Development of a Cancer Risk Prediction Tool for Use in the UK Primary Care and Community Settings

Over recent years, there has been a growth in the development of risk prediction models for cancer and other diseases (1–6). These models provide a range of estimates of future risk of developing disease for applications in prevention, screening, diagnosis, and treatment. Most of them are disease specific. In general, risk algorithms include phenotypic information such as sex, age, and lifestyle factors. Some algorithms may also allow for the incorporation of emerging “omics”-based factors and other biomarkers (7).

Few cancer risk prediction models have included the modifiable risk components such as physical activity, diet, and smoking. One such model that has been developed for the United States is the "YourCancerRisk" model which has been used as a tool for education as well as providing an approach to quantifying the effects of changing key lifestyle exposures. This was subsequently expanded into "YourDiseaseRisk" (8), which extends the range of endpoints to include 12 of the commonest types of cancers in the United States and 6 other chronic diseases (e.g., chronic bronchitis, stroke, emphysema, heart disease, diabetes, and osteoporosis). The model was validated for ovarian, colon, and pancreatic cancers in the United States in Nurses’ Health Study and the U.S. Health Professionals cohorts. The results show the model to be well calibrated for ovarian and colon cancer in women and pancreatic cancer in men and moderately calibrated for colon cancer in men. Discriminatory accuracy for pancreatic cancer showed a concordance index of 0.72, and for colon cancer in men and women, concordance indices were 0.71 and 0.67, respectively (9).

The “YourDiseaseRisk” model aims to predict the risks for individuals (aged 40 and above) of developing the 12 cancers relative to the general population. Uniquely, the approach adopted to develop such models involved extensive systematic reviews of existing studies and finding a consensus of expert opinions to identify risk factors and to the summarize the level of evidence as “definite,” “probable,” and “possible” causes of cancer. Risk points were then allocated according to the strength of the causal association and summed. Population average risk of cancer and cumulative 10-year risk were obtained from the U.S. SEER data (10). Finally, individual ranking relative to the population average was determined.

The “YourDiseaseRisk” online tool has been available in the United States since 2000. It is offered as an educational tool, and in 2005, the site recorded 54 million hits with 6.2 million page views. It was first hosted at Harvard University and, in 2007, transitioned to the Siteman Cancer Center at Barnes-Jewish Hospital and Washington University School of Medicine (CITE Siteman/Wash U/BJH only; ref. 8).

The focus of this article is to describe the steps taken to adapt the "YourDiseaseRisk" models focused on cancers for the UK population for use in Primary Care and community-based settings. We also assess the utility of the approach by scoring the suggested risk factors in the UK Biobank cohort.

We have used the adapted UK version of the "YourDiseaseRisk" models (8) in a pilot study that used individual interviews to assess participants’ understanding and preferences for how such information is offered (publication in press). These results will be presented elsewhere.

We adapted the YourDiseaseRisk models for 11 cancers using the UK data. The 11 cancers chosen were lung, prostate, breast, kidney, bladder, colon, skin, stomach, pancreatic, uterine, and ovarian cancers. Cervical cancer was not included as it required participants to disclose information on sexual history which was considered too sensitive.

In general terms, the information required to develop prediction models are: (1) the identified list of risk factors for inclusion; (2) point estimates of the relative risk for each risk factor; and (3) population prevalence for each of the exposures. To be able to compare individual risk to the population, further information such as cancer incidence by 10-year age bands is required.

To illustrate the comparative relative risks (RR) between the two populations, we used the UKBiobank national cohort and analyzed the RRs for breast cancer. The UKBiobank female cohort consists of 273,467 women with age ranged between 40 and 69 years when recruited. Participants were enrolled in the UK Biobank from April 2007 to July 2010, from 21 assessment centers across England, Wales, and Scotland using standardized procedures (11). The UK Biobank study was approved by the North West Multi-Centre Research Ethics Committee, and all participants provided written informed consent to participate in the UK Biobank study. To date, the cohort has been followed up for 6 years. We computed RRs adjusted only for age. The results are presented in Table 6.

To demonstrate model validation, we selected breast cancer risk prediction as an example. The UKBiobank cohort was used for the validation exercise. For breast cancer cases, we used ICD10, ICD9, and self-reported codes (verified by the UKBiobank health professionals), and only incident cases were included in the analysis. For controls, we used two comparison groups. Firstly only those subjects with no cancer code recorded in ICD10, ICD9 and secondly the coded with no cancer and no other self-reported illness. The total number of incident breast cancer cases was 3,378, and the total number of noncancer controls was 235,603 or healthy controls was 59,731. We coded all variables (except Tamoxifen/Raloxifene usage as such data were not available) present in the breast cancer risk prediction model based on the presence or absence of the exposure for each individual as illustrated in Table 3. To demonstrate model validation, we calculated the area under the curve of the model based on including all factors scored as a binary variable and generated calibration plots -the observed and expected proportions compare within the groups formed by the Hosmer–Lemeshow test. All analyses were performed using STATA 14 (12).

To develop the UK version of the “YourDiseaseRisk” models, we have assumed that the risk factors for these cancers in the United States and UK are the same and have obtained the list of risk factors and point estimates of relative risk for each risk factor from the U.S. “YourDiseaseRisk” models. Since the "YourDiseaseRisk" model was first developed more than a decade ago and to ensure we use the most updated version, we started the process by extracting risk factors for each cancer from the "YourDiseaseRisk" online version (ref. 8; Step 1). Once we created a list of risk factors for each cancer, we assigned point estimates of RR for each factor (Step 2). These point estimates of RR's were extracted from the original publication (13, 14); however if any factors did not have any cited value, we performed a literature search to obtain any missing estimates. To maintain consistency across RR values, we used figures from publications by the Colditz study group and those with a cohort study design (examples of references are depicted in Table 1). For example, as the multivitamin factor was not listed in the original article, we therefore chose and applied an RR of 0.7 from Zhang and colleagues (15).

As the original report applied RRs from compiling evidence from the U.S. cohort studies over a period of time, it is important to justify the use of the RRs published by Colditz and colleagues (14) in the UK models. We demonstrated the similarities/variations of risks between the two populations (US and UK). To do this, we have presented as an example the RRs for breast cancer derived from the UKBiobank study (Table 2). The majority of the point estimates are similar and convert to the same risk score when using values in Table 3. The two exceptions were for multivitamin use and physical activity where the protective effects for both of these factors were less pronounced in the UK as compared to the U.S. estimate.

To obtain the population prevalence (step 3) for each of the exposures, we then reviewed the literature on the UK prevalence of each factor for men and women for all 11 cancers. The criteria for publication selection included (1) UK prevalence data from National surveys (16) or prevalence derived from large cohort studies representative of the general population in the UK, or (2) if no data were available from those sources, information/figures from cohort studies in European countries. As an example, the colon cancer risk factors, RR, and references chosen for prevalence are shown in Table 1.

Once information on the UK prevalence of each risk factor was obtained, we then applied a score to each risk factor using the same scheme as presented in the original article (summary as shown in Table 3).

This risk score is used to compute two further scores—the population average risk score and an individual risk score relative to the population average. The population average score is calculated by multiplying the risk score of each factor by the population prevalence of that particular factor. To prevent negative scores, we chose the direction of each risk factor to make the population average score the highest possible. Taking physical activity, for example, the prevalence of carrying out 3 or more hours of total leisure-time physical activity per week in the UK population is 23%. This figure means that 77% of the population do not do physical activity regularly at this particular level. When we apply a prevalence of 77%, then the assigned score instead of being –10 (for those who are doing regular exercise) will be +10. This conversion allows us to demonstrate the change in the individual risk score following any change in factors that are modifiable. Summation of these scores produces the average population score.

The risk score for a given individual relative to the population average is based on the presence or absence of each factor. An example is illustrated in Table 4. The summation of scores for each risk factor for an individual provides the total risk score for a particular person. That total risk score is then divided by the average population score to give an individual index that is relative to population score (35/27= 1.3). As mentioned earlier, participants can see how their risk changes if they adopt suggested behaviors, for example, by choosing to do regular exercise, the total score for the individual illustrated in Table 4 can be reduced by 10 points (from 35 to 25), making their index score relative to the population reduced from 1.3 to 0.8. This calculation aims to illustrate to each individual the effect of a particular lifestyle or behavior change leading to cancer risk reduction.

The index score for an individual can then be further transformed into a level of risk. A numeric factor indicative of the strength of the risk level is assigned to the individual index score (Table 5). This is done to give an average value of the range of individual index scores as a single numeric estimate that reflects the risk level.

For the individual in Table 3, for example, the individual index score is 1.3, which is equivalent to “above average risk” and gives a numeric factor of 1.5.

To enable estimation of an individual's 10-year estimated cancer risk, we calculated the average 10-year estimated risk for different ages and sexes of the UK population for all 11 cancers. We used the “Current Probability” method proposed by Esteve and colleagues in 1994 (17). This method uses a life-table approach for calculating the risk of developing cancer and takes into account the likelihood of dying from other causes. The method also requires information on deaths from all causes for each age group. This method provides estimate of the 10-year risk of cancer.

We obtained age- and sex-specific cancer incidence and mortality rates and numbers from Cancer Research UK (18) and age- and sex-specific data on all-cause mortality from the Office of National Statistic (ONS) which are available online (19). The following shows the specific data used for the calculation:

1. The annual number of cancer deaths (cancer mortality for males and females); data from CRUK 2010–2012.

2. The annual number of (registered) cancer cases for males and females; data from CRUK 2009–2011.

3. The annual number of deaths (all-cause mortality for males and females); data from ONS 2011.

4. The size of the mid-year population for males and females; data from ONS 2011.

From these data, we computed 10-year cancer risk for each cancer in males and females in 10-year age bands from birth to over 80 years.

Assuming the individual in Table 4 is a 40-year-old woman, the average 10-year risk of breast cancer (age 40–49) is 1.45% in the next 10 years (Table 6). Multiplying this figure by the numeric factor of 1.5 (Table 4) will give a risk of 2.2% (above the population average) or approximately 1 case in 45 women.

The AUC for breast cancer risk prediction model based on utilizing a binary score for each available risk factor in the UKBiobank dataset was 0.58 [95% confidence interval (CI), 0.57–0.60] for a comparison group of controls with no cancer and 0.64 (95% CI, 0.63–0.66) for a comparison group of controls with no cancer or other illnesses. Model calibration curves for both comparisons also suggested both models calibrated well (Fig. 1).

The “YourDiseaseRisk” model has been developed using three key data including an estimate of relative risk of each risk factor, the prevalence figure of exposure in the population, and the 10-year estimated cancer risk. These figures can be acquired through literature review, National data archives, or other organizations that compile these data publically (20).

The “YourDiseaseRisk” site has been launched since 2000 and update/review of information is an ongoing process. The model is an educational tool and can be used for cancer prevention in clinics, community settings, or by individuals simply seeking information on their own. The popularity of “YourDiseaseRisk” is reflected by the large number of hits or page views.

We have adopted their methodology and applied it to build a UK version and to demonstrate the applicability of the approach for other populations as well. The main strength of the approach is the use of large population-based studies to estimate the parameters needed to generate the model.

To demonstrate the relative validity of the approach, we chose breast cancer as an exemplar. We have demonstrated that the quantitative estimates of the individual risk factors apply to current UK population using data from the UKBiobank study. Furthermore, we performed a model validation of the combined risk factors using a binary scoring system, and the results suggest such a model is reasonably calibrated and has moderate discriminatory power [0.58 for a comparison group of controls with no cancer and 0.64 (95% CI, 0.63–0.66) for a comparison group of controls with no cancer or other illnesses]. The prediction performance of a fully specified model using the point estimates of each risk factor rather than a binary value is likely to be higher as it will be derived from more precise estimates of their individual effect. Directly comparative data are not yet available in the UKBiobank for all parameters to undertake such analyses.

There are many breast cancer risk prediction models, but only a few that contained epidemiologic factors and most have similar predictive capabilities (21–30). The majority of these models are extended versions of the Gail model (23). They have been shown to have good calibration but moderate discrimination ranging from 0.56 to 0.89 with few having been assessed in external validation analysis (23, 28, 31, 32). Colditz and colleagues further reported an AUC of 0.64 (95% CI, 0.62–0.66) in their extended validation study using data from the Nurses' Health Study (32). The validation exercise, based on breast cancer example presented here, supports the conclusion that the model can be used in the UK population when the prevalence of risk factors is substituted with the UK figures.

There are, however, potential limitations to the approach that need to be taken into consideration. Firstly, although we were systematic in our literature reviews to obtain data on the point estimate RRs for risk factors missing from the original published “YourDiseaseRisk” model and the prevalence of all risk factors in the UK population, the studies we have chosen may not provide the most accurate estimates. Secondly, we have assumed that the prevalence of each risk factor does not change with age, that the risk associated with each risk factor is the same across all ages and both sexes, and that the risk factors do not interact with each other. These assumptions were also made in the development of the “YourDiseaseRisk” models but, as noted in that report (14), may result in misclassification of risk for exposures across large age ranges and underestimate possible synergistic effects of exposures such as alcohol and smoking. As for the “YourDiseaseRisk” models, therefore, these models should be considered as a guide for assessing an individual's risk of cancer in the UK rather than a precise estimate.

Risk prediction models are widely used for many diseases. With the concept that some cancers can be preventable, the “Your Disease risk” model developed by Colditz and colleagues (14) provides a good platform because their model is based on modifiable factors that have been scrutinized and carefully selected by a panel of experts. The approach has wide utility in allowing for the rapid development of models for educational purposes.

In this article, we have described how we have adapted that model to produce a UK version for 11 cancers using data from the UK population. We are now using the model in the community in the UK. Going forward, we will evaluate how the tool affects perceptions of risk, how to best present the risk, and how the public understands their individual risk. This information will then inform future studies in the community exploring the potential for the use of this model to promote lifestyle change. Finally, we will extend and further validate this risk prediction in the UKBiobank cohort as more precise individual level data and longer term follow-up data become available.

J. Warcaba is Macmillan GP. No potential conflicts of interest were disclosed by the other authors.

Conception and design: A. Lophatananon, J. Usher-Smith, J. Campbell, J. Warcaba, G.A. Colditz, K.R. Muir

Development of methodology: A. Lophatananon, J. Usher-Smith, J. Campbell, J. Warcaba, G.A. Colditz, K.R. Muir

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. Warcaba

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A. Lophatananon, J. Usher-Smith

Writing, review, and/or revision of the manuscript: A. Lophatananon, J. Usher-Smith, J. Campbell, J. Warcaba, B. Silarova, E.A. Waters, G.A. Colditz, K.R. Muir

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): A. Lophatananon

Study supervision: J. Usher-Smith, J. Campbell, K.R. Muir

We would like to thank Kawthar-Al-Ajmi, PhD student at the University of Manchester, for her help with preparation the UKBiobank data. This research has been done using the UK Biobank resources application number 5791.

A. Lophatananon, K.R. Muir, J. Usher-Smith, J. Campbell, J. Warcaba, and B. Silarova were funded by an Innovation Grant from the Cancer Research UK – BUPA Foundation Fund (C55650/A20818). A. Lophatananon and K.R. Muir were also funded by ICEP (“This work was also supported by CRUK [grant number C18281/A19169]”). J. Usher-Smith was also supported by the National Institute for Health Research Clinical Lectureship, and B. Silarova was supported by the Medical Research Council (MC_UU_12015/4). E.A. Waters and G.A. Colditz are funded in part by the Foundation for Barnes-Jewish Hospital, St Louis, MO.

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.