This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board of each participating center (Asan Medical Center, No. 2021-0031; Severance Hospital, No. 4-2021-0376; Seoul National University Hospital, H-2103-091-1205; and Seoul National University Bundang Hospital, B-1707-408-106). The requirement for informed consent was waived as the study is a retrospective, observational study with de-identified data.

This is a retrospective, observational study from four tertiary medical centers in Korea. We identified 2,484 patients with biopsy-proven IgAN from March 2012 to September 2021. Using the electronic health record system (EHRS), we collected all the clinical and pathologic data required to use the International IgA Nephropathy Prediction Tool. Patients were excluded who met the following criteria: 1) missing information for the Oxford classification or eGFR at biopsy or during follow-up, 2) aged <18 years, 3) diagnosed with chronic kidney disease (CKD) G5 requiring KRT at the time of biopsy, and 4) missing information for the International IgA Nephropathy Prediction Tool. Like the original study, we only included patients with one year or more of follow-up care and who developed CKD G5 or eGFR halving from baseline value within 1 year [18]. A total of 2,064 patients was included in the analytic cohort (Supplementary Fig. 1, available online).

Clinical information collected retrospectively from the EHRS comprised demographics, BP, serum creatinine, proteinuria, and use of RASB and immunosuppressants prior to biopsy and during follow-up. Serum creatinine was determined using an isotope dilution mass spectroscopy-traceable method, and eGFR was calculated using the CKD Epidemiology Collaboration equation [19]. Proteinuria was estimated using the spot urine protein-creatinine ratio when 24-hour urinary protein excretion information was missing [20]. Renal biopsies were analyzed by the local pathologists of each participating hospital and scored using the Oxford Classification of IgAN (MEST score) [13]. The pathologists were blinded to the clinical data. The linear predictors for each patient were calculated using predictors at biopsy according to the models with and without race, as previously published (Supplementary Table 1, available online) [18].

The primary outcome was a composite of a persistent decline in eGFR ≥ 50% from the baseline value or onset of ESKD during the follow-up period. ESKD was defined as initiation of KRT including dialysis, renal transplantation, or eGFR of <15 mL/min/1.73 m².

The performances of the two prediction models were evaluated using measures of model fit, discrimination, reclassification, and calibration. The Royston-Sauerbrei D statistic (R²_D) and Akaike information criterion were used to assess model fit [21]. Discrimination was evaluated using C-statistics [22]. We calculated the coefficient of the linear predictor in our cohort data by constructing a linear predictor-only Cox proportional hazard model for the primary outcome to measure discrimination. This estimated coefficient of the linear predictor is known as the “calibration slope,” and the discrimination in the validation dataset is almost the same when the linear predictor slope is approximately 1. If the calibration slope is >1, discrimination is better in the validation data set than in the original data set [23]. Continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI) for censoring were calculated to measure reclassification [24]. The 95% confidence intervals (95% CIs) for C-statistics, NRI, and IDI were obtained from 100 bootstrap samples. The differences between survival curves were examined using the log-rank test. The annual eGFR decline rate for each risk group was calculated using a linear mixed-effects model. To assess the accuracy of risk estimates from the models, we assessed both the entire follow-up period and time-specific calibration. The mean predicted survival curves within risk groups (categorized as <15th, 15th–50th, 51st–85th, and ≥86th percentile according to the linear predictor) were compared with the observed survival curves estimated by the Kaplan-Meier method for calibration over the follow-up period. In addition, subjects were grouped into deciles according to the predicted 5-year risk and compared with the observed 5-year risks in the time-specific calibration. The difference in calibration between the predicted and observed risks was visually compared. Additional analyses were performed in sequential steps to derive a Korean coefficient by recalibrating the prediction tool for Korean patients as follows. First, the International IgA Nephropathy Prediction Tool was applied directly to the Korean analytic cohort using the linear predictor with the ‘other’ race/ethnicity coefficient and baseline survival as previously published [18]. Then, by adjusting the linear predictor’s race coefficient (the ‘other’ race/ethnicity coefficient), the model with the Korean coefficient was evaluated and tested using the calibration plot and the Kaplan-Meier plot [23]. The model with a new coefficient that precisely predicted the observed kidney outcomes during the follow-up period was visually examined, and the coefficient value of overlap between the predicted curve and the observed risk was determined as the final Korean coefficient. Two-tailed p-values of <0.05 were considered statistically significant. Data were analyzed using STATA version. 16.1 (Stata Corp., College Station, TX, USA), R version. 4.1.0 (R Core Team, Vienna, Austria), and SAS version. 9.4 (SAS Institute, Cary, NC, USA).

The study adhered to the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) guidelines (Supplementary material, available online) [25].

The baseline characteristics of the external validation cohort are shown in Table 1. Among the 2,064 subjects, 353 (17.1%) experienced the primary outcome (Supplementary Fig. 2, available online). The median follow-up duration was 3.8 years (interquartile range, 1.8–6.6 years), with 843 subjects (40.8%) undergoing follow-up for 5 years or more. From baseline to year 5, 246 subjects developed the primary outcome, with a corresponding probability of 16.2% (95% CI, 14.3%–18.1%).

The Korean cohort was established more recently than the derivation and validation cohorts of the original study (the median years that biopsies were performed were 2014 vs. 2006 and 1998, respectively), with a smaller follow-up period (3.8 years vs. 4.8 and 5.8 years, respectively) [18]. In addition, the mean age at biopsy was higher in the Korean cohort than in the derivation and validation cohorts of the original study (41.3 years vs. 35.6 and 34.8 years, respectively), and the proportion of men was lower (44.8% vs. 57.8% and 49.3%, respectively) [18]. Our cohort had a median proteinuria of 1.0 g/day compared with 1.2 and 1.3 g/day in the two original cohorts and a median eGFR of 87.8 mL/min/1.73 m² compared with 83.0 and 89.7 mL/min/1.73 m² [18].

The C-statistics of the model without race and the model with race in the Korean cohort were 0.81 (95% CI, 0.77–0.84) and 0.81 (95% CI, 0.78–0.85), respectively. The calibration slopes of the linear predictors were 1.08 (95% CI, 0.96–1.20) and 1.03 (95% CI, 0.92–1.14), respectively, suggesting that discrimination was comparable to that of the original derivation cohort. The risk reclassification for predicting the 5-year primary outcome development risk was slightly improved in the model with race compared to the model without race (NRI, 0.13 [95% CI, 0.08–0.29]; IDI, 0.002 [95% CI, 0–0.007]) (Table 2). The Kaplan-Meier curves between risk subgroups according to model are shown in Fig. 1. The curves were separated distinctively between the lower-risk groups (low- and intermediate-risk) and the high- and highest-risk groups, suggesting that the model without race had a good discriminatory capability. However, the distinction between the low and intermediate-risk groups was less prominent using this model. When the eGFR decline rates were compared according to risk group using the model, there was a graded decrease in eGFR across the four risk groups, with the steepest decline occurring in the highest risk group (Table 3).

Both baseline survival probabilities from the original study underestimated the risk compared with the survival curve observed in our cohort (Fig. 2). When the predicted risk was visually compared with the observed risk, differences were observed in most of the risk groups, suggesting that both models underestimated risk (Fig. 1A, C). Notably, the model with race underestimated risk irrespective of risk group, although the risk was more highly underestimated in the high and highest risk groups (Fig. 1C). Fig. 1B and D show the observed vs. predicted 5-year risks of the primary outcome according to decile of predicted risk. In the model without race, the observed risk in the lower risk deciles approximated a perfect calibration threshold, suggesting that the model estimated risk relatively well. In the higher deciles, the observed risk was above the perfect calibration threshold, suggesting that the model without race underestimated risk for those in the higher deciles of predicted risk (Fig. 1B). In contrast, the model with race underestimated the 5-year risk of the primary outcome for those in all risk deciles (Fig. 1D).

Using the updating strategy as described in the statistical analysis section, we found that the Korean race/ethnicity coefficient was 0.5 (Supplementary Table 1, available online). Since we recalibrated the race/ethnicity coefficient only and our cohort included only Korean people, the C-statistics, NRI, and IDI results for predicting the 5-year primary outcome risk were identical in the two models. The predicted and observed primary outcome risks over the follow-up period and the calibration between the predicted and observed risks using the model updated with the Korean race/ethnicity coefficient are shown in Fig. 1E and F, respectively. Of note, the predicted and observed risks were consistent in the updated model, which was better than the results of the two original models using the International IgA Nephropathy Prediction Tool