Risk Prediction for Renal Cell Carcinoma: Results from the European Prospective Investigation into Cancer and Nutrition (EPIC) Prospective Cohort Study

Kidney cancer causes more than 175,000 deaths worldwide each year, with approximately 400,000 new diagnoses recorded annually (1). Over 80% of these are renal cell carcinoma (RCC). When identified at an early, localized stage, patients with RCC are commonly cured following nephron-sparing nephrectomy, but patients with locally advanced tumors or with regional nodal and/or distant metastases have poor prognosis (2). Most tumors diagnosed at an early stage are asymptomatic and are incidentally detected. Taken together, these facts indicate that screening and early detection for RCC could have a major impact on RCC outcomes. There is currently no recommended screening practice for RCC, except for individuals with a genetic predisposition (3). Given that sporadic RCC is relatively rare, the risk of screening-detected false positives is substantial. In addition to this, screening could also lead to overdiagnosis of renal masses that may not have presented clinically in the lifetime of the patient. Any population screening for RCC would therefore need to focus on a subset of individuals with a priori greater risk of clinically meaningful RCC. A powerful risk prediction model for RCC could be used to identify such high-risk individuals.

There are several established risk factors for RCC, which could be applied in the prediction modeling context. Aside from age, the principal established risk factors for RCC are obesity, smoking, and sex. High body mass index (BMI) is consistently associated with greater risk, with each kg/m² increment associated with approximately 1.05-fold higher rates of RCC incidence (4–7). Current and former tobacco smokers are also at greater risk of RCC, with relative risks compared with never smokers of approximately 1.15 and 1.4, respectively (8). Men have higher incidence of RCC than women, and we have shown that male sex is consistently associated with two-fold higher incidence of RCC worldwide and across all ages and time periods (9). In addition, history of hypertension/high blood pressure is associated with greater risk of RCC (6, 10, 11).

To our knowledge, there exists one published model for estimating absolute risk of RCC in the general population based on established risk factors (12). This model included age, sex, smoking status, and BMI. We evaluated the performance of this model in a large European cohort, and developed and evaluated a more comprehensive model including additional established risk factors: blood pressure and history of hypertension.

We used data from the European Prospective Investigation into Cancer and Nutrition (EPIC). EPIC is an ongoing, multicenter prospective cohort study that recruited 521,330 participants between 1992 and 2000 from 10 European countries. The current study involved participants from eight countries (Denmark, France, Germany, Italy, the Netherlands, Spain, Sweden, and the United Kingdom). Details of recruitment procedures and data collection have been described previously (13). Cancer cases and deaths were ascertained via linkage to population registries or active follow-up, depending on the study center. Participants were followed from recruitment until a first diagnosis of RCC (ICD-10 code C64), death, or end of follow-up (the most recent follow-up linkage was in 2015), whichever occurred first. The EPIC study protocol was approved by the IARC research ethics committee and local committees at each recruiting center.

We used a flexible parametric survival model to estimate log hazard ratio (HR) as well as a parametric estimate of the log baseline cumulative hazard (14). The baseline cumulative hazard was modeled as a function of log follow-up time using a restricted cubic spline basis with a single internal knot at 9 years, and boundary knots at 0.05 and 15 years. The predictors included in the model were selected a priori as established risk factors for RCC: age (modeled using a restricted cubic spline basis with boundary knots at 40 and 70 years, and internal knots at 50 and 60 years), sex (female vs. male), BMI (per 1 kg/m², measured at baseline), smoking status (never, former, current), self-reported history of hypertension (yes vs. no), systolic blood pressure (per 1 mmHg), and diastolic blood pressure (per 1 mmHg). Blood pressure was measured with either a sphygmomanometer or oscillotonometer, independently of self-reported hypertension or any treatment for hypertension. The 10-year predicted cumulative hazard H(10) was obtained directly from the fitted equation, and the 10-year probability of RCC diagnosis was calculated as P(10) = 1−S(10) = 1−exp[−H(10)].

We assessed the calibration of both the EPIC model and the model previously published by Usher–Smith and colleagues (12). For the Usher–Smith model, we calculated the 10-year probability of RCC using their model. For the EPIC model, we conducted leave-one-country-out cross-validation, where we fitted the model to data from seven countries, and used this model to make predictions for the remaining country that had not been used in model fitting, repeating this process until out-of-sample predictions were available for each of the eight countries. To obtain a smooth estimate of the observed probability, we fitted secondary flexible parametric survival models, using a restricted cubic spline basis function of the predicted probabilities with four degrees of freedom. Estimates from these calibration models (“observed”) along with their 95% confidence intervals were plotted against the model predicted probabilities (“predicted”) to provide a visual depiction of model calibration. These models were also used to provide recalibrated probabilities for the Usher–Smith model.

Discrimination was quantified by the C-statistic, which is the probability that for any randomly selected pair of participants, the participant with the shorter disease-free follow-up time will have a higher probability of disease. For the EPIC model, we calculated the observed C-statistic (based on the whole dataset) as well as the bootstrap optimism corrected C-statistic (based on fitting the model to 500 bootstrap samples of the data, and making predictions using the full dataset). This provides an estimate of the out-of-sample discrimination (15).

Among 192,172 participants with complete covariate data (Fig. 1), we identified 588 incident RCC cases during a median follow-up of 15 years (median time-to-diagnosis among cases of 9 years). The distribution of the predictors is presented in Table 1.

Fitted parameters of the EPIC RCC prediction model are presented in Table 2. The hazard of RCC was greater with increasing age, male sex, higher BMI, former and current smokers relative to never smokers, participants with a history of hypertension, and with higher measured blood pressure. Details describing how to use the fitted model to calculate individual absolute risk of RCC are presented in the Supplementary Materials.

The Usher-Smith model was well calibrated when applied to the EPIC cohort (Fig. 2A), with predicted probabilities slightly lower than observed over the range of predictions. Leave-one-country-out cross-validation indicated that the EPIC model was also well calibrated, with some suggestion of miscalibration for the highest predicted probabilities (Fig. 2B).

The Usher–Smith model had good discrimination in the EPIC data (C-statistic [95% CI]: 0.699 [0.679–0.721]). The EPIC model had slightly higher discrimination (0.714 [0.694–0.735], bootstrap optimism corrected C-statistic: 0.709). The distribution of predicted probabilities from EPIC had greater variance than that of the Usher–Smith model (SD 0.00172 vs. 0.00149, Fig. 2C). Despite this, the difference between probabilities from the EPIC and Usher–Smith models was small for the majority of participants (Fig. 2D). The absolute difference in probabilities between the models was < 0.001 for 84% of participants, with less than 1% of participants having a difference of 0.005 or more (Table 3).

The distribution of predicted probabilities from the EPIC model is presented in Table 4. The vast majority of participants had predicted 10-year risk less than 1%. Only 30% of participants had a predicted 10-year risk of 0.0025 or greater.

We showed that a previously developed RCC prediction model (Usher–Smith; ref. 12) incorporating age, sex, smoking status, and BMI performed well in terms of discrimination and calibration. We also showed that the EPIC model, augmented with blood pressure and history of hypertension, slightly improved discrimination. Leave-one-country-out validation and bootstrap optimism corrected C-statistics indicate that the EPIC model will perform well when applied to new observations.

The discrimination of both the EPIC and Usher–Smith models is not substantially lower than the Pooled Cohort Equations or Qrisk which are used in practice for prediction of cardiovascular diseases (16, 17); however, compared with cardiovascular diseases, the 10-year risk of RCC is relatively low. In the context of low expected risk, any screening test will need to have very high specificity for detection of tumors that would have presented clinically in the lifetime of the patient. Focused renal ultrasound has been proposed as a screening modality for RCC with a specificity possibly as high as 0.98 (18, 19). Even if screening via ultrasonography was restricted to people predicted to have a 10-year probability of RCC ≥ 0.0025 (30% of EPIC participants meet this criteria), we would expect approximately eight false positives to be detected for every RCC diagnosis in the EPIC cohort. While this is a substantial improvement over the approximately 13 false positives we would expect per RCC diagnosis if screening was not targeted, it would still represent a substantial burden on health systems and many thousands of individuals. Further restricting screening to individuals with predicted 10-year probability ≥ 0.005 (approximately 5% of EPIC participants), we would expect approximately four false positives for every RCC diagnosis in the EPIC cohort.

Clearly, the utility of risk models will depend on the relative costs and potential harms of the subsequent procedures. For example, if the next step after focused renal ultrasound is further imaging, then false positives will be less of an issue. On the other hand, if the next step is a more invasive procedure such as renal biopsy, a risk model with a high positive predictive value will be required. In this case, practical and feasible screening for RCC is likely to require either more powerful risk stratification tools or screening/detection methods with near-perfect specificity.

One approach to improving risk stratification would be to incorporate biomarkers of risk or disease into the models. Promising urinary biomarkers for RCC include aquaporin 1 (AQP1) and perilipin 2 (PLIN2), which have been shown to be elevated in patients with RCC compared with healthy controls, as well as patients with non-RCC renal masses, prostate cancer, or bladder cancer (20). Kidney injury molecule 1 (KIM-1) has been evaluated as both a urinary and circulating biomarker. Urinary KIM-1 is elevated in patients with RCC compared with healthy controls and patients with prostate cancer, but may suffer from poor specificity to differentiate RCC from benign renal masses (21). In a previous study, we prospectively evaluated circulating concentrations of KIM-1 using data and samples from the EPIC cohort, showing that prediagnostic circulating KIM-1 is strongly predictive of 5-year RCC risk (22). Future studies investigating these or novel candidate biomarkers should evaluate their performance against the predictive power of risk models such as the EPIC RCC prediction model presented here.

Our study has several strengths. We used data from a large, multicenter Eurpoean prospective cohort study with complete follow-up for cancer incidence and death, and could include baseline measurements of BMI and blood pressure in the model. Our approach to validating the model through leave-one-country-out cross-validation and bootstrap optimism correction provide confidence that our model will perform well when applied to new observations. One limitation of our study is that RCC is composed of different types that have specific genetic and histologic characteristics (23). Different tumor types might have specific risk factor profiles, but most studies have analyzed RCC as a single entity. As risk factors become established for RCC subtypes, it may be possible to enhance the predictive ability of the models. Another limitation is that we lack data on incidence of benign renal masses in the cohort. It has been estimated that approximately 15% of resected renal masses are benign or indolent (24). It is possible that screening could lead to increased detection of benign renal masses, with associated unnecessary treatment of disease unlikely to present clinically in the lifetime of the patient. If benign renal masses and RCC have shared risk factors, then risk-based screening would lead to greater detection of both RCC and benign renal masses. Future studies are needed to investigate whether the models are sufficiently specific to clinically meaningful RCC—and can differentiate between this risk and the risk of benign renal masses—before the models can be applied for risk stratification. Models with high specificity for clinically meaningful RCC would not only reduce overdiagnosis, but would substantially reduce the required sample size for screening efficacy trials.

In summary, we evaluated the performance of a risk prediction model for RCC, and developed a more comprehensive model. Both models perform well in terms of calibration, and the updated EPIC model has slightly improved discrimination. Further work is needed to incorporate new, strong predictors into models before risk-stratified screening for sporadic RCC will be feasible.

B. Ljungberg reports personal fees from Novartis, Ipsen, BMS, Janssen, and MSD outside the submitted work. R.T. Fortner reports grants from German Cancer Aid and German Federal Ministry of Education and Research during the conduct of the study. M.B. Schulze reports grants from German Federal Ministry of Education and Research (BMBF) and German Cancer Aid during the conduct of the study. R.C. Travis reports grants from Cancer Research UK during the conduct of the study. D.C. Muller reports grants from Cancer Research UK during the conduct of the study and grants from NIH/NCI outside the submitted work. No disclosures were reported by the other authors.

The funders had no role in the design and conduct of the study; the collection, analysis, and interpretation of the data; the preparation, review, and approval of the manuscript; or the decision to submit the manuscript for publication.

R.K. Singleton: Formal analysis, investigation, methodology, writing-original draft. A.K. Heath: Supervision, investigation, writing-review and editing. J.L. Clasen: Investigation, writing-review and editing. G. Scelo: Investigation, writing-review and editing. M. Johansson: Investigation, writing-review and editing. F. Le Calvez-Kelm: Investigation, writing-review and editing. E. Weiderpass: Investigation, writing-review and editing. F. Liedberg: Investigation, writing-review and editing. B. Ljungberg: Investigation, writing-review and editing. J. Harbs: Investigation, writing-review and editing. A. Olsen: Investigation, writing-review and editing. A. Tjønneland: Investigation, writing-review and editing. C.C. Dahm: Investigation, writing-review and editing. R. Kaaks: Investigation, writing-review and editing. R.T. Fortner: Investigation, writing-review and editing. S. Panico: Investigation, writing-review and editing. G. Tagliabue: Investigation, writing-review and editing. G. Masala: Investigation, writing-review and editing. R. Tumino: Investigation, writing-review and editing. F. Ricceri: Investigation, writing-review and editing. I.T. Gram: Investigation, writing-review and editing. C. Santiuste: Investigation, writing-review and editing. C. Bonet: Investigation, writing-review and editing. M. Rodriguez-Barranco: Investigation, writing-review and editing. M.B. Schulze: Investigation, writing-review and editing. M.M. Bergmann: Investigation, writing-review and editing. R.C. Travis: Investigation, writing-review and editing. I. Tzoulaki: Investigation, writing-review and editing. E. Riboli: Resources, investigation, writing-review and editing. D.C. Muller: Formal analysis, supervision, funding acquisition, investigation, methodology, writing-original draft, project administration.

The authors thank all participants of the EPIC study for their ongoing contributions. The authors also acknowledge the National Institute for Public Health and the Environment (RIVM), Bilthoven, the Netherlands, for their contribution and ongoing support of the EPIC Study. D.C. Muller is supported by a Cancer Research UK Population Research Fellowship (C57955/A24390). The coordination of EPIC is financially supported by the European Commission (DG-SANCO) and the International Agency for Research on Cancer (Lyon, France). The national cohorts are supported by Danish Cancer Society (Denmark); Ligue Contre le Cancer, Institut Gustave Roussy, Mutuelle Générale de l'Education Nationale, Institut National de la Santé et de la Recherche Médicale (INSERM) (France); German Cancer Aid, German Cancer Research Center (DKFZ), Federal Ministry of Education and Research (BMBF), Deutsche Krebshilfe, Deutsches Krebsforschungszentrum and Federal Ministry of Education and Research (Germany); Associazione Italiana per la Ricerca sul Cancro-AIRC-Italy and National Research Council (Italy); Dutch Ministry of Public Health, Welfare and Sports (VWS), Netherlands Cancer Registry (NKR), LK Research Funds, Dutch Prevention Funds, Dutch ZON (Zorg Onderzoek Nederland), World Cancer Research Fund (WCRF), Statistics Netherlands (the Netherlands); Health Research Fund (FIS-ISCIII), Regional Governments of Andalucía, Asturias, Basque Country, Murcia, Navarra, and the Catalan Institute of Oncology (Barcelona) (Spain); Swedish Cancer Society, Swedish Research Council and County Councils of Skåne and Västerbotten (Sweden); Cancer Research UK (14136 to EPIC-Norfolk; C570/A16491 and C8221/A19170 to EPIC-Oxford), Medical Research Council (1000143 to EPIC-Norfolk, MR/M012190/1 to EPIC-Oxford) (United Kingdom).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.