- Email: [email protected]
Can we really predict risk of cancer?
G Model
CANEP-565; No. of Pages 4 Cancer Epidemiology xxx (2013) xxx–xxx
Contents lists available at SciVerse ScienceDirect
Cancer Epidemiology The International Journal of Cancer Epidemiology, Detection, and Prevention journal homepage: www.cancerepidemiology.net
Review
Can we really predict risk of cancer?§ Aaron P. Thrift, David C. Whiteman * Population Health Department, Queensland Institute of Medical Research, Queensland, Australia
A R T I C L E I N F O
A B S T R A C T
Article history: Received 27 February 2013 Accepted 4 April 2013 Available online xxx
Background: Growing awareness of the potential to predict a person’s future risk of cancer has resulted in the development of numerous algorithms. Such algorithms aim to improve the ability of policy makers, doctors and patients to make rational decisions about behaviour modification or surveillance, with the expectation that this activity will lead to overall benefit. There remains debate however, about whether accurate risk prediction is achievable for most cancers. Methods: We conducted a brief narrative review of the literature regarding the history and challenges of risk prediction, highlighting our own recent experiences in developing tools for oesophageal adenocarcinoma. Results and conclusions: While tools for predicting future risk of cardiovascular outcomes have been translated successfully to clinical practice, the experience with cancer risk prediction has been mixed. Models have now been developed and validated for predicting risk of melanoma and cancers of the breast, colo-rectum, lung, liver, oesophagus and prostate, and while several of these have adequate performance at the population-level, none to date have adequate discrimination for predicting risk in individual patients. Challenges of individual risk prediction for cancer are many, and include long latency, multiple risk factors of mostly small effect, and incomplete knowledge of the causal pathways. ß 2013 Published by Elsevier Ltd.
Keywords: Neoplasms Risk assessment Risk factors Esophageal neoplasms Risk reduction behavior
Clinical decision-making is generally based on estimating average risks within population subgroups and applying the same strategy for all within these groups. However, doctors want to know which diseases are most likely to affect their individual patients. Equally important, a patient wants to know if they have higher than average risk of certain diseases. If doctors can tailor prevention and treatment strategies for a patient based on their absolute risk for a disease, derived from the individual’s specific characteristics (e.g., a panel of factors including age, sex, body mass index, and diet), there is potential for better clinical outcomes [1]. This is the aim of the term ‘personalized medicine’. To fully realize personalized medicine, the scientific community has synthesized information about established risk factors (clinical, environmental and genetic) for specific diseases into statistical models for risk prediction [2–4]. These models use information on multiple risk factors to estimate absolute risk. Risk models are generally developed by selecting the most important panel of predictive factors from among a larger list of candidate
§ Grant support: APT is supported by the Cancer Council NSW STREP grant 08-04. DCW is supported by a Future Fellowship from the Australian Research Council (FT0990987). * Corresponding author at: Cancer Control Group, Queensland Institute of Medical Research, Locked Bag 2000, Royal Brisbane Hospital, Queensland 4029, Australia. Tel.: +61 7 33620279; fax: +61 7 38453502. E-mail addresses: [email protected], [email protected] (D.C. Whiteman).
risk factors. Each predictor is then assigned a relative weight in a combined risk score [2,4]. Individual risk profiles can then be generated and used to identify those at high risk of having (diagnosis) or developing (prognosis) a particular disease. The models have various applications, including use in development of health policy, education, clinical decision-making, and designing future research (e.g., establishing eligibility criteria for intervention and screening trials) [5]. They are not however intended to replace doctors’ clinical decision-making, but rather allow for more objective estimates of risk and uniform decision-making across different centers [6]. Risk prediction first received attention in the cardiovascular literature. One of the most widely validated and used prediction tools is the Framingham Risk Score (FRS) [7]. The FRS was designed to predict 10-year absolute risk for coronary heart disease. Absolute risk assessment charts based on the FRS and subsequent risk scores [8–12] are now included with many cardiovascular disease prevention guidelines [13–20]. Absolute risk estimates are used to guide the management of high-risk patients for cardiovascular disease and have improved the efficacy of medical interventions. The overall aim is to see their use translated into lower incidence and mortality from cardiovascular disease. However, decades after first being recommended, use of the absolute risk assessment charts in clinical practice tends to be highly variable. There remain many barriers to overcome (e.g., time constraints, over simplification of risk assessment, and overmedication) before they are routinely used [21–24].
1877-7821/$ – see front matter ß 2013 Published by Elsevier Ltd. http://dx.doi.org/10.1016/j.canep.2013.04.002
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002
G Model
CANEP-565; No. of Pages 4 2
A.P. Thrift, D.C. Whiteman / Cancer Epidemiology xxx (2013) xxx–xxx
The first risk prediction model for cancer was described by Gail et al. in 1989 [25]. The Gail breast cancer model was developed using data from a nested case-control study (the Breast Cancer Detection and Demonstration Project) and predicted risk of developing invasive or in situ breast cancer in a defined age interval. The model was modified in 1999 to predict absolute 5year risks for invasive breast cancer only [26]. The Gail breast cancer models included terms for age, hormonal or reproductive history, previous history of breast disease, and family history. Investigators have since further modified the original Gail model by adding modifiable risk factors (e.g., body mass index), terms for risk biomarkers (e.g., breast density) and multiple genetic variants, and by developing racial and ethnic specific models [27–29]. The National Cancer Institute (NCI) recognizes risk prediction as an ‘‘area of extraordinary opportunity’’ [30] and cancer risk prediction models are increasingly common in the medical literature. Population-based models have now been developed and validated for predicting risk of colorectal cancer [31], melanoma [32–36], lung cancer [37–41], liver cancer [42,43], and prostate cancer [44,45]. Many have associated online risk calculators available on the NCI’s website (http://epi.grants.cancer.gov/cancer_risk_prediction/). More and more of these traditional models (i.e., those with only clinical and epidemiologic risk factors) are being modified to incorporate biologic and genetic data to estimate cancer risk more accurately. It is difficult however to measure the value of these models. While most claim statistically significant results, few models have translated into clinically useful tools. Before being used clinically, a prediction model must be shown to provide accurate and generalisable risk predictions. The accuracy and generalisability of a model is assessed using measures of discrimination (how it performs at the individual-level) and calibration (populationlevel) [2,4]. These should be assessed in a similar, but external population to that in which the model was developed [3]. However, it is not always possible to find a suitable external validation dataset, and thus, models continue to be validated within the development dataset using various techniques (e.g., bootstrapping, and cross-validation) [2,3]. Several breast cancer models are currently used in clinical practice. In the United States, absolute risk estimates derived from the Gail model are used to identify high-risk women for mammography screening [46] and interventions trials [47]. However, while well calibrated at the population-level, the breast cancer models lack adequate discrimination in predicting risk for individuals [27,28,48]. We recently developed a model to estimate absolute risk for esophageal adenocarcinoma (EAC) [49]. EAC is a relatively rare cancer but is the fastest rising cancer in Western populations [50,51] and has poor survival [52]. EAC has a number of clearly established risk factors with reasonably large effect sizes (male sex, white ethnicity, reflux, obesity, and tobacco smoking) and thus ought to lend nicely to risk prediction [53]. However, despite the identification of ‘high-risk phenotypes’, it remains the case that very few people with one or more of these factors develop EAC. For instance, the absolute 5-year risk for EAC among obese, white, 50year-old males with reflux remains only 0.04% [49], and more than 40% of EAC patients do not report reflux symptoms at diagnosis [54]. It comes as no surprise then that our model had only moderate discriminatory accuracy when internally validated. That is, it failed to effectively differentiate between those who will develop EAC and those who will not. Why are we seeing this time and again for these cancer risk prediction models? Like other cancers, there are a number of issues as to why we may not be able to estimate EAC risk accurately at the individual level. Firstly, the primary risk factors for EAC are common in the population (for example, up to 20% of the adult population in
Western countries suffer from weekly reflux symptoms [55], and approximately 30% are obese [56,57] and 20% are current smokers [58,59]) and are therefore prevalent in individuals who will not develop the disease. Secondly, while the relative risks for these factors are consistent across populations, and statistically significantly associated with EAC risk, the effects are typically modest (relative risks < 5) [60]. As a general rule of thumb, risk factors require relative risks greater than 10 to be good predictors of individual risk, although this does not guarantee high discriminatory accuracy [61,62]. Whilst knowledge of these factors helps advance our understanding of disease mechanisms, these risk factors alone do not accurately discriminate those who will develop EAC from those who will not. Thirdly, most cancers have long latent periods and arise in individuals with risk close to the population average. Therefore, identifying high-risk groups for EAC on the basis of a probability threshold (e.g., using 5-year absolute risks) with high sensitivity is difficult. For rare cancers such as EAC, it may not be cost-effective to use risk prediction models in clinical practice. For example, even predictors with high sensitivity and specificity can still have low positive predictive values for rare diseases. That is, the number of people who actually develop cancer as a proportion of those predicted to do so will be low. As such, even if we had predictors with high discriminatory power (i.e., the test can correctly classify most people as ‘affected’ or ‘not affected’), most people exposed to the risks of further investigation or treatment will not see any benefit. Therefore, such a model appears well suited for population prevention strategies, but may not be able to identify individuals at high (or low) risk in a clinically meaningful and cost-effective way. With extensive research efforts already spent trying to improve current cancer models, is it possible to improve their clinical utility and move towards personalized medicine? Genetic risk profiling, in particular the addition of genetic variants to existing cancer risk models (as well as gene gene and gene environment interaction terms), has been proposed as a solution to increase discriminatory ability. However, it has been shown that, even when alleles with modest effect sizes are combined, the addition of genetic variants to these models has added only modestly to discrimination [29,63]. Thus, integrating genetic risk profiles to traditional models does not appreciably improve performance, at least with current knowledge [64]. Perhaps there are other risk factors (with high relative risks) not yet identified that, if added to existing cancer model, might add substantially to predictive ability over and above currently used risk factors. This seems unlikely for most cancers. One of the major lessons that we have learnt from risk prediction is that our knowledge of disease pathways is imperfect. Assuming we had a validated cancer risk prediction model and before widespread implementation, we need robust external validation, and of course, we need evidence that these tools actually improve patient outcomes when used clinically. Such evidence is best sourced from randomized trials, which are not without challenges. The goal of personalized medicine, at least in terms of predicting risks for the purposes of targeted prevention, may therefore be illusory. Recent experience with cancer risk prediction adds a layer of complexity to the ‘prevention paradox’ described by Geoffrey Rose more than 30 years ago, in which he made the observation that the bulk of disease arises in people with low risk [65]. The most effective strategies might be to target prevention at the entire population, and then put efforts into better diagnosis and better treatment for those who do develop illness. Conflict of interest None declared.
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002
G Model
CANEP-565; No. of Pages 4 A.P. Thrift, D.C. Whiteman / Cancer Epidemiology xxx (2013) xxx–xxx
References [1] Cooney MT, Dudina A, D’Agostino R, Graham IM. Cardiovascular risk-estimation systems in primary prevention: do they differ? Do they make a difference? Can we see the future? Circulation 2010;122:300–10. [2] Moons KG, Kengne AP, Woodward M, Royston P, Vergouwe Y, Altman DG, et al. Risk prediction models. I. Development, internal validation, and assessing the incremental value of a new (bio)marker. Heart 2012;98:683–90. [3] Moons KG, Kengne AP, Grobbee DE, Royston P, Vergouwe Y, Altman DG, et al. Risk prediction models. II. External validation, model updating, and impact assessment. Heart 2012;98:691–8. [4] Steyerberg EW. Clinical prediction models: a practical approach to development, validation and updating. New York: Springer, 2009. [5] Freedman AN, Seminara D, Gail MH, Hartge P, Colditz GA, Ballard-Barbash R, et al. Cancer risk prediction models: a workshop on development, evaluation, and application. J Natl Cancer Inst 2005;97:715–23. [6] Moons KG, Royston P, Vergouwe Y, Grobbee DE, Altman DG. Prognosis and prognostic research: what, why, and how? Br Med J; 2009;338. [7] Kannel WB, McGee D, Gordon T. A general cardiovascular risk profile: the Framingham Study. Am J Cardiol 1976;38:46–51. [8] Ridker PM, Paynter NP, Rifai N, Gaziano JM, Cook NR. C-reactive protein and parental history improve global cardiovascular risk prediction: the Reynolds Risk Score for men. Circulation 2008;118:2243–4. [9] Conroy RM, Pyorala K, Fitzgerald AP, Sans S, Menotti A, De Backer G, et al. Estimation of ten-year risk of fatal cardiovascular disease in Europe: the SCORE project. Eur Heart J 2003;24:987–1003. [10] Hippisley-Cox J, Coupland C, Vinogradova Y, Robson J, May M, Brindle P. Derivation and validation of QRISK, a new cardiovascular disease risk score for the United Kingdom: prospective open cohort study. Br Med J 2007;335:136–41. [11] Woodward M, Brindle P, Tunstall-Pedoe H, SIGN group on risk estimation. Adding social deprivation and family history to cardiovascular risk assessment: the ASSIGN score from the Scottish Heart Health Extended Cohort (SHHEC). Heart 2007;93:172–6. [12] Assmann G, Cullen P, Schulte H. Simple scoring scheme for calculating the risk of acute coronary events based on the 10-year follow-up of the Prospective Cardiovascular Munster (PROCAM) study. Circulation 2002;105: 310–5. [13] National Vascular Disease Prevention Alliance. Guidelines for the assessment of absolute cardiovascular disease risk. Canberra: National Heart Foundation of Australia, 2009. [14] New Zealand Guidelines Group. Best practice evidence-based guideline: the assessment and management of cardiovascular risk. Wellington: NZGG, 2003. [15] Wood D, Wray R, Poulter N, Williams B, Kirby M, Patel V, et al. JBS 2: Joint British Societies’ guidelines on prevention of cardiovascular disease in clinical practice. Heart 2005;91:v1–52. [16] Grundy SM, Cleeman JI, Merz CN, Brewer HB, Clark LT, Hunninghake DB, et al. Implications of recent clinical trials for the National Cholesterol Education Program Adult Treatment Panel III guidelines. Circulation 2004;110:227–39. [17] Cooper A, O’Flynn N, Guideline Development Group. Risk assessment and lipid modification for primary and secondary prevention of cardiovascular disease: summary of NICE guidance. Br Med J 2008;336:1246–8. [18] De Backer G, Ambrosioni E, Borch-Johnsen K, Brotons C, Cifkova R, Dallongeville J, et al. European guidelines on cardiovascular disease prevention in clinical practice. Third Joint Task Force of European and Other Societies on Cardiovascular Disease Prevention in Clinical Practice. Eur Heart J 2003;24: 1601–10. [19] Perk J, De Backer G, Gohlke H, Graham I, Reiner Z, Verschuren WM, et al. European guidelines on cardiovascular disease prevention in clinical practice (version 2012). Eur Heart J 2012;33:1635–701. [20] Scottish Intercollegiate Guidelines Network. Risk estimation and the prevention of cardiovascular disease. A national clinical guideline. Edinburgh: SIGN, 2007. [21] Jackson R. Cardiovascular risk prediction: are we there yet? Heart 2008;94: 1–3. [22] Eichler K, Zoller M, Tschudi P, Steurer J. Barriers to apply cardiovascular prediction rules in primary care: a postal survey. BMC Fam Pract 2007;8:1. [23] van Steenkiste B, van der Weijden T, Stoffers H, Grol R. Barriers to implementing cardiovascular risk tables in routine general practice. Scand J Prim Health Care 2004;22:32–7. [24] Marshall T. Estimating the value of information in strategies for identifying patients at high risk of cardiovascular disease. Inform Prim Care 2006;14: 85–92. [25] Gail MH, Brinton LA, Byar DP, Corle DK, Green SB, Schairer C, et al. Projecting individualized probabilities of developing breast cancer for white females who are being examined annually. J Natl Cancer Inst 1989;81:1879–86. [26] Costantino JP, Gail MH, Pee D, Anderson S, Redmond CK, Benichou J, et al. Validation studies for models projecting the risk of invasive and total breast cancer incidence. J Natl Cancer Inst 1999;91:1541–8. [27] Anothaisintawee T, Teerawattananon Y, Wiratkapun C, Kasamesup V, Thakkinstian A. Risk prediction models of breast cancer: a systematic review of model performances. Breast Cancer Res Treat 2012;133:1–10. [28] Meads C, Ahmed I, Riley RD. A systematic review of breast cancer incidence risk prediction models with meta-analysis of their performance. Breast Cancer Res Treat 2012;132:365–77.
3
[29] Wacholder S, Hartge P, Prentice R, Garcia-Closas M, Feigelson HS, Diver WR, et al. Performance of common genetic variants in breast-cancer risk models. N Engl J Med 2010;362:986–93. [30] National Cancer Institute. The nation’s investment in cancer research. A plan and budget proposal for the fiscal year 2006. Available from: http:// www.cancer.gov/aboutnci/budget_planning_leg/plan-archives [accessed 18.02.13]. [31] Win AK, MacInnis RJ, Hopper JL, Jenkins MA. Risk prediction models for colorectal cancer: a review. Cancer Epidemiol Biomarkers Prev 2012;21: 398–410. [32] Williams LH, Barlow WE, Shors AR, Solomon C, White E. Identifying persons at highest risk of melanoma using self-assessed risk factors. J Clin Exp Dermatol Res; 2011;2. [33] Fortes C, Mastroeni S, Bakos L, Antonelli G, Alessandroni L, Pilla MA, et al. Identifying individuals at high risk of melanoma: a simple tool. Eur J Cancer Prev 2010;19:393–400. [34] Fears TR, Guerry D, Pfeiffer RM, Sagebiel RW, Elder DE, Halpern A, et al. Identifying individuals at high risk of melanoma: a practical predictor of absolute risk. J Clin Oncol 2006;24:3590–6. [35] Mar V, Wolfe R, Kelly JW. Predicting melanoma risk for the Australian population. Australas J Dermatol 2011;52:109–16. [36] Cho E, Rosner BA, Feskanich D, Colditz GA. Risk factors and individual probabilities of melanoma for whites. J Clin Oncol 2005;23:2669–75. [37] Hoggart C, Brennan P, Tjonneland A, Vogel U, Overvad K, Ostergaard JN, et al. A risk model for lung cancer incidence. Cancer Prev Res (Phila Pa) 2012;5: 834–46. [38] Tammemagi CM, Pinsky PF, Caporaso NE, Kvale PA, Hocking WG, Church TR, et al. Lung cancer risk prediction: Prostate, Lung Colorectal and Ovarian Cancer Screening Trial models and validation. J Natl Cancer Inst 2011; 103:1058–68. [39] Spitz MR, Hong WK, Amos CI, Wu X, Schabath MB, Dong Q, et al. A risk model for prediction of lung cancer. J Natl Cancer Inst 2007;99:715–26. [40] Cassidy A, Myles JP, van Tongeren M, Page RD, Liloglou T, Duffy SW, et al. The LLP risk model: an individual risk prediction model for lung cancer. Br J Cancer 2008;98:270–6. [41] Bach PB, Kattan MW, Thornquist MD, Kris MG, Tate RC, Barnett MJ, et al. Variations in lung cancer risk among smokers. J Natl Cancer Inst 2003;95: 470–8. [42] Wen CP, Lin J, Yang YC, Tsai MK, Tsao CK, Etzel C, et al. Hepatocellular carcinoma risk prediction model for the general population: the predictive power of transaminases. J Natl Cancer Inst 2012;104:1599–611. [43] Michikawa T, Inoue M, Sawada N, Iwasaki M, Tanaka Y, Shimazu T, et al. Development of a prediction model for 10-year risk of hepatocellular carcinoma in middle-aged Japanese: the Japan Public Health Center-based Prospective Study Cohort II. Prev Med 2012;55:137–43. [44] Thompson IM, Ankerst DP, Chi C, Goodman PJ, Tangen CM, Lucia MS, et al. Assessing prostate cancer risk: results from the prostate cancer prevention trial. J Natl Cancer Inst 2006;98:529–34. [45] Nam RK, Toi A, Klotz LH, Trachtenberg J, Jewett MA, Appu S, et al. Assessing individual risk for prostate cancer. J Clin Oncol 2007;25:3582–8. [46] Smith RA, Cokkinides V, Brooks D, Saslow D, Shah M, Brawley OW. Cancer screening in the United States, 2011: a review of current American Cancer Society guidelines and issues in cancer screening. CA Cancer J Clin 2011;61: 8–30. [47] Fisher B, Costantino JP, Wickerham DL, Cecchini RS, Cronin WM, Robidoux A, et al. Tamoxifen for the prevention of breast cancer: current status of the National Surgical Adjuvant Breast and Bowel Project P-1 study. J Natl Cancer Inst 2005;97:1652–62. [48] Amir E, Freedman OC, Seruga B, Evans DG. Assessing women at high risk of breast cancer: a review of risk assessment models. J Natl Cancer Inst 2010;102:680–91. [49] Thrift AP, Kendall BJ, Pandeya N, Whiteman DC. A model to determine absolute risk for esophageal adenocarcinoma. Clin Gastroenterol Hepatol 2013; 11:138–44. [50] Pohl H, Welch HG. The role of overdiagnosis and reclassification in the marked increase of esophageal adenocarcinoma incidence. J Natl Cancer Inst 2005; 97:142–6. [51] Thrift AP, Whiteman DC. The incidence of esophageal adenocarcinoma continues to rise: analysis of period and birth cohort effects on recent trends. Ann Oncol 2012;23:3155–62. [52] Gavin AT, Francisci S, Foschi R, Donnelly DW, Lemmens V, Brenner H, et al. Oesophageal cancer survival in Europe: a EUROCARE-4 study. Cancer Epidemiol 2012;36:505–12. [53] Reid BJ, Li X, Galipeau PC, Vaughan TL. Barrett’s oesophagus and oesophageal adenocarcinoma: time for a new synthesis. Nat Rev Cancer 2010;10:87–101. [54] Lagergren J, Bergstrom R, Lindgren A, Nyren O. Symptomatic gastroesophageal reflux as a risk factor for esophageal adenocarcinoma. N Engl J Med 1999;340:825–31. [55] Dent J, El-Serag HB, Wallander MA, Johansson S. Epidemiology of gastrooesophageal reflux disease: a systematic review. Gut 2005;54:710–7. [56] Flegal KM, Carroll MD, Kit BK, Ogden CL. Prevalence of obesity and trends in the distribution of body mass index among US adults, 1999–2010. JAMA 2012; 307:491–7. [57] The Information Centre. Statistics on obesity, physical activity and diet: England, 2013. London: The Information Centre, 2013.
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002
G Model
CANEP-565; No. of Pages 4 4
A.P. Thrift, D.C. Whiteman / Cancer Epidemiology xxx (2013) xxx–xxx
[58] Centers for Disease Control and Prevention. Current Cigarette Smoking Among Adults – United States, 2011. MMWR Morb Mortal Wkly Rep 2011;61:889–94. [59] Office for National Statistics. General lifestyle survey overview: a report on the 2010 general lifestyle survey; 2012. [60] Whiteman DC, Sadeghi S, Pandeya N, Smithers BM, Gotley DC, Bain CJ, et al. Combined effects of obesity, acid reflux and smoking on the risk of adenocarcinomas of the oesophagus. Gut 2008;57:173–80. [61] Ware JH. The limitations of risk factors as prognostic tools. N Engl J Med 2006;355:2615–7.
[62] Wald NJ, Hackshaw AK, Frost CD. When can a risk factor be used as a worthwhile screening test? Br Med J 1999;319:1562–5. [63] Aschard H, Chen J, Cornelis MC, Chibnik LB, Karlson EW, Kraft P. Inclusion of gene–gene and gene–environment interactions unlikely to dramatically improve risk prediction for complex diseases. Am J Hum Genet 2012;90:962–72. [64] Janssens AC, van Duijn CM. Genome-based prediction of common diseases: advances and prospects. Hum Mol Genet 2008;17:R166–73. [65] Rose G. Strategy of prevention: lessons from cardiovascular disease. Br Med J 1981;282:1847–51.
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002
CANEP-565; No. of Pages 4 Cancer Epidemiology xxx (2013) xxx–xxx
Contents lists available at SciVerse ScienceDirect
Cancer Epidemiology The International Journal of Cancer Epidemiology, Detection, and Prevention journal homepage: www.cancerepidemiology.net
Review
Can we really predict risk of cancer?§ Aaron P. Thrift, David C. Whiteman * Population Health Department, Queensland Institute of Medical Research, Queensland, Australia
A R T I C L E I N F O
A B S T R A C T
Article history: Received 27 February 2013 Accepted 4 April 2013 Available online xxx
Background: Growing awareness of the potential to predict a person’s future risk of cancer has resulted in the development of numerous algorithms. Such algorithms aim to improve the ability of policy makers, doctors and patients to make rational decisions about behaviour modification or surveillance, with the expectation that this activity will lead to overall benefit. There remains debate however, about whether accurate risk prediction is achievable for most cancers. Methods: We conducted a brief narrative review of the literature regarding the history and challenges of risk prediction, highlighting our own recent experiences in developing tools for oesophageal adenocarcinoma. Results and conclusions: While tools for predicting future risk of cardiovascular outcomes have been translated successfully to clinical practice, the experience with cancer risk prediction has been mixed. Models have now been developed and validated for predicting risk of melanoma and cancers of the breast, colo-rectum, lung, liver, oesophagus and prostate, and while several of these have adequate performance at the population-level, none to date have adequate discrimination for predicting risk in individual patients. Challenges of individual risk prediction for cancer are many, and include long latency, multiple risk factors of mostly small effect, and incomplete knowledge of the causal pathways. ß 2013 Published by Elsevier Ltd.
Keywords: Neoplasms Risk assessment Risk factors Esophageal neoplasms Risk reduction behavior
Clinical decision-making is generally based on estimating average risks within population subgroups and applying the same strategy for all within these groups. However, doctors want to know which diseases are most likely to affect their individual patients. Equally important, a patient wants to know if they have higher than average risk of certain diseases. If doctors can tailor prevention and treatment strategies for a patient based on their absolute risk for a disease, derived from the individual’s specific characteristics (e.g., a panel of factors including age, sex, body mass index, and diet), there is potential for better clinical outcomes [1]. This is the aim of the term ‘personalized medicine’. To fully realize personalized medicine, the scientific community has synthesized information about established risk factors (clinical, environmental and genetic) for specific diseases into statistical models for risk prediction [2–4]. These models use information on multiple risk factors to estimate absolute risk. Risk models are generally developed by selecting the most important panel of predictive factors from among a larger list of candidate
§ Grant support: APT is supported by the Cancer Council NSW STREP grant 08-04. DCW is supported by a Future Fellowship from the Australian Research Council (FT0990987). * Corresponding author at: Cancer Control Group, Queensland Institute of Medical Research, Locked Bag 2000, Royal Brisbane Hospital, Queensland 4029, Australia. Tel.: +61 7 33620279; fax: +61 7 38453502. E-mail addresses: [email protected], [email protected] (D.C. Whiteman).
risk factors. Each predictor is then assigned a relative weight in a combined risk score [2,4]. Individual risk profiles can then be generated and used to identify those at high risk of having (diagnosis) or developing (prognosis) a particular disease. The models have various applications, including use in development of health policy, education, clinical decision-making, and designing future research (e.g., establishing eligibility criteria for intervention and screening trials) [5]. They are not however intended to replace doctors’ clinical decision-making, but rather allow for more objective estimates of risk and uniform decision-making across different centers [6]. Risk prediction first received attention in the cardiovascular literature. One of the most widely validated and used prediction tools is the Framingham Risk Score (FRS) [7]. The FRS was designed to predict 10-year absolute risk for coronary heart disease. Absolute risk assessment charts based on the FRS and subsequent risk scores [8–12] are now included with many cardiovascular disease prevention guidelines [13–20]. Absolute risk estimates are used to guide the management of high-risk patients for cardiovascular disease and have improved the efficacy of medical interventions. The overall aim is to see their use translated into lower incidence and mortality from cardiovascular disease. However, decades after first being recommended, use of the absolute risk assessment charts in clinical practice tends to be highly variable. There remain many barriers to overcome (e.g., time constraints, over simplification of risk assessment, and overmedication) before they are routinely used [21–24].
1877-7821/$ – see front matter ß 2013 Published by Elsevier Ltd. http://dx.doi.org/10.1016/j.canep.2013.04.002
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002
G Model
CANEP-565; No. of Pages 4 2
A.P. Thrift, D.C. Whiteman / Cancer Epidemiology xxx (2013) xxx–xxx
The first risk prediction model for cancer was described by Gail et al. in 1989 [25]. The Gail breast cancer model was developed using data from a nested case-control study (the Breast Cancer Detection and Demonstration Project) and predicted risk of developing invasive or in situ breast cancer in a defined age interval. The model was modified in 1999 to predict absolute 5year risks for invasive breast cancer only [26]. The Gail breast cancer models included terms for age, hormonal or reproductive history, previous history of breast disease, and family history. Investigators have since further modified the original Gail model by adding modifiable risk factors (e.g., body mass index), terms for risk biomarkers (e.g., breast density) and multiple genetic variants, and by developing racial and ethnic specific models [27–29]. The National Cancer Institute (NCI) recognizes risk prediction as an ‘‘area of extraordinary opportunity’’ [30] and cancer risk prediction models are increasingly common in the medical literature. Population-based models have now been developed and validated for predicting risk of colorectal cancer [31], melanoma [32–36], lung cancer [37–41], liver cancer [42,43], and prostate cancer [44,45]. Many have associated online risk calculators available on the NCI’s website (http://epi.grants.cancer.gov/cancer_risk_prediction/). More and more of these traditional models (i.e., those with only clinical and epidemiologic risk factors) are being modified to incorporate biologic and genetic data to estimate cancer risk more accurately. It is difficult however to measure the value of these models. While most claim statistically significant results, few models have translated into clinically useful tools. Before being used clinically, a prediction model must be shown to provide accurate and generalisable risk predictions. The accuracy and generalisability of a model is assessed using measures of discrimination (how it performs at the individual-level) and calibration (populationlevel) [2,4]. These should be assessed in a similar, but external population to that in which the model was developed [3]. However, it is not always possible to find a suitable external validation dataset, and thus, models continue to be validated within the development dataset using various techniques (e.g., bootstrapping, and cross-validation) [2,3]. Several breast cancer models are currently used in clinical practice. In the United States, absolute risk estimates derived from the Gail model are used to identify high-risk women for mammography screening [46] and interventions trials [47]. However, while well calibrated at the population-level, the breast cancer models lack adequate discrimination in predicting risk for individuals [27,28,48]. We recently developed a model to estimate absolute risk for esophageal adenocarcinoma (EAC) [49]. EAC is a relatively rare cancer but is the fastest rising cancer in Western populations [50,51] and has poor survival [52]. EAC has a number of clearly established risk factors with reasonably large effect sizes (male sex, white ethnicity, reflux, obesity, and tobacco smoking) and thus ought to lend nicely to risk prediction [53]. However, despite the identification of ‘high-risk phenotypes’, it remains the case that very few people with one or more of these factors develop EAC. For instance, the absolute 5-year risk for EAC among obese, white, 50year-old males with reflux remains only 0.04% [49], and more than 40% of EAC patients do not report reflux symptoms at diagnosis [54]. It comes as no surprise then that our model had only moderate discriminatory accuracy when internally validated. That is, it failed to effectively differentiate between those who will develop EAC and those who will not. Why are we seeing this time and again for these cancer risk prediction models? Like other cancers, there are a number of issues as to why we may not be able to estimate EAC risk accurately at the individual level. Firstly, the primary risk factors for EAC are common in the population (for example, up to 20% of the adult population in
Western countries suffer from weekly reflux symptoms [55], and approximately 30% are obese [56,57] and 20% are current smokers [58,59]) and are therefore prevalent in individuals who will not develop the disease. Secondly, while the relative risks for these factors are consistent across populations, and statistically significantly associated with EAC risk, the effects are typically modest (relative risks < 5) [60]. As a general rule of thumb, risk factors require relative risks greater than 10 to be good predictors of individual risk, although this does not guarantee high discriminatory accuracy [61,62]. Whilst knowledge of these factors helps advance our understanding of disease mechanisms, these risk factors alone do not accurately discriminate those who will develop EAC from those who will not. Thirdly, most cancers have long latent periods and arise in individuals with risk close to the population average. Therefore, identifying high-risk groups for EAC on the basis of a probability threshold (e.g., using 5-year absolute risks) with high sensitivity is difficult. For rare cancers such as EAC, it may not be cost-effective to use risk prediction models in clinical practice. For example, even predictors with high sensitivity and specificity can still have low positive predictive values for rare diseases. That is, the number of people who actually develop cancer as a proportion of those predicted to do so will be low. As such, even if we had predictors with high discriminatory power (i.e., the test can correctly classify most people as ‘affected’ or ‘not affected’), most people exposed to the risks of further investigation or treatment will not see any benefit. Therefore, such a model appears well suited for population prevention strategies, but may not be able to identify individuals at high (or low) risk in a clinically meaningful and cost-effective way. With extensive research efforts already spent trying to improve current cancer models, is it possible to improve their clinical utility and move towards personalized medicine? Genetic risk profiling, in particular the addition of genetic variants to existing cancer risk models (as well as gene gene and gene environment interaction terms), has been proposed as a solution to increase discriminatory ability. However, it has been shown that, even when alleles with modest effect sizes are combined, the addition of genetic variants to these models has added only modestly to discrimination [29,63]. Thus, integrating genetic risk profiles to traditional models does not appreciably improve performance, at least with current knowledge [64]. Perhaps there are other risk factors (with high relative risks) not yet identified that, if added to existing cancer model, might add substantially to predictive ability over and above currently used risk factors. This seems unlikely for most cancers. One of the major lessons that we have learnt from risk prediction is that our knowledge of disease pathways is imperfect. Assuming we had a validated cancer risk prediction model and before widespread implementation, we need robust external validation, and of course, we need evidence that these tools actually improve patient outcomes when used clinically. Such evidence is best sourced from randomized trials, which are not without challenges. The goal of personalized medicine, at least in terms of predicting risks for the purposes of targeted prevention, may therefore be illusory. Recent experience with cancer risk prediction adds a layer of complexity to the ‘prevention paradox’ described by Geoffrey Rose more than 30 years ago, in which he made the observation that the bulk of disease arises in people with low risk [65]. The most effective strategies might be to target prevention at the entire population, and then put efforts into better diagnosis and better treatment for those who do develop illness. Conflict of interest None declared.
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002
G Model
CANEP-565; No. of Pages 4 A.P. Thrift, D.C. Whiteman / Cancer Epidemiology xxx (2013) xxx–xxx
References [1] Cooney MT, Dudina A, D’Agostino R, Graham IM. Cardiovascular risk-estimation systems in primary prevention: do they differ? Do they make a difference? Can we see the future? Circulation 2010;122:300–10. [2] Moons KG, Kengne AP, Woodward M, Royston P, Vergouwe Y, Altman DG, et al. Risk prediction models. I. Development, internal validation, and assessing the incremental value of a new (bio)marker. Heart 2012;98:683–90. [3] Moons KG, Kengne AP, Grobbee DE, Royston P, Vergouwe Y, Altman DG, et al. Risk prediction models. II. External validation, model updating, and impact assessment. Heart 2012;98:691–8. [4] Steyerberg EW. Clinical prediction models: a practical approach to development, validation and updating. New York: Springer, 2009. [5] Freedman AN, Seminara D, Gail MH, Hartge P, Colditz GA, Ballard-Barbash R, et al. Cancer risk prediction models: a workshop on development, evaluation, and application. J Natl Cancer Inst 2005;97:715–23. [6] Moons KG, Royston P, Vergouwe Y, Grobbee DE, Altman DG. Prognosis and prognostic research: what, why, and how? Br Med J; 2009;338. [7] Kannel WB, McGee D, Gordon T. A general cardiovascular risk profile: the Framingham Study. Am J Cardiol 1976;38:46–51. [8] Ridker PM, Paynter NP, Rifai N, Gaziano JM, Cook NR. C-reactive protein and parental history improve global cardiovascular risk prediction: the Reynolds Risk Score for men. Circulation 2008;118:2243–4. [9] Conroy RM, Pyorala K, Fitzgerald AP, Sans S, Menotti A, De Backer G, et al. Estimation of ten-year risk of fatal cardiovascular disease in Europe: the SCORE project. Eur Heart J 2003;24:987–1003. [10] Hippisley-Cox J, Coupland C, Vinogradova Y, Robson J, May M, Brindle P. Derivation and validation of QRISK, a new cardiovascular disease risk score for the United Kingdom: prospective open cohort study. Br Med J 2007;335:136–41. [11] Woodward M, Brindle P, Tunstall-Pedoe H, SIGN group on risk estimation. Adding social deprivation and family history to cardiovascular risk assessment: the ASSIGN score from the Scottish Heart Health Extended Cohort (SHHEC). Heart 2007;93:172–6. [12] Assmann G, Cullen P, Schulte H. Simple scoring scheme for calculating the risk of acute coronary events based on the 10-year follow-up of the Prospective Cardiovascular Munster (PROCAM) study. Circulation 2002;105: 310–5. [13] National Vascular Disease Prevention Alliance. Guidelines for the assessment of absolute cardiovascular disease risk. Canberra: National Heart Foundation of Australia, 2009. [14] New Zealand Guidelines Group. Best practice evidence-based guideline: the assessment and management of cardiovascular risk. Wellington: NZGG, 2003. [15] Wood D, Wray R, Poulter N, Williams B, Kirby M, Patel V, et al. JBS 2: Joint British Societies’ guidelines on prevention of cardiovascular disease in clinical practice. Heart 2005;91:v1–52. [16] Grundy SM, Cleeman JI, Merz CN, Brewer HB, Clark LT, Hunninghake DB, et al. Implications of recent clinical trials for the National Cholesterol Education Program Adult Treatment Panel III guidelines. Circulation 2004;110:227–39. [17] Cooper A, O’Flynn N, Guideline Development Group. Risk assessment and lipid modification for primary and secondary prevention of cardiovascular disease: summary of NICE guidance. Br Med J 2008;336:1246–8. [18] De Backer G, Ambrosioni E, Borch-Johnsen K, Brotons C, Cifkova R, Dallongeville J, et al. European guidelines on cardiovascular disease prevention in clinical practice. Third Joint Task Force of European and Other Societies on Cardiovascular Disease Prevention in Clinical Practice. Eur Heart J 2003;24: 1601–10. [19] Perk J, De Backer G, Gohlke H, Graham I, Reiner Z, Verschuren WM, et al. European guidelines on cardiovascular disease prevention in clinical practice (version 2012). Eur Heart J 2012;33:1635–701. [20] Scottish Intercollegiate Guidelines Network. Risk estimation and the prevention of cardiovascular disease. A national clinical guideline. Edinburgh: SIGN, 2007. [21] Jackson R. Cardiovascular risk prediction: are we there yet? Heart 2008;94: 1–3. [22] Eichler K, Zoller M, Tschudi P, Steurer J. Barriers to apply cardiovascular prediction rules in primary care: a postal survey. BMC Fam Pract 2007;8:1. [23] van Steenkiste B, van der Weijden T, Stoffers H, Grol R. Barriers to implementing cardiovascular risk tables in routine general practice. Scand J Prim Health Care 2004;22:32–7. [24] Marshall T. Estimating the value of information in strategies for identifying patients at high risk of cardiovascular disease. Inform Prim Care 2006;14: 85–92. [25] Gail MH, Brinton LA, Byar DP, Corle DK, Green SB, Schairer C, et al. Projecting individualized probabilities of developing breast cancer for white females who are being examined annually. J Natl Cancer Inst 1989;81:1879–86. [26] Costantino JP, Gail MH, Pee D, Anderson S, Redmond CK, Benichou J, et al. Validation studies for models projecting the risk of invasive and total breast cancer incidence. J Natl Cancer Inst 1999;91:1541–8. [27] Anothaisintawee T, Teerawattananon Y, Wiratkapun C, Kasamesup V, Thakkinstian A. Risk prediction models of breast cancer: a systematic review of model performances. Breast Cancer Res Treat 2012;133:1–10. [28] Meads C, Ahmed I, Riley RD. A systematic review of breast cancer incidence risk prediction models with meta-analysis of their performance. Breast Cancer Res Treat 2012;132:365–77.
3
[29] Wacholder S, Hartge P, Prentice R, Garcia-Closas M, Feigelson HS, Diver WR, et al. Performance of common genetic variants in breast-cancer risk models. N Engl J Med 2010;362:986–93. [30] National Cancer Institute. The nation’s investment in cancer research. A plan and budget proposal for the fiscal year 2006. Available from: http:// www.cancer.gov/aboutnci/budget_planning_leg/plan-archives [accessed 18.02.13]. [31] Win AK, MacInnis RJ, Hopper JL, Jenkins MA. Risk prediction models for colorectal cancer: a review. Cancer Epidemiol Biomarkers Prev 2012;21: 398–410. [32] Williams LH, Barlow WE, Shors AR, Solomon C, White E. Identifying persons at highest risk of melanoma using self-assessed risk factors. J Clin Exp Dermatol Res; 2011;2. [33] Fortes C, Mastroeni S, Bakos L, Antonelli G, Alessandroni L, Pilla MA, et al. Identifying individuals at high risk of melanoma: a simple tool. Eur J Cancer Prev 2010;19:393–400. [34] Fears TR, Guerry D, Pfeiffer RM, Sagebiel RW, Elder DE, Halpern A, et al. Identifying individuals at high risk of melanoma: a practical predictor of absolute risk. J Clin Oncol 2006;24:3590–6. [35] Mar V, Wolfe R, Kelly JW. Predicting melanoma risk for the Australian population. Australas J Dermatol 2011;52:109–16. [36] Cho E, Rosner BA, Feskanich D, Colditz GA. Risk factors and individual probabilities of melanoma for whites. J Clin Oncol 2005;23:2669–75. [37] Hoggart C, Brennan P, Tjonneland A, Vogel U, Overvad K, Ostergaard JN, et al. A risk model for lung cancer incidence. Cancer Prev Res (Phila Pa) 2012;5: 834–46. [38] Tammemagi CM, Pinsky PF, Caporaso NE, Kvale PA, Hocking WG, Church TR, et al. Lung cancer risk prediction: Prostate, Lung Colorectal and Ovarian Cancer Screening Trial models and validation. J Natl Cancer Inst 2011; 103:1058–68. [39] Spitz MR, Hong WK, Amos CI, Wu X, Schabath MB, Dong Q, et al. A risk model for prediction of lung cancer. J Natl Cancer Inst 2007;99:715–26. [40] Cassidy A, Myles JP, van Tongeren M, Page RD, Liloglou T, Duffy SW, et al. The LLP risk model: an individual risk prediction model for lung cancer. Br J Cancer 2008;98:270–6. [41] Bach PB, Kattan MW, Thornquist MD, Kris MG, Tate RC, Barnett MJ, et al. Variations in lung cancer risk among smokers. J Natl Cancer Inst 2003;95: 470–8. [42] Wen CP, Lin J, Yang YC, Tsai MK, Tsao CK, Etzel C, et al. Hepatocellular carcinoma risk prediction model for the general population: the predictive power of transaminases. J Natl Cancer Inst 2012;104:1599–611. [43] Michikawa T, Inoue M, Sawada N, Iwasaki M, Tanaka Y, Shimazu T, et al. Development of a prediction model for 10-year risk of hepatocellular carcinoma in middle-aged Japanese: the Japan Public Health Center-based Prospective Study Cohort II. Prev Med 2012;55:137–43. [44] Thompson IM, Ankerst DP, Chi C, Goodman PJ, Tangen CM, Lucia MS, et al. Assessing prostate cancer risk: results from the prostate cancer prevention trial. J Natl Cancer Inst 2006;98:529–34. [45] Nam RK, Toi A, Klotz LH, Trachtenberg J, Jewett MA, Appu S, et al. Assessing individual risk for prostate cancer. J Clin Oncol 2007;25:3582–8. [46] Smith RA, Cokkinides V, Brooks D, Saslow D, Shah M, Brawley OW. Cancer screening in the United States, 2011: a review of current American Cancer Society guidelines and issues in cancer screening. CA Cancer J Clin 2011;61: 8–30. [47] Fisher B, Costantino JP, Wickerham DL, Cecchini RS, Cronin WM, Robidoux A, et al. Tamoxifen for the prevention of breast cancer: current status of the National Surgical Adjuvant Breast and Bowel Project P-1 study. J Natl Cancer Inst 2005;97:1652–62. [48] Amir E, Freedman OC, Seruga B, Evans DG. Assessing women at high risk of breast cancer: a review of risk assessment models. J Natl Cancer Inst 2010;102:680–91. [49] Thrift AP, Kendall BJ, Pandeya N, Whiteman DC. A model to determine absolute risk for esophageal adenocarcinoma. Clin Gastroenterol Hepatol 2013; 11:138–44. [50] Pohl H, Welch HG. The role of overdiagnosis and reclassification in the marked increase of esophageal adenocarcinoma incidence. J Natl Cancer Inst 2005; 97:142–6. [51] Thrift AP, Whiteman DC. The incidence of esophageal adenocarcinoma continues to rise: analysis of period and birth cohort effects on recent trends. Ann Oncol 2012;23:3155–62. [52] Gavin AT, Francisci S, Foschi R, Donnelly DW, Lemmens V, Brenner H, et al. Oesophageal cancer survival in Europe: a EUROCARE-4 study. Cancer Epidemiol 2012;36:505–12. [53] Reid BJ, Li X, Galipeau PC, Vaughan TL. Barrett’s oesophagus and oesophageal adenocarcinoma: time for a new synthesis. Nat Rev Cancer 2010;10:87–101. [54] Lagergren J, Bergstrom R, Lindgren A, Nyren O. Symptomatic gastroesophageal reflux as a risk factor for esophageal adenocarcinoma. N Engl J Med 1999;340:825–31. [55] Dent J, El-Serag HB, Wallander MA, Johansson S. Epidemiology of gastrooesophageal reflux disease: a systematic review. Gut 2005;54:710–7. [56] Flegal KM, Carroll MD, Kit BK, Ogden CL. Prevalence of obesity and trends in the distribution of body mass index among US adults, 1999–2010. JAMA 2012; 307:491–7. [57] The Information Centre. Statistics on obesity, physical activity and diet: England, 2013. London: The Information Centre, 2013.
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002
G Model
CANEP-565; No. of Pages 4 4
A.P. Thrift, D.C. Whiteman / Cancer Epidemiology xxx (2013) xxx–xxx
[58] Centers for Disease Control and Prevention. Current Cigarette Smoking Among Adults – United States, 2011. MMWR Morb Mortal Wkly Rep 2011;61:889–94. [59] Office for National Statistics. General lifestyle survey overview: a report on the 2010 general lifestyle survey; 2012. [60] Whiteman DC, Sadeghi S, Pandeya N, Smithers BM, Gotley DC, Bain CJ, et al. Combined effects of obesity, acid reflux and smoking on the risk of adenocarcinomas of the oesophagus. Gut 2008;57:173–80. [61] Ware JH. The limitations of risk factors as prognostic tools. N Engl J Med 2006;355:2615–7.
[62] Wald NJ, Hackshaw AK, Frost CD. When can a risk factor be used as a worthwhile screening test? Br Med J 1999;319:1562–5. [63] Aschard H, Chen J, Cornelis MC, Chibnik LB, Karlson EW, Kraft P. Inclusion of gene–gene and gene–environment interactions unlikely to dramatically improve risk prediction for complex diseases. Am J Hum Genet 2012;90:962–72. [64] Janssens AC, van Duijn CM. Genome-based prediction of common diseases: advances and prospects. Hum Mol Genet 2008;17:R166–73. [65] Rose G. Strategy of prevention: lessons from cardiovascular disease. Br Med J 1981;282:1847–51.
Please cite this article in press as: Thrift AP, Whiteman DC. Can we really predict risk of cancer? Cancer Epidemiology (2013), http:// dx.doi.org/10.1016/j.canep.2013.04.002