The authors would like to correct several descriptions in the published article (https://doi.org/10.23876/j.krcp.26.055). The first correction concerns the description of the data source in the Methods section, under the subheading “Standardization via Observational Medical Outcomes Partnership Common Data Model.” The remaining corrections were identified when, while preparing the published model for external validation at other institutions, the authors reconciled the derivation-set variable values with the source records and the analysis code: the derivation of two input variables (estimated glomerular filtration rate [eGFR] slope and cumulative medication exposure), the handling of physiologically impossible values, and several OMOP concept identifiers in Supplementary Table 2 differed from the descriptions in the article. The original and corrected texts, and the consequences for interpretation, are provided below.
1. Methods, “Standardization via Observational Medical Outcomes Partnership Common Data Model”: data source
The original text read:
“To ensure interoperability and standardization, SNUH electronic health records were converted to the Observational Medical Outcomes Partnership Common Data Model (OMOP-CDM) version 5.3. We used standard OMOP Concept IDs to map local hospital codes to international standard vocabularies: SNOMED-CT for condition diagnoses, RxNorm for medication prescriptions, and LOINC for laboratory measurements (Supplementary Table 2, available online).”
The corrected text should read:
“Pseudonymized electronic health record data (January 2000–July 2025) were extracted from the SNUH clinical data warehouse through the institutional research retrieval system (SUPREME 2.0) after Data Review Board approval. Within the hospital’s secure analysis environment, the research team mapped local codes for condition diagnoses and laboratory measurements to the Observational Medical Outcomes Partnership Common Data Model (OMOP-CDM) version 5.3 standard concepts (SNOMED-CT and LOINC, respectively). Medication exposure variables were derived from ingredient names using rule-based string matching; RxNorm mapping of medications was performed for documentation (Supplementary Table 2, available online). The institutional OMOP-CDM database was not queried, and no full CDM conversion was performed.”
This correction clarifies the data extraction and standardization pathway only and does not affect the study cohort, model variables, analyses, results, or conclusions.
2. Methods, “Feature engineering”: derivation of the eGFR slope and medication exposure
The original text read:
“• eGFR slope: Calculated as a simple 2-point difference—(eGFR at index date – eGFR at most recent prior visit)/inter-measurement interval (years)—without regression fitting. Critically, only the index-date and prior-visit eGFR were used; the outcome eGFR (subsequent visit) was never incorporated into slope calculation, precluding data leakage. Patients without a prior measurement (Wave 1 entries) had eGFR slope set to missing and imputed with the SNUH training-set median.
• Drug duration: Cumulative exposure to renin-angiotensin-aldosterone system (RAAS) inhibitor (74.8% of derivation cohort) and statins (71.1%) was calculated as days from first prescription to index date, included as continuous features reflecting disease chronicity.”
The corrected text should read:
“• eGFR slope: In the SNUH derivation cohort, the eGFR slope used for model training and internal validation was computed as the least-squares linear regression slope of all serum creatinine–derived eGFR values available for each patient in the extracted record, not as a 2-point difference. The regression window was not restricted to measurements made on or before the index date: in 26,473 of 28,209 patients (93.8%), it included measurements made after the index date, within the 2-year outcome window. The statement that only index-date and prior-visit eGFR values were used therefore does not apply to the derivation cohort, and the derivation-cohort slope cannot be regarded as free of outcome-period information. In the KoGES external validation cohort, the slope was computed as the 2-point difference between consecutive wave measurements, as originally described; the slope definition applied at external validation thus differed from the definition under which the model was trained. Because the external validation used only pre-index information whereas the model had been trained with a slope that incorporated post-index information, this mismatch would be expected to reduce rather than inflate the reported external performance; the reported KoGES results are retained as the results of the published model evaluated as described. Patients without a prior measurement (Wave 1 entries) had the slope set to missing and imputed with the SNUH training-set median.
• Drug duration: In the derivation cohort, cumulative renin-angiotensin-aldosterone system (RAAS) inhibitor and statin exposure, and the corresponding use indicators, were derived from the entire extracted prescription record, including prescriptions issued after the index date, rather than from first prescription to the index date. Cumulative exposure was calculated as the sum of prescribed days across all prescriptions, so that overlapping prescriptions (e.g., repeated inpatient orders) were counted more than once, and the ingredient-name rule for statins also matched non-statin ingredients containing the string “statin” (e.g., nystatin and cilastatin). The reported prevalence of use (RAAS inhibitors 74.8%, statins 71.1%) reflects these definitions. With drug classes defined by Anatomical Therapeutic Chemical (ATC) codes, statin use over the full record was 66.6% (RAAS inhibitors unchanged at 74.8%), and when restricted to prescriptions up to the index date, 58.2% and 47.1% of patients, respectively, had been exposed.”
To quantify the effect of these definitions on the reported performance, we retrained the ensemble with the identical cohort, training/validation split, preprocessing, architecture, and hyperparameters after redefining the eGFR slope as the annualized change between the index-date eGFR and the most recent eGFR measured at least 90 days earlier (set to missing when unavailable) and medication exposure as the union of prescription intervals of the ATC-defined drug classes (RAAS inhibitors, C09; statins, C10AA) up to the index date. In the internal validation set (n = 8,463), the retrained model achieved an area under the receiver operating characteristic curve (AUROC) of 0.930 (95% CI, 0.925–0.936), an area under the precision-recall curve (AUPRC) of 0.890, and a Brier score of 0.094, compared with the reported AUROC of 0.939 (95% CI, 0.934–0.944), AUPRC of 0.901, and Brier score of 0.089; it remained superior to the 4-variable and 8-variable Kidney Failure Risk Equation (KFRE; AUROC, 0.879 and 0.884). Alternative pre-index slope definitions, and omission of the slope, yielded AUROCs of 0.931–0.932. The reported discrimination of the published model therefore includes a contribution of approximately 0.01 AUROC attributable to outcome-period information in the slope feature. The published model and the results reported in the article are retained as the results of the model as it was actually trained; the corrected variable definitions are being used for subsequent external validation studies of this model. Statements in the Abstract, Results, and Discussion that attribute predictive importance to the eGFR slope (including its SHAP ranking and its role in the subgroup and case-study results), as well as the internal calibration and decision-curve results, reflect the slope as actually computed and should be interpreted in light of this section.
3. Results, “Clinical utility and case study,” and the legend of Fig. 6
The original text read:
“To illustrate the practical utility of the model, Fig. 6 presents an individual prediction case in which the ML model correctly classified a ‘rapid progressor’ who might have been overlooked by traditional assessments because of preserved baseline function. Despite a relatively stable baseline eGFR (73.2 mL/min/1.73 m2), the model correctly assigned high risk (predicted probability 82%), driven principally by a markedly rapid eGFR slope of –30.6 mL/min/1.73 m2/yr (SHAP, +0.90), hypoalbuminemia (serum albumin, 3.1 g/dL; SHAP, +0.69), anemia (hemoglobin, 9.6 g/dL; SHAP, +0.22), hypocalcemia (calcium, 8.2 mg/dL; SHAP, +0.22), and severe metabolic acidosis (serum bicarbonate, 16 mmol/L; SHAP, +0.14). This demonstrates the capability of the model to capture dynamic risk factors that static scoring systems often miss.”
The corrected text should read:
“To illustrate the behavior of the model, Fig. 6 presents an individual prediction case of a patient with preserved baseline kidney function who subsequently showed rapid progression. Despite a relatively stable baseline eGFR (73.2 mL/min/1.73 m2), the model assigned high risk (predicted probability 82%), driven principally by an eGFR slope of –30.6 mL/min/1.73 m2/yr (SHAP, +0.90), hypoalbuminemia (serum albumin, 3.1 g/dL; SHAP, +0.69), anemia (hemoglobin, 9.6 g/dL; SHAP, +0.22), hypocalcemia (calcium, 8.2 mg/dL; SHAP, +0.22), and severe metabolic acidosis (serum bicarbonate, 16 mmol/L; SHAP, +0.14). As described in the corrected Methods, this slope was computed by regression over the patient’s full record and incorporated 93 creatinine measurements obtained after the index date, including the decline in eGFR that occurred within the outcome window; restricted to measurements before the index date, this patient’s 2-point slope was +20.9 mL/min/1.73 m2/yr. The case therefore illustrates the behavior of the model as published, but it should not be interpreted as a prospective identification of a rapid progressor from information available at the index date.”
Fig. 6 itself (the SHAP waterfall plot) is unchanged, as it depicts the model’s actual input values and contributions. In the legend of Fig. 6, the sentences beginning “A representative” and “Despite a relatively preserved” should be corrected as follows.
The original text read:
“A representative ‘Rapid Progressor’ case correctly identified by the ensemble machine learning (ML) model despite preserved baseline kidney function.”
“Despite a relatively preserved baseline estimated glomerular filtration rate (eGFR) of 73.2 mL/min/1.73 m2 (SHAP –0.46, protective), the model correctly assigned high risk driven principally by a markedly rapid eGFR slope of –30.6 mL/min/1.73 m2/yr (SHAP +0.90) and reduced serum albumin of 3.1 g/dL (SHAP +0.69).”
The corrected text should read:
“A representative case of a patient with preserved baseline kidney function who subsequently showed rapid progression and was assigned high risk by the ensemble machine learning (ML) model.”
“Despite a relatively preserved baseline estimated glomerular filtration rate (eGFR) of 73.2 mL/min/1.73 m2 (SHAP –0.46, protective), the model assigned high risk driven principally by an eGFR slope of –30.6 mL/min/1.73 m2/yr (SHAP +0.90), which was computed from the patient’s full record including measurements obtained after the index date (see corrected Methods), and reduced serum albumin of 3.1 g/dL (SHAP +0.69).”
4. Methods, “Data preprocessing and quality control”: physiologically impossible values and Fig. 1
The original text read:
“Conversely, physiologically impossible values were excluded prior to model training, including potassium >10 mmol/L, hemoglobin <3 g/dL, and eGFR slope values outside the biologically plausible range of ±60 mL/min/1.73 m2 per year (n = 6, 0.02%).”
The corrected text should read:
“Physiologically impossible laboratory values (potassium >10 mmol/L, hemoglobin <3 g/dL) were treated as missing at the variable level, consistent with the handling of other extreme values described above; no patient was excluded on this basis. No patient had a potassium value >10 mmol/L at the index measurement, and the two patients whose index hemoglobin was <3 g/dL were retained with another recorded value. The six patients (0.02%) whose eGFR slope lay outside ±60 mL/min/1.73 m2 per year were neither excluded nor set to missing; their computed slope values remained in the training data.”
Accordingly, the exclusion step for physiologically impossible values depicted in Fig. 1 did not remove any patient, and the cohort counts shown in Fig. 1 (derivation cohort, n = 28,209) are unaffected.
5. Supplementary Table 2: OMOP concept identifiers
In Supplementary Table 2 (available online), the concept identifiers listed for 12 laboratory variables designate concepts other than those named in the table (in most cases the identifier of a different laboratory test or an unrelated concept), and the ATC identifiers listed for the two RAAS inhibitor variables (use indicator and cumulative days) designate other drug classes. The concept names were correct. Because the derivation-cohort variables were derived from institutional laboratory codes and ingredient-name rules rather than by querying these identifiers (as stated in Section 1), the model, the analyses, and the results are not affected; the error concerns only the reproducibility of the variable definitions at other OMOP-CDM sites. The corrected identifiers for these 14 variables (13 rows; the two RAAS inhibitor variables share one row) are as follows (standard concept; LOINC or ATC code in parentheses):
|
Variable |
Published concept ID (actual concept designated) |
Corrected concept ID (concept) |
|
Baseline eGFR |
3049187 (eGFR, MDRD) |
1619025 (eGFR, CKD-EPI 2021; LOINC 98979-8)—at external sites eGFR is recalculated from serum creatinine (3016723) |
|
Calcium |
3020509 (albumin/globulin ratio) |
3006906 (calcium, serum/plasma; 17861-6) |
|
Phosphorus |
3024629 (urine glucose, test strip) |
3011904 (phosphate, serum/plasma; 2777-1) |
|
Bicarbonate |
3015469 (urine cortisol/creatinine) |
3015632 (total CO2, serum/plasma; 2028-9); bicarbonate 3016293 (1963-8) where reported as such |
|
Uric acid |
3022192 (triglyceride) |
3037556 (urate, serum/plasma; 3084-1) |
|
Total protein |
3020460 (C-reactive protein) |
3020630 (protein, serum/plasma; 2885-2) |
|
GGT |
3006923 (ALT) |
3026910 (GGT, serum/plasma; 2324-2) |
|
HDL cholesterol |
3004416 (procedure concept) |
3007070 (HDL cholesterol, serum/plasma; 2085-9) |
|
Triglycerides |
3022190 (observation concept) |
3022192 (triglyceride, serum/plasma; 2571-8) |
|
UACR |
3045479 (urine estrone) |
3034485 (albumin/creatinine, urine; 9318-7) |
|
Urine creatinine |
3016200 (unrelated concept) |
3017250 (creatinine, urine; 2161-8) |
|
Urine microalbumin |
3016407 (unrelated concept) |
3000034 (microalbumin, urine; 14957-5) |
|
RAAS inhibitors (use indicator and cumulative days) |
21601461, 21601560 (ATC C03, C04) |
21601783 (C09A), 21601801 (C09B), 21601822 (C09C), 21601832 (C09D), with descendants |
In addition, three identifiers are valid but narrower than intended and are clarified: age was derived from year of birth and the index date rather than from concept 4265453; ALT may be represented by 3006923 (LOINC 1742-6) in addition to 46236949; and the history of hypertension corresponds to concept 316866 (hypertensive disorder, with descendants) rather than 320128 (essential hypertension only). A corrected Supplementary Table 2, in which the definitions of the eGFR slope and the medication variables are also corrected in accordance with Section 2, is provided online.
Summary
These corrections concern the description of the data source, the description of how two input variables were derived, the handling of physiologically impossible values, the concept identifiers in Supplementary Table 2, and the interpretation of the case study. The study cohort, the trained model, and the results reported in the article are unchanged. The additional analysis reported in Section 2 indicates that, when the slope and medication variables are restricted to information available at the index date, the model’s discrimination in the derivation cohort is modestly lower (AUROC, 0.930 vs. 0.939) and its superiority over the KFRE is maintained. The authors apologize for these errors and for any inconvenience caused.