BMI), and systolic and diastolic blood pressure), laboratory results, and procedure (i.e., surgery or coronary artery intervention) data were collected at baseline. Hospital-related data, including the clinical department where the SCr test was conducted (internal medicine, surgical, or emergency medicine) and the type of hospital visit (admission, outpatient unit, or emergency room) were collected. Comorbidities, including hypertension, diabetes mellitus (DM), cardiovascular disease, and malignancy, were defined using the International Classification of Diseases, 10th revision (ICD-10) diagnostic codes, and relevant medical prescriptions. Prescription drugs included renin–angiotensin–aldosterone system blockers, beta-blockers, calcium channel blockers, diuretics, oral hypoglycemic agents, and insulin. Details of the ICD-10 and prescribed drug codes are presented in Supplemental Table 1. Among the participants who started kidney replacement therapy after an AKI event, the type and date of dialysis were collected. The medical records of the study participants were obtained either during admission, stay in the emergency room, or in the community setting.

Body mass index was categorized into < 18.5 (underweight), ≥ 18.5 and < 25 (reference range), ≥ 25 and < 30 (overweight), and ≥ 30 kg/m2 (obese). Systolic blood pressure was categorized into normal (≥ 100 and < 130 mmHg) and abnormal (< 100 or ≥ 130 mmHg). Anemia was defined as a hemoglobin level of < 12 g/dL for female and < 13 g/dL for male. Dysnatremia was defined as a serum sodium level < 135 or ≥ 145 mEq/dL.

The eGFR was calculated using baseline SCr, sex, and age at baseline with three GFR estimating equations: 2009 CKD-EPIcr [7], 2021 CKD-EPIcr [10], and EKFC [22]. Based on each eGFR equation, participants were classified into the five categories of the KDIGO classification: ≥ 90, ≥ 60 and < 90, ≥ 45 and < 60, ≥ 30 and < 45, and ≥ 15 and < 30 mL/min/1.73 m2. Differences between the 2009 and 2021 CKD-EPIcr equations were calculated by subtracting the eGFR of the 2021 CKD-EPIcr from that of the 2009 CKD-EPIcr.

The study outcome was the development of AKI within 1 week of baseline. Acute kidney injury was defined according to KDIGO SCr criteria as an increase in SCr by ≥ 0.3 mg/dL within 48 h or an increase in SCr to ≥ 1.5 times baseline within 7 days.

Categorical variables were reported as numbers (percentages) and continuous variables as means (± SD) or medians [interquartile range (IQR)]. The bias of the 2009 and 2021 CKD-EPIcr equations was calculated by extracting the 2021 CKD-EPIcr from the 2009 CKD-EPIcr value and is presented in the histogram.

A multivariate regression prediction model was developed for the dataset. Model 1 was adjusted for age and sex, and Model 2 was adjusted for predictor variables, including demographic (age and sex) and clinical (BMI and systolic BP) values, laboratory values (presence of anemia, eGFR, and dysnatremia), clinical department (internal medicine, surgery, and emergency medicine), type of hospital visit (admission, outpatient unit, or emergency room), comorbidities (hypertension, DM, and malignancy), and procedures within 2 weeks before baseline. Two models were constructed based on the eGFR equations, wherein each model evaluated both continuous and categorical eGFR values.

The AKI prediction performance of the models was assessed by calculating the area under the receiver operator curve (AUROC) with 95% confidence interval (CI), and pairwise comparisons of AUROC were determined by the Delong method [23]. Analyses were repeated within subgroups of age, sex, DM, presence of CKD stage ≥ 3 (2009 CKD-EPIcr-estimated GFR threshold at 60 mL/min/1.73 m2), procedure (within 14 days before baseline), and year of baseline (2011–2013, 2014–2017, and 2018–2021).

To further assess reclassification, the net reclassification improvement analysis for AKI was implemented [24]. We set the 2009 CKD-EPIcr equation as a standard model, and the reclassification index of the 2021 CKD-EPIcr and EKFC equations was evaluated using continuous eGFR values. Bootstrapping with 100 replicates was used to construct 95% CIs for net reclassification improvement.

All statistical analyses were performed with R (version 4.2.3), and two-sided P values < 0.05 were considered statistically significant. The 1:1 exact matching was performed using “MatchIt” package in R. ROC and net reclassification improvement analyses were performed using the “pROC” and “nricens” package in R, respectively.