Summary of key findings

We have developed and validated a series of risk prediction algorithms—collectively known as the QCancer 10 year risk algorithms—to quantify future absolute risk of 10 common cancers in women and eight common cancers in men. The algorithms incorporate predictor variables which are associated with risk of cancer including sociodemographic variables, lifestyle, morbidity, medications, family history and previous diagnoses of other cancers. The algorithms can be applied to any adult aged 25–84 years in a primary care setting regardless of whether they have had a prior cancer. For nine of the 11 cancers, the ROC values exceeded 0.8 which is generally considered to be very good. For two of the cancers in women (breast and ovarian) the ROC values were lower at 0.76 and 0.77 though this is still considered acceptable.

The QCancer 10 year risk algorithms are designed to quantify future risk of cancer and differ from the existing QCancer algorithms3–10 which combine symptoms such as appetite loss and abdominal pain with risk factors to quantity the absolute risk that a patient has an existing cancer as yet undiagnosed to help inform the decision for further diagnostic tests. These algorithms to predict existing cancer are designed to be used when patients present with symptoms that might be indicative of cancer.

Comparisons with the literature

We included established predictor variables in our analysis and found that our HRs were similar in magnitude and direction to those reported in other studies. This increases the clinical face validity of the variables included as predictors in the final algorithms. We have summarised the key relationships below from the perspective of the relevant set of risk factors (rather than by each cancer as described in our results section above).

Thresholds

Generally we envisage that cancer risk prediction values would be kept continuous for assessment of an individual although at some point there needs to be a cut-off if a clinician is going to take action for an individual patient. At national level, policymakers and commissioners tend to make recommendations around absolute thresholds to ensure equitable access and consistent management across a health community. This is already common place for cardiovascular disease risk where the latest National Institute for Health and Care Excellence (NICE) guidelines from 2014 recommend treatment with statins for patients with a 10-year cardiovascular disease risk above 10%.79 However, in this paper, we have not provided definite comment on what threshold of absolute risk should be used to define ‘high risk’ since that would require (1) consideration of the balance of risk/benefit for an individual and their choice and (2) cost-effectiveness analyses which are outside the scope of this study. We have, however, provided analyses using the top 10% of absolute risk as a threshold of risk which can be used to help inform future analyses. Sensitivity is important as it is a measure of how well the algorithm performs in finding cases that might be suitable for intervention. If the risk threshold is set too high, then the sensitivity will be low and a large number of patients at increased risk of cancer will be ‘missed’ by the algorithm. Although a high-risk threshold is likely to result in a higher positive predictive value which means a higher proportion of those identified are likely to go on to develop cancer over the next 10 years. A lower risk threshold would have higher sensitivity but could lead to unnecessary interventions and anxiety in people who will not develop cancer over 10 years. So, at the population level, there is a balance to be struck between the sensitivity and positive predictive value of the score which depends on the risk threshold selected, resources available; likely risks and benefits of the interventions and the chance that patients might become anxious about being classified as ‘high risk’.

Methodological considerations

The methods to derive and validate these models are the same as for a range of other clinical risk prediction tools derived from the QResearch database.19 ,80–83 The strengths and limitations of the approach have already been discussed in detail21 ,80 ,83–86 including information on multiple imputation of missing data. In summary, key strengths include size, duration of follow-up, representativeness, and lack of selection, recall and respondent bias. UK general practices have good levels of accuracy and completeness in recording clinical diagnoses and prescribed medications.87 We think our study has good face validity since it has been conducted in the setting where the majority of patients in the UK are assessed, treated and followed up. Our database has linked hospital, mortality and cancer records for nearly all patients and is therefore likely to have picked up the majority of cancer diagnoses thereby minimising ascertainment bias. We excluded patients without a valid deprivation score since this group may represent a more transient population where follow-up could be unreliable or unrepresentative. Their deprivation scores are unlikely to be missing at random so we did not think it would be appropriate to impute them.

The present validation has been carried out on a completely separate set of practices and individuals to those which were used to develop the score although the practices all use the same GP clinical computer system (EMIS—the computer system used by 55% of UK GPs). An independent validation study would be a more stringent test and should be carried out, but when such independent studies have examined other risk algorithms,85 ,86 ,88 ,89 they have demonstrated comparable performance compared with the validation in the QResearch database.19 ,80 ,84 We (or another academic team) intend to conduct a separate validation on an independent database using a different clinical system (CPRD) as part of a separate project.

Other limitations of our study include the lack of formal adjudication of cancer diagnoses, information bias, and potential for bias due to missing data. However, we think our ascertainment of cancer is likely to be high given the combination of the four data sources and the similarity between our rates and those published by the Office for National Statistics.11 While we acknowledge that there may be discrepancies between self-reported features (such as number of cigarettes smoked), the information recorded on the GP record has a face validity as it is still predictive of relevant outcomes, exhibits a ‘dose-response’ relationships (with higher doses being associated with increased cancer risks) and similar to HRs from the published literature as we describe above. We have also only based behavioural variables (for smoking and alcohol) on a simple assessment of current exposure rather than more complex measures reliant on recall of past exposure which would be less reliable.

We measured family history of cancer based on information which may have been recorded opportunistically rather than systematically. Patients with a positive family history of cancer may be more likely to report this to their GP and for this to be recorded and we have assumed that where no positive family history is recorded then the patients has a negative family history which will lead to some misclassification. Also, there is no Read code to record family history of some types of cancer (eg, pancreatic cancer) so we were unable to include this in our list of predictor variables as this information is not captured in coded form and available on the QResearch database. Overall, the recording of family history is therefore likely to be subject to both ascertainment and recording bias. However, it is likely that the recording of family history will improve over time particularly if (1) clinicians can be prompted to ask for this information by the use of structured clinical templates which can be offered to the clinician for use during the consultation or if (2) patients can record this information directly once they are able to access their medical record electronically.

Genetic information is likely to be important in a full assessment of cancer risk although the information is not currently routinely recorded in electronic health records. Genetic information cannot therefore be used either to derive or validate a new score using large representative UK clinical databases. The lack of routinely available genetic results would also cause practical problems in implementing a tool dependent on such variables into clinical practice. Previous studies examining the benefits of adding genetic variations to risk assessment tools have shown only modest improvements in the predictive power. For example adding genetic variant information to the Gail Breast Cancer Risk assessment tool only increased the ROC value from 0.58 to 0.62.90 The ROC value for the breast cancer algorithm based on clinical data in our study was 0.76 which is significantly higher than the ROC value of 0.62 reported for the generic variant of the Gail model.90

A recent study by Tomasetti et al 91 concluded that “only a third of the variation in cancer risk among tissues is attributable to environmental factors or inherited predispositions. The majority is due to ‘bad luck’, that is, random mutations arising during DNA replication in normal, noncancerous stem cells”. Our study focuses on methods to predict absolute risk of cancer taking account of the individual's age, sex, ethnicity, lifestyle, family history and comorbidities. We found that our models explained well over 33% of the variation in cancer risk for 9 of the 11 cancers studied as shown by the R² values presented in table 8. This would tend to refute the hypothesis by Tomasetti et al 91 that only a third of the variation in cancer risk among tissues is attributable to environmental factors or inherited predispositions with the remainder being due to random mutations for the majority of cancers. However, our models only explained 22% of the variation for breast cancer and 29% for ovarian cancer which would support it.

The prediction models in our study have been derived from routinely recorded electronic health data using variables which are accessible in everyday practice in order that they can be applied in real world clinical settings to identify high-risk patients for further screening or prevention. Overall the validation statistics indicate that performance of all the algorithms is either good (breast, oral and ovarian cancers have ROC values 0.76–0.79); very good (blood, bowel, gastro-oesophageal, pancreas prostate, renal, uterine cancers have ROC values of 0.80 to 0.89) or excellent (lung cancer has an ROC value >0.9). Our prediction models have included clinical and lifestyle variables (such as age/sex/ethnicity/coexisting illnesses/smoking/alcohol and medication). They also include family history which was a significant predictor for six cancers (prostate, breast, bowel, blood, lung, ovarian). While there is clearly a difference between family history information and a genetic sample, family history does reflect genetic factors and is routinely recorded in general practice settings. It would be impractical, both from a cost and governance perspective to undertake genetic tests on all patients so that this information could be included in derivation of the models. It would also be impractical to require genetic information to be available routinely for when the scores are applied. It would be possible, however, to use the QCancer 10 year risk tool to identify a subset of high-risk patients for whom further genetic testing might be warranted and to quantify the level of absolute risk to help inform patient choice.

The QCancer 10 year risk algorithms have been developed using linked data from general practices in England and some of them (breast, gastro-oesophageal, lung, oral, pancreatic, prostate and renal tract) include a postcode related deprivation score (Townsend score). We included the Townsend score in the algorithms since there were clear relationships between deprivation and risk of some cancers which is captured by this variable. If we omitted it, it would tend to under-estimate cancer risk in patients from deprived areas for most cancers except breast cancer and renal cancer in women and prostate cancer in men, where the risk may be slightly over-estimated. In terms of international use of these algorithms where the Townsend score is not available a locally relevant deprivation score could be constructed or adapted for the relevant country to have a range between −7 (most affluent) and +11 (most deprived) to correspond to the range of Townsend scores which could be used instead (subject to local validation). For other predictor variables such as age, smoking, alcohol, family history, prior cancer, then we have compared the magnitude of the HRs to other international studies and found them to be similar in direction and magnitude (see Comparison with literature). This would tend to support the utility and face validity of the algorithms outside the UK although best practice would be that they are externally and independently validated in the settings where they would be used to ensure they are appropriately calibrated and have good discrimination.

Lastly changes in environmental factors may occur over time and this underlines the need to update the prediction algorithms on an ongoing basis as has been carried out with QRISK292 cardiovascular score and other related prediction scores.93 Regular updates, with a moving time window, will also help ensure that the algorithms will benefit from improvements in the scope and quality of the underlying database which is likely to occur over time. This is an important strength of using routine databases for the development of risk prediction algorithms that is not feasible with prospective study cohorts that are assembled at one point in time.

Clinical implications

The algorithms have been designed to work in a primary care setting, making use of information which is already recorded on the GP clinical computer system. The algorithms can be integrated into the clinical computer system alongside similar algorithms which already quantify risk of other clinical conditions in everyday clinical practice such as QRISK2,19 QDiabetes,80 QStroke,94 QFracture,81 QThrombosis,83 QBleed95 and QAdmissions.12 They can be used in ‘batch process’ mode to generate a list of patients at high risk of each cancer for further assessment or they could be used within the consultation. For example, policy makers and commissioners are likely to want to use the tumour site specific algorithms in ‘batch mode’ to risk stratify populations to better target screening programmes.

To indicate the potential value of using the QCancer tool for identifying a high-risk population for screening compared to an approach based on risk factors alone, we calculated the sensitivity and predictive values for lung cancer in heavy and moderate smokers. In the validation data set for women there were 3007 lung cancer cases over 10 years, and 812 occurred in moderate/heavy smokers (9.3% of cohort) giving a sensitivity of 27%, this contrasts with a sensitivity of 67% for the top 10% of women at highest predicted risk of lung cancer using the QCancer score (table 9). In men the corresponding sensitivity values are 22% in moderate/heavy smokers compared with 67% using QCancer (table 10). So QCancer is better at identifying future lung cancer cases than an approach targeting moderate/heavy smokers would be. Further work would be needed to determine any risk thresholds for screening, which as we state in the paper would require cost-effectiveness analyses which are outside the scope of this study and consideration of available resources and impacts on patients.

The algorithms could help inform the discussions between doctor and patient within the consultation regarding the future risk of cancer associated with existing diagnoses (such as the risk of bowel cancer among patients with ulcerative colitis); prior cancers (which might indicate an increased risk of a second primary cancer); family history of cancer (where additional screening may be justified) or lifestyle related variables (such as BMI, smoking and alcohol intake which can be moderated). For example, the doctor could use the algorithms to assess the patient's 10 year risk of cancer to highlight the higher risk associated with heavy or moderate smoking compared with that for lighter smokers or ex-smokers. Currently there is no easy to use widely available calculator, such as that described in this paper, which will allow a patient to quantify their absolute risk of getting different types of cancer taking account of their age, sex, family history and other risk factors such as alcohol consumption and smoking status. While identifying effective interventions to reduce alcohol intake or increase smoking cessation remains a challenge, there is evidence96 that physician advice has some effect on smoking cessation rates albeit small and that brief alcohol interventions in primary care populations can reduce alcohol consumption.97 It is possible that providing patients with information on how these factors influence their personal cancer risk might have an additional impact on smoking and alcohol consumption. There is also some evidence to support the use of biomedical risk assessment feedback such as ‘lung age’ to increase rates of smoking cessation98 which suggests that patients may respond to information presented in an accessible format.

As another example, the doctor and patient could also review treatments such as the use of HRT in a woman at higher risk of breast cancer due to other factors such as family history. However, if a patient presents with potential symptoms of cancer such as appetite loss or haematuria the existing QCancer scores we developed previously3–10 would be more suitable for assessing current cancer risk and informing decisions regarding further investigation and referral.