Predicting the future in immunoglobulin A nephropathy: a new international risk prediction tool

How can you give advice about prognosis to a patient recently diagnosed with immunoglobulin A nephropathy (IgAN)? How can you make immunosuppression treatment decisions without knowing the patient’s risk of disease progression? The challenge of accurate individual risk stratification in IgAN is a clinical problem regularly faced by nephrologists, and further exacerbated by the wide range of clinical outcomes ranging from complete remission to end-stage kidney disease (ESKD). Despite 50 years of study since IgAN was first described, until now we have had very limited capacity to address this problem. Most patients newly diagnosed with IgAN are young adults with pressing work and family responsibilities, who will inevitably ask their healthcare providers stark and poignant questions: will they develop kidney failure? And how quickly? Will immunosuppression reduce that risk? It is incumbent upon the nephrology community to do better in helping patients address these life-changing concerns. It is conventionally stated that around 25% of those with IgAN will reach ESKD within 25 years of diagnosis [1]. Although such numbers have the advantage of being easy to remember, they are based on an average outcome across heterogeneous research cohorts and thus are of little benefit to individual patients. Kidney outcome estimates are also heavily dependent on the case mix of the cohorts from which they are derived. For example, a cohort comprised predominantly of patients diagnosed with IgAN as a result of asymptomatic abnormalities found on urinalysis will have much better prognosis than cohorts presenting with more advanced kidney disease [2, 3]. There are many risk factors identified at the time of biopsy that have been associated with renal outcome in IgAN, including age, sex, proteinuria, estimated glomerular filtration rate (eGFR), blood pressure and the MEST-C histology score [4]. However, it has not been known how to combine these risk factors in individual patients in order to generate an accurate assessment of long-term renal prognosis. In a previous study, we showed that patients with low-risk proteinuria and eGFR levels but high-risk kidney biopsy features (as assessed by the MEST score) have a substantial 33% risk of subsequent decline in kidney function 10 years after biopsy [5]. This demonstrates the importance of combining clinical and histological risk factors together to determine more accurately the risk of disease progression in IgAN. To this end, several previous attempts have been made to generate prediction models that include multiple different risk factors, but these have often used pathology scoring systems, which are not widely accepted, require long periods of follow-up prior to risk stratification, use small cohorts of restricted ethnicity or have not been robustly validated [6–14]. We propose that the recently published International IgAN Risk Prediction Tool, developed by the International IgA Nephropathy Network, is an important step toward addressing all these limitations, and offers a robust tool for individual risk stratification that can be easily implemented in clinical practice [15]. The study leveraged an international collaboration of researchers to generate the largest IgAN cohort yet (n = 3297) with detailed longitudinal clinical data that are suitable for prediction modeling analysis. The cohort included Caucasian, Japanese and Chinese patients followed for >5 years, and was split into derivation (n = 2781) and validation (n = 1146) subgroups that had similar clinical and pathology characteristics although with some differences in multi-ethnic representation. A large majority of patients (81%) were receiving renin–angiotensin system (RAS) blockade at the time of biopsy or soon after, consistent with the KDIGO Glomerulonephritis guideline recommendations [16]. The primary study endpoint was a composite of ESKD (defined as GFR <15 mL/min/1.73 m2 or the need for renal replacement therapy) or a 50% reduction in eGFR from the time of biopsy. The derivation and validation cohorts had a similar risk of primary outcome 5 years after kidney biopsy (14.7 and 13.3%, respectively). Several risk prediction models were tested. A ‘clinical model’ included eGFR, proteinuria and blood pressure at the time of biopsy because these are the most consistent clinical predictors of outcome [4]. A full model, which included the clinical model variables and the MEST score in order to be consistent with our previous study [5], and allowed the data to select additional predictor variables that were independently associated with the primary outcome chosen among age, sex, self-reported race, body mass index, the presence of cellular or fibrocellular crescents, and immunosuppression and RAS blockade at or prior to the time of biopsy. The predictor variables that were preferentially chosen included age, race and use of immunosuppression and RAS blockade. Because the ethnicity of the cohorts was predominantly Caucasian, Chinese and Japanese, a second full model was created in the same way but without considering race as a predictor variable, so that it might be applicable in other ethnic groups. Therefore, the final predictor variables included in both full models were age, eGFR, proteinuria and blood pressure at the time of biopsy; the MEST score; the use of immunosuppression and RAS blockade at or prior to biopsy; and with or without self-reported race. All the prediction models were Cox proportional hazards models of time from kidney biopsy to the primary outcome censored at death or the end of follow-up, and were assessed using multiple different prediction modeling analytics that test goodness-of-fit, discrimination, reclassification and calibration. Goodness-of-fit reflects how well the model fits the data and was assessed using R2D, which is a measure of explained variation that accounts for censoring, and the Akaike information criterion (AIC), which must be compared between two models [17]. Larger values of R2,D and smaller values of the AIC suggest better model fit. Discrimination is the ability of the model to differentiate between patients that did or did not experience the primary outcome, which was assessed using the C-statistic with higher values suggesting better discrimination [17]. Reclassification improvement between two models is a relative comparison of the ability to move patients who experience the primary outcome up a risk category and move patients who do not experience an outcome down a risk category [17]. This was assessed using the net reclassification improvement and integrated discrimination improvement, with values above zero indicating better reclassification. Calibration is the agreement between observed and predicted risk values, which was visually assessed using a calibration plot in which predicted values closer to observed values indicate better calibration [17]. These metrics assess different aspects of prediction performance; all have limitations, and none should be considered on its own. However, a minimum requirement is that a risk prediction model should be well calibrated, and it is critically important to assess prediction performance in a sufficiently large external validation cohort [18–20]. In the derivation analysis, both full risk prediction models with or without race demonstrated an improvement in all measures of prediction performance compared with the clinical model, as assessed by the R2D, AIC, C-statistic, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) (see Table 1). Calibration plots for both full models demonstrated excellent calibration, with very similar predicted and observed risks across the full spectrum of risk of disease progression. When compared with each other, there was no prediction benefit seen for the full model with race compared with the model without race. In the external validation cohort, both full models had better R2D (both 35.3%) than was seen in the derivation cohort, with similar C-statistics [0.82, 95% confidence interval (CI) 0.81–0.83; and 0.81, 95% CI 0.80–0.82], and maintained excellent calibration. These results therefore demonstrated that both full models better predict the risk of a 50% decline in eGFR or ESKD compared with the clinical model, with robust external validation in a sufficiently large and separate cohort from that in which the models were derived. This modeling strategy demonstrates that the complete repertoire of clinical factors, demographics and MEST histology scores contained in the full models significantly improved risk stratification in IgAN compared with using only measures of eGFR, proteinuria and blood pressure. Measures of prediction performance in the derivation analysis for the clinical and full models with and without race [15] Measures of prediction performance in the derivation analysis for the clinical and full models with and without race [15] The International IgAN Prediction Tool models have complex formulae with several interaction terms, so a mobile app calculator was developed by QxMD (available for free on Calculate by QxMD through the app store for iOS and Android). A web-based version of the calculator is also available at https://qxmd.com/calculate-by-qxmd. To use the calculators, the user must select the time horizon up to 7 years after biopsy at which risk prediction is desired, and enter values for all the predictor variables, including age; race; blood pressure, eGFR and proteinuria at biopsy; MEST score; and use of RAS blockade and immunosuppression. The importance of crescents independently predicting disease progression in IgAN has been controversial. A study by Haas et al. [21] in 2017 evaluated this issue in 3096 patients from Japan, China, Europe and North America. In the overall cohort, the presence of cellular or fibrocellular crescents was not associated with a faster rate of eGFR decline. However, they were associated with an increased risk of a 50% decline in eGFR or ESKD [hazard ratio (HR) 1.37, 95% CI 1.07–1.75], an effect that was more pronounced in patients not treated with immunosuppression (HR 1.51, 95% CI 1.13–2.02) and in those with crescents involving ≥25% of glomeruli (HR 2.29, 95% CI 1.35–3.91). Importantly, these results were adjusted for baseline eGFR, the MEST score and time-averaged proteinuria and blood pressure, but not other predictor variables such as age, race or use of RAS blockade. Based on these results, although crescents were considered as a candidate predictor, it is likely that they were not selected for retention in the International IgAN Prediction Tool for several reasons. First, the prediction tool included multiple variables not contained in the adjusted analyses done in the Haas study. Second, crescents were highly correlated with race such that the inclusion of race in the models diminished the predictive value of crescents. Third, the association of crescents with the primary outcome was confounded by the use of immunosuppression during follow-up. Because the prediction tool was designed for use at the time of biopsy, it would not have been possible to include subsequent immunosuppression exposure in the models, which resulted in further attenuation of the observed predictive value of crescents. Finally, the datasets used for the prediction tool analysis did not contain details on the percentage of glomeruli involved with crescents. It remains possible that a higher burden of crescents in ≥25% of glomeruli may further improve prediction; however, this requires additional research. There are several limitations to the prediction tool that need to be considered. Because the derivation and validation cohorts included only adults, the prediction tool should not be used in children until it has been validated in a pediatric cohort. Similarly, the prediction tool should not be used in IgA vasculitis, or in presumed but not biopsy-confirmed cases of IgAN. The predictor variables were constrained to be within 6 months of kidney biopsy, so the prediction models should not be applied at later time points in the disease trajectory. Although the derivation and validation cohort were multi-ethnic, they were dominated by Caucasian, Japanese and Chinese representation. Further validation will be required in other ethnic groups. Because of the follow-up duration of the analytic cohorts, the prediction tool should not be used to predict renal outcome beyond 7 years after biopsy. This needs to be considered in the context of IgAN as a slowly progressive disease in which it may be desirable to quantify the risk of disease progression over longer time horizons. Importantly, the models were designed to predict the risk of renal function decline and not to predict response to immunosuppression. These two outcomes may be distinctly discordant; for example, interstitial fibrosis may be a risk factor for disease progression but may not respond to immunosuppression. Therefore, at the current time, the prediction tool should not be used to determine immunosuppression treatment decisions. Now that we have the International IgAN Prediction Tool, what is the best approach for future research to further improve risk stratification? High-quality databases in IgAN are rare, resource intensive to generate and highly dependent on the generosity of patients who participate. As such, the nephrology community has an obligation to use these databases as effectively as possible. Further research to generate de novo prediction models in smaller single-ethnicity cohorts may not achieve this goal. A recent study used a novel machine learning approach to derive and validate a prediction tool in two Chinese cohorts (n = 1022 and 1025) [22]. This model does not address all the limitations of previous risk prediction studies in which it is limited to a single-ethnic group, did not assess or validate model calibration, was derived using very few outcome events and used predictor variables (such as quantification of urine red blood cell concentration) that might not be accurately available in all jurisdictions [23]. A more effective strategy to utilize new research cohorts may instead be to further validate the available International IgA Prediction Tool models, update the parameters in the models or assess the addition of new predictor variables that are more specific to local circumstances. This can be done with more certainty using smaller datasets compared with generating an entirely new model [24]. The global research community is more likely to further advance risk stratification in IgAN in a coordinated manner by sequentially improving on the existing prediction models in order to address their limitations so that they can be applied with confidence to improve patient care. There are several interesting future applications of the International IgAN Prediction Tool beyond simple risk stratification. Clinical trial design in glomerular diseases is complicated by the feasibility of recruiting sufficient patients across multiple sites to power a study capable of observing renal outcome events over a reasonable timeframe at an acceptable cost [25]. This challenge is demonstrated by the supportive versus immunosuppressive therapy of progressive IgA nephropathy (STOP-IgAN) trial [26], which was underpowered to detect an effect on the primary renal function-based outcome because only 27% of patients in the control group experienced a decline in eGFR ≥15 mL/min/1.73 m2. Accurate risk stratification using the prediction tool offers the opportunity to enrich future clinical trial cohorts with higher risk patients who are more likely to experience eGFR decline during the trial period, thereby improving study power. However, this will require that clinical trials recruit patients close to the time of biopsy so that histological and clinical predictor variables can be used in the prediction tool. The fully specified prediction models are publicly available and can be used to validate novel biomarkers for disease progression by providing a baseline model with clinical and histological predictor variables to which a biomarker can be added to test for additional prediction benefit. This requires much smaller cohorts than de novo model development, thereby improving the feasibility of biomarker research with clinical validation. The prediction tool offers an opportunity to improve immunosuppression treatment decisions by targeting therapy to patients at highest risk of disease progression using personalized risk stratification instead of simplistic categories of proteinuria and eGFR as is currently the standard of care [16]. Although this will require additional research in better understanding the risks versus benefits of immunosuppression (which are currently uncertain) and the impact of specific treatments on patient quality of life, the International IgAN Prediction Tool is the first step toward a longer term goal of achieving precision medicine treatment decisions in IgAN. Although there are limitations to the International IgAN Prediction Tool with room for future improvement, it is more effective than any existing approach in predicting at the time of biopsy the risk of disease progression over 5–7 years. We encourage nephrologists to download the tool now for use in clinical practice while waiting for future publications that may further refine its role and relevance. None declared.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.