External Validation of International Risk-Prediction Models of IgA Nephropathy in an Asian-Caucasian Cohort.

INTRODUCTION: Two prediction models for IgA nephropathy (IgAN) using clinical variables and the Oxford MEST scores were developed and validated in 2 multiethnic cohorts. Additional external validation is required.
METHODS: Biopsy-proven Chinese and Argentinian patients with IgAN were included. The primary outcome was defined as a 50% decline in estimated glomerular filtration rate (eGFR) or end-stage renal disease. C-statistics and stratified analyses were used for model discrimination, coefficient of determination (R2 D) for model fit, and calibration plots for model calibration. Baseline survival function was also evaluated.
RESULTS: A total of 1275 patients were enrolled, with a mean age of 34 (interquartile range: 27-42) years, 50% of whom (638 of 1275) were men. Use of renin-angiotensin system blockers was higher than in previously reported cohorts, whereas other variables were comparable. The C-statistic of the models was 0.81, and R2 D was higher than reported. Survival curves in the subgroups (<16th, ∼16th to <50th, ∼50th to <84th, and ≥84th percentiles of linear predictor) were well separated. Most of the predictor variables, including hazard ratio, predicted 5-year risk, and eGFR decline slope, were worse with risk increasing. The baseline survival function was comparable in our cohort and the reported cohorts. The calibration was acceptable for the full model without race. However, the risk probability over 3 years was overestimated in the full model with race included.
CONCLUSION: The prediction models showed good performance on personalized risk assessment, which may be used as drug-specific, precision-medicine approaches to treatment decisionmaking.

IgA nephropathy (IgAN) is the most common form of primary glomerulonephritis worldwide and is characterized by clinical heterogeneity and ethnic variation., Nearly 10% and >20% of patients progress to end-stage renal disease (ESRD) within 5 and 20 years after diagnosis, respectively., Although a number of clinical variables, such as eGFR, proteinuria, and blood pressure, have been suggested as reliable prognostic factors of IgAN, none has accurately discriminated the disease risk individually or jointly., The diagnosis of IgAN depends on renal biopsy. The Oxford MEST (M = mesangial hypercellularity, E = endothelial hypercellularity, S = segmental sclerosis, T = tubular atrophy and interstitial fibrosis) histologic scores for IgAN were derived from a multiethnic population with high reproducibility and can provide prognosis evaluation., Thus, a prediction model combining clinical features and histologic scores would be valuable to better discriminate patients with a progressive disease course and permit assessment of targeted therapies in clinical trials and risk stratification in clinical practice.

Recently, the International IgA Nephropathy Network took advantage of the existing large multiethnic cohorts with long-term follow-up to develop and validate 2 full prediction models without and with race included. The 2 full models have the advantage of simplicity, being derived from factors routinely available, and have shown sufficient discrimination and calibration. The authors also provided a mobile app (QxMD) and a web-based calculator (https://qxmd.com/calculate-by-qxmd) for clinical use. We believe the development of such models in patients from diverse ethnic backgrounds will aid clinicians in patient stratification, treatment decisionmaking, clinical trial recruitment, and biomarker validation.

As emphasized in the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement, for all the prediction models, development and validation is only a first step. For any prediction model to be widely applicable, it must be validated in cohorts outside that in which it was derived but similar to the target population. Clearly, more validation studies detecting the generalizability of prediction models will be required., Although the full models performed well in the reported validation cohort, additional external validation would provide further evidence of prediction model performance in target populations. Notably, as it was stated that the study cohorts included patients from an “old era” (the 1980s), during which only about 30% of patients received renin-angiotensin system blockers (RASBs) at biopsy and a further 30% to 50% had these drugs added during follow-up. Because RASBs have greatly changed the progression of IgAN, in current clinical practice nearly all patients with IgAN routinely receive RASBs at or soon after biopsy, as recommended by the 2012 Kidney Disease International Global Outcomes (KDIGO) guideline. In recent large, randomized, controlled trials, including the Supportive Versus Immunosuppressive Therapy for the Treatment of Progressive IgA Nephropathy (STOP-IgAN), Targeted-release Budesonide Versus Placebo in Patients with IgA Nephropathy (NEFIGAN), and Therapeutic Evaluation of Steroids in IgA Nephropathy Global (TESTING) studies, approximately 90% patients were given RASBs. Recognizing that these models may be used in more current populations, the prediction performance of this model needs to be validated in cohorts from a “new era,” especially those consistent with the current KDIGO guideline. This will help clinicians to predict the risk of failure for routine treatment and to determine the level of treatment with immunosuppressants. Risk stratification is essential for the care of patients with IgAN to avoid unnecessary exposure to toxic therapies while reducing the risk of chronic kidney disease progression. It is also of vital importance for patients and physicians in making optimal health-related and life decisions.

Thus, in this study, we further evaluated the performance of 2 full models without and with race in a large external cohort of IgAN patients from northern China and Argentina.

Methods

Patients

To validate the full model without or with race, we enrolled an Asian-Caucasian cohort, including 1360 Chinese patients who were registered and with long-term follow-up in the Peking University First Hospital IgAN database (www.renal-online.org) since January 2003 and 116 patients with a long-term follow-up who were diagnosed and treated at the Hospital Británico de Buenos Aires since 1995. All patients were diagnosed by biopsy and those with <8 glomeruli per biopsy section were excluded. Our study was approved by the ethics committee of Peking University First Hospital and by the institutional review board of the Hospital Británico de Buenos Aires. Written informed consent was provided by all participants.

Variable Definitions

Baseline characteristics, including proteinuria, systolic blood pressure (SBP), diastolic blood pressure (DBP), eGFR (calculated using the Chronic Kidney Disease Epidemiology Collaboration formula), age, and previous use of RASBs and/or immunosuppressants, were collected at the time of biopsy. MEST histologic scores were evaluated according to the Oxford classification system by 3 pathologists independently, who were blinded to clinical data. Mean arterial pressure (MAP) was calculated as 1/3 × SBP + 2/3 × DBP. The primary outcome was a composite of the first occurrence of either ESRD (eGFR <15 ml/min per 1.73 m2, dialysis, or kidney transplantation) or a 50% decline in eGFR from the value at biopsy. The decline slope of eGFR was calculated using a mixed-effects model as reported elsewhere. For validation, each covariate and outcome were defined exactly according to the original publication using the same measurement units (Table 1).

Table 1

Definitions and units of predictors and outcome

Variables	Reported cohorts	Current cohort
Proteinuria, g per 24 h	Proteinuria was estimated from 24-h urine collections within 6 mo of renal biopsy; if not available, the value was estimated from spot urine protein:creatine ratios	Proteinuria was estimated from 24-h urine collections within 2 wk before or after renal biopsy
MAP, mm Hg	Blood pressure was taken within 6 mo of renal biopsy. MAP was calculated as 1/3 × SBP + 2/3 × DBP	The blood pressure was measured within 2 weeks of renal biopsy, and MAP was calculated as 1/3 × SBP + 2/3 × DBP
eGFR, ml/min per 1.73 m2	eGFR calculated by CKD-EPI formula using variants collected within 6 months of renal biopsy	eGFR was calculated by CKD-EPI formula, based on variants collected within 2 wk of renal biopsy
Age, yr	Age of patient at time of renal biopsy	Same as reported
RASB use (0/1)	Use of RASBs (ACE inhibitors or ARBs) before or at the time of renal biopsy	Same as reported
Immunosuppressant use (0/1)	Use of immunosuppressive therapy (included but not limited to corticosteroids, azathioprine, mycophenolate mofetil, etc.) before or at time of renal biopsy	Same as reported
M	M0: mesangial score <0.5; M1: mesangial score >0.5	Same as reported
E	E1, presence of endothelial hypercellularity	Same as reported
S	S1, present of segmental glomerular sclerosis; presence or absence of podocyte hypertrophy/tip lesions	Same as reported
T	T0, <25% tubular atrophy or interstitial fibrosis; T1, ∼25%–50% tubular atrophy or interstitial fibrosis; T2, ≥50% tubular atrophy or interstitial fibrosis	Same as reported
Outcome
50% decline of eGFR	First occurrence of eGFR <50% of baseline eGFR at renal biopsy	Same as reported
ESRD	eGFR was <15 ml/min per 1.73 m2, or the start of renal replacement therapy (dialysis or kidney transplant)	Same as reported

ACE, angiotensin-converting enzyme; ARB, aldosterone receptor blocker; CKD-EPI, Chronic Kidney Disease Epidemiology Collaboration; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; ESRD, end-stage renal disease; MAP, mean arterial pressure; MEST, M = mesangial hypercellularity, E = endothelial hypercellularity, S = segmental sclerosis, T = tubular atrophy or interstitial fibrosis; RASB, renin-angiotensin system blocker; SBP, systolic blood pressure.

Prediction Models for External Validation

The published prediction models for validation were derived as follows:where log is the natural log function.

For the full model without race:

For the full model with race:

Statistical Analysis

For model validation, we initially calculate the linear predictor for each patient in the current cohort based on the exact predictors and coefficient values as mentioned. We then assessed the model performance of discrimination and calibration according to Royston and Altman.

For discrimination, we first estimated the regression coefficient on the linear predictor coefficient by fitting a Cox proportional hazards model for the full model without race and an interval format Cox proportional hazards model for the full model with race in our data set. If βlinear predictor ≥ 1, then discrimination of the models would be acceptable. Second, the C-statistic was used to determine how well the model could distinguish those with an endpoint from those without an endpoint. Considering the censoring issue, the AUC.cd, according to Chambless and Diao, was calculated accordingly. Coefficient of determination (R2D) was calculated according to a method evaluating model fit performance. Third, we divided patients into risk groups, including <16th (low risk), ∼16th to <50th (intermediate risk), ∼50th to <84th (higher risk), and ≥84th (the highest risk) percentiles of the linear predictor from the full model without or with race. Subgroup analyses were performed and and survival curves derived. As suggested, in contrast to P values for comparing risk groups, the hazard ratios were suggested to be a sensible verification of model discrimination. Thus, hazard ratios were evaluated by fitting a Cox model with a dummy variable representing each risk group referring to the lowest risk group. When survival curves are more widely separated, the hazard ratio tends to be greater.

For calibration, because it is preferred to have patients with similar baseline risks on average, we first investigated the accuracy of the baseline survival function itself. The reported baseline survival function was obtained directly from the publication. A Kaplan-Meier‒like estimate of the baseline survival function in our data was obtained by standard methods after fitting a Cox model with no covariates other than the linear predictor, with regression coeffect constrained to 1. We then applied the averaging method to obtain predicted mean survival curves in our cohort and compared them with the Kaplan-Meier survival curves in the risk groups. Finally, we assessed calibration graphically by comparing the predicted survival probability with the observed percentage stratified by deciles and calculated Kaplan-Meier statistics using val.surv embedded in the “rms” package in R software. All statistical analyses were performed using R version 3.3.0 (https://www.R-project.org/).

Results

Baseline Characteristics and Outcomes

The flowchart for patient enrollment is shown in Figure 1a. There were 1275 patients enrolled in the analysis, including 1169 Chinese and 106 Argentinians. In our cohort, the percentage of combined outcomes was 14.2%. Among these, 13.5% achieved halving of their eGFR and 8.6% reached ESRD during the median 3.8-year follow-up. All rates were lower than the reported derivation and reported validation cohorts due to our relatively shorter follow-up period (P < 0.05, considering the general clinical slow progressive course of IgAN) (Figure 1b). The patient characteristics are displayed in Table 2. As stated earlier, the rates of RASB application ranged from 30.0% to 32.4% at biopsy and 66.4% to 86.7% after biopsy in the previously reported cohorts, but were 72.6% at biopsy and reached 91.3% after biopsy in our cohort. Similarly, rates of immunosuppressant use ranged from 7.1% to 9.1% in the earlier cohorts and was 11.1% in our cohort at biopsy. The distributions of other clinical parameters, including age, male ratio, baseline eGFR, blood pressure, proteinuria, and Oxford MEST histologic scores, were broadly similar between the present and previously reported cohorts. The baseline features of Chinese and Argentinian patients in our cohort are presented separately in Supplementary Table S1.

Figure 1

(a) Enrollment flowchart and (b) cumulative incidence of the primary outcome in the current cohort. Overall, 1275 of the original 1476 patients remained in the final cohort, including 1169 Chinese patients and 106 Argentinian patients. Among the excluded patients, 92 had other forms of glomerulopathy, 59 were <18 years old, 12 had end-stage renal disease at the time of renal biopsy, 22 were without available MEST score, and 16 lacked medication information. ESRD, end-stage renal disease; MEST, M = mesangial hypercellularity, E = endothelial hypercellularity, S = segmental sclerosis, T = tubular atrophy and interstitial fibrosis; RASB, renin-angiotensin system blocker.

Table 2

Description and comparison of the current and reported cohorts

Characteristics	Reported derivation cohort	Reported validation cohort	Our validation cohort
Number of patients	2781	1146	1275
Follow-up time, median (IQR), yr	4.8 (3.0–7.6)	5.8 (3.4–8.5)	3.8 (2.1–6.9)
Year of biopsy, median (IQR)	2006 (2004–2008)	1998 (1993–2003)	2010 (2006–2013)
Age, median (IQR), yr	35.6 (28.2–45.4)	34.8 (26.9–45.0)	34 (27–42)
Male, n (%)	1608 (57.8)	565 (49.3)	638 (50.0)
Race, n (%)
Caucasian	1167 (42.0)	176 (15.5)	106 (8.3)
Chinese	1021 (36.7)	292 (25.8)	1169 (91.7)
Japanese	569 (20.5)	616 (54.4)	—
Other	22 (0.8)	49 (4.3)	—
SCr at biopsy, median (IQR), μmol/l	92.0 (70.7–123.8)	84.0 (66.2–111.4)	90.20 (71.3–120.2)
eGFR at biopsy, median (IQR), ml/min per 1.73 m2	83.0 (56.7–108.0)	89.7 (65.3–112.7)	82.8 (59.9–104.6)
<30, n (%)	142 (5.1)	37 (3.2)	44 (3.5)
∼30–60, n (%)	657 (23.6)	191 (16.7)	276 (21.6)
∼60–90, n (%)	800 (28.%)	350 (30.5)	417 (32.7)
≥90 n (%)	1182 (42.5)	568 (49.6)	538 (42.2)
MAP at biopsy, median (IQR), mm Hg	96.7 (88.7–106.3)	93.3 (85.0–103.3)	93.3 (86.7–100.0)
Proteinuria at biopsy, median (IQR), g/d	1.2 (0.7–2.2)	1.3 (0.6–2.4)	1.2 (0.7–2.3)
<0.5, n (%)	383 (13.9)	221 (19.4)	217 (17.0)
∼0.5–1, n (%)	772 (28.1)	209 (18.3)	306 (24.0)
∼1–2, n (%)	817 (29.7)	352 (30.8)	370 (29.0)
∼2–3, n (%)	360 (13.1)	145 (12.7)	158 (12.4)
≥3, n (%)	415 (15.1)	215 (18.8)	224 (17.6)
MEST histologic score, n (%)
M1	1054 (38.0)	481 (42.0)	570 (44.7)
E1	478 (17.3)	476 (41.5)	385 (30.2)
S1	2137 (77.0)	912 (79.6)	768 (60.2)
T1	686 (24.7)	207 (18.1)	306 (24.0)
T2	128 (4.6)	122 (10.6)	112 (8.8)
RASB use, n (%)
At biopsy	862 (32.4)	320 (30.0)	926 (72.6)
After biopsy	2400 (86.7)	708 (66.4)	1164 (91.3)
Immunosuppressant use, n (%)
At biopsy	252 (9.1)	81 (7.1)	142 (11.1)
After biopsy	1209 (43.5)	359 (31.3)	432 (33.9)
Primary outcomea, n (%)
50% decline in eGFR	420 (15.1)	210 (18.3)	173 (13.5)
ESRD	372 (13.4)	155 (13.5)	110 (8.6)
Total primary outcomes	492 (17.7)	213 (18.6)	181 (14.2)

eGFR, estimated glomerular filtration rate; ESRD, end-stage renal disease; IQR, interquantile range; MAP, mean arterial pressure; MEST, M = mesangial hypercellularity, E = endothelial hypercellularity, S = segmental sclerosis, T = tubular atrophy or interstitial fibrosis; RASB, renin-angiotensin system blocker; SCr, serum creatinine.

Total primary outcomes defined as the first event of either 50% decline in eGFR or ESRD.

Regression on Linear Predictor in Validation Data

The calibration slopes of linear prediction were 0.87 and 0.89 for the full model without and with race, respectively. Although the slopes were smaller than those in the other cohorts, they were close to 1 and not significantly different from 1 (P > 0.05; Table 3), so the discrimination appeared to be preserved.

Table 3

Discrimination measures in the current and reported cohorts

Group	Regression slope on linear prediction	C-statistic	R2D (%)
Full model without race
Reported derivation cohort	—	0.81 (0.80–0.81)	25.3
Reported validation cohort	1.19	0.81 (0.80–0.82)	35.3
Our validation cohort	0.87 (0.73–1.00)	0.81 (0.80–0.81)	37.6
Full model with race
Reported derivation cohort	—	0.82 (0.81–0.82)	26.3
Reported validation cohort	1.12	0.82 (0.81–0.83)	35.3
Our validation cohort	0.89 (0.75–1.03)	0.81 (0.81–0.82)	42.2

Measures of Discrimination and Model Fit

When the reported models were applied directly to our current cohort, the C-statistic was 0.81 for both full models without and with race, with similar values in the reported cohorts. In addition, R2D values were 37.6% and 42.2% for the full model without and with race in our cohort, indicating an increase compared with ∼25% in the reported cohorts, suggesting good performance of the model fit (Table 3).

Comparison of Risk Groups

Figure 2 shows Kaplan-Meier curves according to risk groups based on the percentiles of the linear predictor (<16th for low-risk group [red], ∼16th to <50th for intermediate-risk group [green], ∼50th to <84th for higher risk group [blue], and ≥84th for the highest risk group [purple]). We found that the Kaplan-Meier curves in the risk groups were well separated, especially those for the highest risk group of the 2 full models, confirming our earlier conclusion that the models have preserved discrimination. By visual comparison of our validation results with the original publication, the discrimination was broadly similar, but the full model with race seemed less able to distinguish between the 2 lowest risk groups in our validation cohort. Accordingly, the hazard ratios between risk groups were well-maintained, confirming the impression in Figure 2. The predicted 5-year risks for patients in the 4 groups defined in our cohort were 2.51%, 5.37%, 15.57%, and 48.13% for the full model without race, and 5.22%, 10.25%, 25.81%, and 61.22% for the full model with race, respectively. Similarly, the eGFR decline slopes in the 4 groups were −0.67, −1.56, −2.27, and −3.32 for the full model with race, and −1.67, −1.87, −1.95, and −3.72 for the full model without race (Table 4). The demographic characteristics of the patients in the 4 risk groups based on the full model without/with race are presented in Supplementary Tables S2 and S3, respectively. In addition, we found that all predictor variables were worse with increasing risk, such as more proteinuria, more Oxford MEST lesions, and lower eGFR.

Figure 2

Kaplan-Meier curves for survival probability of primary outcome in 4 risk groups based on percentile of the linear predictor. Full model without race (a). Full model with race (b). The 4 risk groups were defined as <16th (low risk), ∼16th to <50th (intermediate risk), ∼50th to <84th (higher risk), and ≥84th (the highest risk) percentiles of the linear predictor from the full model without and with race, respectively.

Table 4

Hazard ratios, mean predicted 5-year risk of primary outcome, and rate of kidney function decline in subgroups based on linear predictor

Measure	Hazard ratio	Mean predicted 5-yr risk, %	eGFR decline slope
Full model without race
Low-risk group	Reference	2.51	−1.67
Intermediate-risk group	2.57 (1.16–5.72)	5.37	−1.87
Higher risk group	3.85 (1.75–8.46)	15.57	−1.95
Highest risk group	20.35 (9.29–44.57)	48.13	−3.72
Full model with race
Low-risk group	Reference	5.22	−0.67
Intermediate-risk group	1.33 (0.55–3.23)	10.25	−1.56
Higher risk group	2.97 (1.28–6.87)	25.81	−2.27
Highest risk group	8.29 (3.84–20.71)	61.22	−3.32

eGFR, estimated glomerular filtration rate.

Subgroups were <16th (low risk), ∼16th to <50th (intermediate risk), ∼50th to <84th (higher risk), and ≥84th (the highest risk) percentiles of the linear predictor from the full models without and with race, respectively.

Model Calibration

As shown in Figure 3, the Kaplan-Meier‒like estimates of baseline survival function estimated after fitting the model were similar to the reported ones. Model calibration performance was obtained by comparing observed and predicted risks of primary outcome over the duration of follow-up. For the full model without race (Figure 4a), the predicted curves (red) showed an acceptable fit with the observed curves (black). For the full model with race, however, calibration was better earlier on, whereas the predicted risk curves displayed a creep at 3 years in all risk groups (Figure 4b). Consistent with this, a comparison of 5-year predicted and observed survival probability of primary outcome is shown in Figure 5. Results at 1, 2, 3, and 4 years are presented in Supplementary Figures S1 and S2. Differences between observed and predicted survival probabilities of primary outcome over 3 years were apparent for the full model with race.

Figure 3

Estimates of the baseline survival function in the reported and the current data sets. Full model without race (a). Full model with race (b). Baseline survival function of the reported data set is shown in red, and the baseline survival function of the current data set is shown in blue.

Figure 4

Comparison of observed and predicted survival probability of primary outcome over the duration of follow-up. Full model without race (a). Full model with race (b). The 4 risk groups were defined as <16th (low risk), ∼16th to <50th (intermediate risk), ∼50th to <84th (higher risk), and ≥84th (the highest risk) percentiles of the linear predictor from the full model without and with race, respectively. Predicted risks were mean predicted risk calculated according to the reported models (in red), and observed risks were Kaplan-Meier estimated risk probability (in black).

Figure 5

Calibration curves depicting the predicted-vs.-observed survival probability of the 5-year primary outcome. Full model without race (a). Full model with race (b).

Discussion

The 2 reported full prognostic models without and with race provided us with a useful prediction tool for clinical IgAN progression. In this study, we further evaluated model performance using an external cohort from China and Argentina. For the full models without race and with race, good discrimination (C-statistic >0.80) was observed and the models fit well. The survival curves of patients stratified by percentiles of linear predictor were well separated. Our external validation provides further evidence that the clinical and pathologic variables used in the model appeared to be sufficient for patient discrimination. The full model without race showed acceptable calibration. Although the full model with race showed a similar regression slope on linear predictor (0.89), it seemed to overestimate the prognostic risk over 3 years.

Compared with the vigorous progress in model development, only about 25% (32 of 127) of them have been validated and few have been used in clinical practice. Although there is an increasing number of clinical prediction models for IgAN, most were performed in single ethnicity and few were approved to accurately identify high-risk patients.25, 26, 27, 28, 29, 30 Considering that the diagnosis of IgAN depends on renal biopsy, a prediction model with histologic variants would help increase the model accuracy. However, although there were some prediction models developed with pathologic variables, these models either developed with a relatively earlier cohort or in a single population. Moreover, some of these models used different pathologic scoring systems that are not widely used. For example, Goto et al derived a prediction model based on the Japanese population, with both clinical variables and pathologic variables that were not internationally committed., Chen et al used the XGBoost system and stepwise Cox regression to develop a prediction model, based on Chinese patients at different centers. The predictors included demographic, clinical, and pathologic variables. However, this model was based on single ethnic population, and external validation based on non-Asian population is not available. In this context, 2 full models integrating clinical variables and Oxford MEST histologic scores were derived and validated in 2 multiple ethnic cohorts. The well-established factors for disease progression of IgAN, including eGFR, proteinuria, blood pressure, Oxford MEST histologic scores, age, and use of RASB/immunosuppressant, could be “easy” and consistently obtained in clinical practice, demonstrating its potential in clinical practice.

In this study, we have further performed external validation of the full models in a Chinese-Argentinian cohort from a relatively “new era.” Although it was shown that 86% patients received RASBs after diagnosis in the reported derivation cohort, which is not that different from the 91% in our current cohort, the rate of immunosuppressant use between the reported and our current cohort were also similar after biopsy. Compared with the reported cohorts, there was a significantly higher rate of RASB initiation before diagnosis (30% vs. 70%). This strengthens the analysis because our current cohort is much more representative of patients receiving the current treatment regimens. We found that the prediction models consistently performed well for discrimination in our validation cohort. Our subgroup analysis suggests that survival curves of different risk groups were quite well-separated in both models. Accordingly, the eGFR decline slope was relatively larger with risk increasing. This was consistent with the worse predictor variables, such as more proteinuria, more Oxford MEST lesions, and lower eGFR, across the risk groups. In this way, the “simple, robust” models validated in our cohort could be applied to improve risk prediction, which was approved to both increase treatment allocation to patients at high risk of disease progression and avoid treatment in patients with nonprogressive disease. Moreover, for the full model without race, the calibration was acceptable. However, for the full model with race, it seemed to overestimate the prediction risk over 3 years in our cohort. Considering that, for a given high-risk clinical decision, a well-calibrated model providing a wider risk stratification is likely to have greater clinical utility, we suggest using the full model without race for further assessment of setting the thresholds, which is used to estimate the benefits and costs of specific interventions.

The strength of this study is that we did external validation based on a Chinese/Argentinian population from a relatively “new era,” which enabled us to validate the performance of the full models in the populations with treatment under current guidelines. However, there are also some limitations. First, our final cohort excluded patients who did not have a renal biopsy performed or for whom the Oxford MEST histologic scores were not available, meaning that we may have missed some very high-risk patients because the Oxford MEST histologic scores were hard to evaluate due to few glomeruli in the biopsy. Second, as there was a large proportion of Chinese patients in our validation cohort, the model performance in other new ethnic populations and recalibration of the full model with race are to be evaluated in the future. Third, a limitation of the prediction model is that, taking into account that IgAN is an entity (not a disease) with a long-term evolution, the model offers only short-term prognosis, up to 8 years at the most.

In summary, we externally validated the full prediction models to risk stratify patients after an initial diagnosis of IgAN. The prediction models showed good performance on personalized risk assessment, which will help allocate immunosuppression to those patients at high risk of disease progression and avoid treatment in those with nonprogressing disease.

Disclosure

All the authors declared no competing interests.