Risk Prediction Models for Colorectal Cancer: A Review

Colorectal cancer (CRC) is one of the most frequently diagnosed cancers in the world with more than one million new cases diagnosed (9.8% of worldwide cancer diagnoses) and 600,000 deaths (8.1% of all worldwide cancer deaths) caused by the disease in 2008 (1). There is likely to be a large spectrum of risk across populations (2). While part of this is due to individual difference in exposure to environmental risk factors, in theory, there is also likely a large variation of risk due to underlying familial risk factors, which we refer to as the “familial risk profile” (3–7). Mathematical modeling of the relationship between family history as a risk factor and underlying familial risk profile suggests that (i) the risk of developing CRC varies approximately 20-fold between the people in the lowest quartile (average 1.25% lifetime risk of CRC) versus the highest quartile for familial risk profile (average 25% risk; ref. 4) and (ii) 90% of all CRCs occurs in people who are above the median familial risk profile (6, 8).

Many risk factors have been implicated in the development of CRC (9). Family history is a well-established risk factor (10). Studies have reported that individuals with one affected first-degree relative (parent, offspring, sibling) have, on average, a 2-fold increased risk of CRC compared with those with no family history (10–12) and this association increases to 4-fold for individuals with 3 or more affected first-degree relatives (10). However, the above statistics are mere averages. Heterogeneity in both absolute and relative CRC cancer risk associated with family history depends on the age of the individual at risk, the age(s) at diagnosis of affected relative(s) (the earlier the age at diagnosis the more likely to be at risk), and the genetic relationship between them (the greater the number and/or the closer the relationship to, the more likely to be at risk; ref. 10).

Approximately 3% to 5% of CRC is now known to be due to mutations in the known high-risk cancer susceptibility genes (13). The strongest ones in terms of risk and mutation frequency comprise the DNA mismatch repair (MMR) genes that cause Lynch syndrome (14), the adenomatous polyposis coli gene that causes familial adenomatous polyposis (15), and the MUTYH gene that causes colorectal polyps and subsequently cancer (MUTYH-associated polyposis; ref. 16). Although rare (in aggregate, 1 in 500 to 1 in 10,000 of the population carry a mutation in these genes; refs. 17–19), these mutation carriers are at high risk of developing cancer, particularly CRC (20–28).

At most, only half of the familial risk of CRC can be explained by mutations in known high-risk genes (29). While much research capital has been spent on the search for other CRC “high-risk” susceptibility genes, none have been confirmed. On the other hand, genome-wide association studies have identified at least 15 common genetic susceptibility markers (single-nucleotide polymorphisms, SNP) that have been reliably shown to be associated with small increments in risk of developing CRC (ORs for homozygotes of the minor allele vs. noncarriers ranging from 0.80 to 1.70; refs. 30–38). The causes of these weak associations, let alone of the residual of the familial risk, are unknown, but they could include less common genetic variants and/or other risk factors shared by relatives.

The etiology of developing CRC is also attributable to nongenetic, personal, and environmental factors. Meta-analyses have reported positive associations with obesity [relative risk (RR) per 5 kg/m²: colon cancer, 1.24 (95% confidence interval, CI, 1.20–1.28) for men and 1.09 (95% CI, 1.05–1.13) for women; and rectal cancer, 1.09 (95% CI, 1.06–1.12) for men and 1.02 (95% CI, 1.00–1.05) for women; ref. 39], diabetes [RR, 1.29 (95% CI, 1.15–1.44) for men and 1.33 (95% CI, 1.23–1.44) for women; ref. 40], consumption of red meat [RR for 120 g/d, 1.28 (95% CI, 1.18–1.39); ref. 41] and processed meat [RR for 30 g/d, 1.09 (95% CI, 1.05–1.13); ref. 41], and negative associations with physical activity [RR of colon cancer for occupational activities vs. none, 0.79 (95% CI, 0.72–0.87) for men, and for recreational activities vs. none, 0.78 (95% CI, 0.68–0.91) for men and 0.71 (95% CI, 0.57–0.88) for women; ref. 42], postmenopausal hormone replacement therapy [RR for ever user vs. never user, 0.80 (95% CI, 0.74–0.86) for colon cancer and 0.81 (95% CI, 0.72–0.92) for rectal cancer; ref. 43], calcium intake [RR for the highest vs. the lowest quintile of intake, 0.86 (95% CI, 0.79–0.95) for dietary calcium and 0.78 (95% CI, 0.69–0.88) for total calcium; ref. 44], vitamin D intake [RR for the highest vs. lowest categories, 0.88 (95% CI, 0.80–0.96); ref. 45] and aspirin use [RR for ≥2–3 times weekly for >1 year vs. never user, 0.78 (95% CI, 0.63–0.97); ref. 46].

Risk prediction models are important because they can be used (i) to reduce the burden of disease by enabling targeted preventive measures to those at highest risk; (ii) to identify those most likely to carry a genetic predisposition, who then can be triaged for genetic testing (47); and (iii) to increase the power of observation studies to identify new risk factors for the disease (provided that such risk prediction models do have high predictive power for identifying individuals at high risk for CRC). Risk prediction models are typically developed using data from observational study designs including case–control and cohort studies (48, 49).

Health policy makers and clinicians rely on risk classifications to decide which individuals to screen for CRC. To date, these classification schemes are predominantly based only on age and/or a simple classification of family history. The advantage of using risk prediction models that incorporate multiple variables (including novel predictors such as genomic data) is that they generally have greater accuracy than clinical stage alone (50). CRC risk prediction models identify individuals most likely to benefit from CRC screening or other preventative interventions. The majority of CRCs derive from a polyp, a well-recognized premalignant form that may be present for many years before symptoms (such as rectal bleeding, change in bowel habit, or anemia) manifest. These polyps can be removed during colonoscopy, thereby reducing the risk of CRC (51, 52). Furthermore, as CRCs are generally asymptomatic before the cancer has reached a relatively advanced stage, screening by fecal occult blood test and colonoscopy is an effective method to reduce metastasis and death by CRC by detecting the disease at an early stage (51, 53). Accurate CRC risk predictions could help with clinical decisions on the frequency and mode of CRC screening and preventive strategies for those who are at high risk while leaving those who are at low risk of disease unexposed to potentially invasive screening modalities.

A small proportion of the population have inherited a genetic mutation that puts them at a high risk of CRC (18, 19). The relevant genes identified to date include the MMR genes (MLH1, MSH2, MSH6, and PMS2), APC, MUTYH, STK11 (LB1), BMPR1A, and PTEN. Carriers of a mutation in any one of these genes have a lifetime risk of CRC of 30% to 100% (compared with a 5% lifetime risk for the general population) and benefit greatly from CRC screening (54, 55). Likely carriers of a mutation in an MMR gene can be identified from among CRC cases by pathologic examination of their tumors (56), whereas likely carriers of a mutation in the APC gene and a proportion of carriers of 2 mutations in the MUTYH gene can be identified by colonoscopic examination of the bowel for multiple polyps (57). However, for CRC-unaffected individuals and those who are not screened by colonoscopy, the identification of likely carriers requires a predictive model. Once identified, those most likely to be carriers could be triaged to genetic testing (which would be prohibitively expensive if provided to the whole population given less than 1 in 500 of the population carry a mutation; ref. 19).

Traditionally, observational studies of CRC risk have either compared the frequency of environmental, lifestyle, and genetic risk factors of affected with unaffected (case–control design) or have compared CRC incidence for those exposed with that for those not exposed (cohort design). In the vast majority of these studies, participants have been sampled irrespective of their family history of CRC, and as a consequence, most of the population does not have a family history of the disease. Therefore, finding from such research might not be generalizable to people with a family history of the disease. Furthermore, as analyses are usually adjusted for a family history, the observed associations are merely the average association across the broad spectrum of familial risk profile (8). Consequently, little is known about the specific role of risk factors within any category of familial risk profile or whether the associations differ by familial risk profile.

Risk prediction models can be used as a tool to try to identify new risk factors for disease. One potential method would be to classify individuals into those who had high risk of disease and those with a low risk of disease and then to compare the cases and controls within these low- and high-risk categories for a given exposure (provided it was not used to classify the risk). This method could have greater statistical power to identify additional risk factors for the disease and may provide evidence for gene–environment interactions (8).

Several CRC risk prediction models have been developed on the basis of known genetic and environmental risk factors. The aim of this review is to summarize these models and the studies that have evaluated them in terms of their applicability, strengths, and weaknesses. Table 1 summarizes risk prediction models for CRC that incorporate personal, environmental, and lifestyle risk factors and/or cancer susceptibility genes. To focus our review, we excluded models that predict stage of CRC based on patient and tumor characteristics (e.g., Cai and colleagues; ref. 58), models that predict type of colorectal neoplasia (benign or malignant) based on clinical symptoms (e.g., Brazer and colleagues; ref. 59), models that predict CRC incidence based on polyp growth rates after first polypectomy (e.g., Wilson and Lightwood; ref. 60), models that predict symptomatic CRC based on bowel symptoms of a patient assessed by a general practitioner (e.g., Selvachandran and colleagues; ref. 61), and models that only predict risk of metachronous CRC (i.e., subsequent primary cancer) based on characteristics of first cancer and gene signature (e.g., Peng and colleagues; ref. 62).

Most of the previously developed models that predict CRC risk use scoring systems based on regression models incorporating family history, lifestyle, and environmental risk factors. Such models are straightforward to implement and nongenetic risk factors can readily be incorporated. However, it is difficult to incorporate detailed information on family history including number of relatives, ages of unaffected relatives, ages of diagnoses, beyond first-generation relatives, and risk factors correlated between family members (both genetic and nongenetic). Consequently, these “nongenetic” models may not give precise or accurate estimates of future CRC risk for important groups. For example, while the models that take account family history of colon cancer (63) or CRC (64, 65) as a binary variable (yes/no) provide an average risk for groups of individuals at a low risk, they are unable to provide accurate risk estimates for individuals at high-risk due to having a strong family history or carrying a known genetic mutation.

Another important issue with respect to history of colonoscopy is the occurrence of screening or diagnostic colonoscopies. A negative colonoscopy indicates a reduced risk of disease (compared with someone who has had no colonoscopy; ref. 66) and a positive colonoscopy indicates an increased risk of disease (evidence of susceptibility to further adenomas; ref. 67) as well as a potential decreased risk of disease if it was followed by polypectomy to remove pre-cancerous adenoma (68). Currently, it is not clear how to incorporate such information into a risk prediction model, and this will be important as colonoscopies are being increasingly used as a diagnostic or screening test.

We are aware of only one model that bases predictions on inherited genetic risk factors for the disease. MMRpro (previously known as CRCAPRO; ref. 69) was developed by Chen and colleagues (70) and it assumes that all familial aggregation is due to dominantly inherited, highly penetrant mutations in MLH1, MSH2, or MSH6. It predicts the probability of being a mutation carrier, by gene, as well as age-specific CRC (and endometrial cancer) risks for unaffected individuals. This model assumes a mutation carrier frequency of 0.0009 (∼1 in 1,100) for MLH1, 0.0010 (∼1 in 1,000) for MSH2, and 0.00036 (∼1 in 2,800) for MSH6 with penetrance functions based on published estimates (70). There are several limitations of this model: (i) it uses family history of CRC only up to second-degree relatives (Table 1); (ii) it does not incorporate the MMR gene PMS2, which accounts for 15% of MMR gene mutations (71), although this probably has little impact given risk estimates are robust to mutation frequencies in this range; (iii) it does not include any environmental risk factors; and (iv) it does not predict second primary cancer risk for affected individuals.

None of the models described above (summarized in Table 1) adequately account for any complexities of the genetic susceptibility of CRC. For example, high-risk mutations in the known CRC susceptibility genes only account for half of the familial risk at most (29). Genome-wide association studies have identified SNPs located in 15 loci that individually are associated with small increments in risk of CRC (37, 72). Modeling suggests that should the residual familial risk be explained by similarly weak genetic effects, then hundreds of genes may exist that alter CRC risk (73). While cancer prediction models have been developed to account for a polygenic component representing the effects of a large number of genetic variants each of small effect on risk for breast (74, 75) and prostate cancers (76), no comparable tool has been developed for CRC.

Before a risk prediction model can be recommended as a useful tool for individualized decision making in a clinical setting, it needs to be validated in an independent population than that was used to create the model (77). The most important characteristics used to evaluate the performance of a risk prediction model are as follows:

• Calibration (or reliability) assesses the ability of a model to predict the number of events (CRCs in this setting) it purports to predict. This is commonly described by using the goodness-of-fit or χ² statistic to compare the expected number of events with the observed number of events (78).

• Discrimination (or precision) measures the ability of a model to distinguish individuals more likely to develop disease from those less likely to develop disease, by using the concordance statistic (c-statistic) that corresponds to the area under a receiver operating characteristic curve (AUC; ref. 78) or the net reclassification index (NRI), which is the probability that the enhanced model reclassifies a person correctly (i.e., increases the estimate of risk of a person who develops CRC or decreases the estimate of risk of a person who does not) minus the probability of an incorrect reclassification (i.e., decreases the estimate of risk of a person who develops CRC or increases the estimate of risk of one who does not; ref. 79).

• Accuracy evaluates how well a model categorizes specific individuals into those more and less likely to develop disease and the usefulness on an individual basis in predicting disease risk. Sensitivity and specificity as well as positive and negative predictive values are important indices for this test.

• Utility, which is the ability of the model to be completed by those it is designed for, such as clinicians, patients, general population, and health policy makers. This is commonly evaluated by results from surveys or interviews of users.

Table 2 summarizes the studies that have evaluated the models for CRC risk prediction. To our knowledge, the models of Driver and colleagues (80) and Wei and colleagues (65) have not been validated. For MMRpro (70), some studies (e.g., refs. 81, 82) have evaluated the probability of a mismatch repair gene mutation status, but not CRC risk. The other 4 models (63, 64, 83, 84) were validated using an independent data set from a separate study or a subset of the same data set used to generate a model. There have been no publications comparing these models with each other.

Three studies (85–87) have assessed this model including 2 qualitative studies (86, 87). While this model was well calibrated for women, it overestimated the number of CRC for men at low risk (85). Utility was evaluated by assessing the lay understanding of risk, perception of risk and interpretation of the results, and was determined to be well received by users (86). A computer-based tool that provides an estimate of personal absolute and relative risk for CRC (the Harvard Colorectal Cancer Risk Assessment and Communication Tool for Research) was evaluated qualitatively by Emmons and colleagues (87) for accuracy of risk perception and level of worry and satisfaction by users. They concluded that the tool was useful for correcting misperceptions about personal risk. Of those with inaccurate risk perception at baseline, more than half of the participants in intervention groups had corrected risk perceptions at posttest, compared with only 12% in the control group (Table 2).

Imperiale and colleagues (84) validated their own model using an independent data set, from the same source used to develop the model; 1,031 men and women without bowel symptoms who underwent a screening colonoscopy. They observed that the respective actual risk of advanced lesion in the proximal colon (tubular adenomas larger than 1 cm, any polyp with villous histology or severe dysplasia, or cancer) was similar to the predicted low-, intermediate- and high-risk estimated by the model, albeit with modest discrimination indicated by the c-statistic of 0.74 (84). However, the validity of the model for different populations is unknown, given that both the model and validation data sets were from the same population.

Park and colleagues (88) evaluated Freeman's model (64) using prospective data from a cohort of ∼260,000 individuals over a mean follow-up of 7 years. They observed that Freeman's model was well calibrated overall for both men and women and in most categories of risk factors. However, the model overestimated risk for individuals with a family history of CRC, with the expected number of CRCs being 35% (95% CI, 17%–55%) higher than the observed number for men who had one affected relative with CRC, and 42% (24%–63%) higher for men who had a history of screening and polyps. The model underestimated the risk for men who had a history of screening but no polyps by 33% (95% CI, 28%–38%; Table 2). The model had a modest discrimination indicated by the c-statistic of 0.61 for men and women.

Ma and colleagues (83) validated their own model using the Japan Public Health Center–based study. They found that their model underestimated the number of colon cancers by 19% (95% CI, 3%–37%), but there was good agreement for rectal cancer and CRC overall. This model underestimated CRC cases with 4 categories of environmental exposures (age, body mass index, alcohol consumption, and smoking).

The models that do exist for CRC risk are limited in terms of incorporated risk factors and their usefulness in terms of validity. More complex modeling is required, based on known risk factors for disease AND hypothesized but residual risk factors. In this section, we discuss the issues to consider for developing a comprehensive model to predict CRC risk.

Risk of CRC varies greatly depending on the presence or absence of a germ line mutation in a cancer risk susceptibility gene. For example, according to the studies that appropriately conditioned on ascertainment of mutation carrying families, the carriers of mutations in MMR genes are estimated to have at least 10 times higher cumulative risk (penetrance) to age 70 years than for the general population: MLH1 and MSLH2, 56% (95% CI, 37%–75%) for men and 48% (95% CI, 26%–65%) for women (21); MSH6, 22% (95% CI, 14%–32%) for men and 10% (95% CI, 5%–17%) for women (20); and PMS2, 20% (95% CI, 11%–34%) for men and 15% (95% CI, 8%–26%) for women (22). Therefore, it is important that future models, particularly those to be applied to populations with a proportion of people who carry these mutations, should have the ability to derive genotype-specific cancer risks. Only the MMRpro model (70) incorporates such risks for MLH1, MSH2, and MSH6 mutations.

The role of physical characteristics and environmental exposures on CRC risk could depend on the presence or absence of carrying an inherited genetic mutation (89). Previous models have not accounted for gene–environment interactions or gene–gene interactions. Incorporation of these interactions is challenging. Studies have observed that the strengths of associations between environmental risk factors and CRC for those with a family history differs compared with those randomly selected from the population (90–92), which is consistent with the existence of gene–environment interactions. However to date, there have been few studies directly comparing strengths of associations for carriers with that for noncarriers. We have shown that body mass index in early adulthood is similarly associated with CRC risk for both carriers and noncarriers (93). Studies of alcohol consumption (94), fruit consumption (95), dietary fiber intake (95), smoking (95, 96) have not directly compared strengths of associations for carriers with noncarriers.

Most existing CRC risk prediction models were developed using data from predominantly Caucasian populations and therefore they may not be applicable to other racial/ethnic populations. Further studies are needed for such populations. There is a well-recognized high degree of heterogeneity by nationality in CRC incidence with an up to 10-fold difference internationally (1, 2). Although some genetic differences may exist for CRC risk, studies have suggested that much of this difference may be due to differences in environmental risk factors as the incidence rate of CRC in migrants approach that of the host country within one or two generations (97).

Risk prediction models need to incorporate different subtypes of CRC (proximal colon, distal colon, and rectal cancer) as distinct disease endpoints as it is possible that the genetic and environmental risk factors may differ by anatomic location. For instance, beer consumption is associated with rectal cancer in men but not in women, whereas there is little evidence of an association of any dimension of alcohol consumption being associated with the risk of colon cancer (98).

The incidence of metachronous CRC, that is, a subsequent primary cancer for people with a previous diagnosis of CRC, is estimated to be 8% over 5 years (99, 100), which is higher than the incidence of CRC for people who have never been diagnosed with CRC (0.6% at age 50 years and 5.1% at age 70 years over 5 years). Future risk prediction models will need to be extended to include estimates of cancer risk subsequent to the first diagnosis if they are to be useful to those with a previous diagnosis of CRC.

Inherited CRC predisposition syndromes are seldom confined to the colorectum. Increased risks of various extracolonic cancers have been reported: cancers of the uterus, stomach, ovary, ureter, renal pelvis, brain, small bowel, and hepatobiliary tract in Lynch syndrome (101); cancers of the brain, thyroid, and liver in familial adenomatous polyposis (102); and cancers of the uterus and stomach in monoallelic MUTYH mutation carriers (103), and cancers of the duodenum, bladder, skin, and ovary in biallelic MUTYH mutation carriers (104). Even within families not known to be segregating a high-risk mutation, CRC risk is higher for those with a family history of extracolonic cancers (105–107). The effects of these other cancers in relatives on CRC risk should therefore be incorporated into risk models.

To determine individual risk of CRC precisely, it is important to develop a genetically based prediction model that incorporates the multigeneration family history and environmental risk factors, and allowing for the simultaneous effects of different MMR genes and other high cancer risk susceptibility genes including MUTYH, and the effect of low-penetrance genes and residual genetic or other familial risk factors. Only such a model can provide precise estimates of future cancer risks, and predict high-risk genetic mutation status for each individual across the full CRC risk spectrum. Validated risk prediction models should be easily assessable and user friendly for clinicians and genetic counselors as well as the general public by implementing them as a web-based application.

No potential conflicts of interests were disclosed.

A.K. Win was supported by a grant from the Cancer Council Victoria and the Picchi Brothers Foundation, Australia. R.J. MacInnis is a Sidney Sax Post Doctoral Research Fellow and J.L. Hopper is an Australia Fellow of the National Health and Medical Research Council (NHMRC), Australia.