By combining case–control data with regional incidence rates, we have developed a model to project individual 5-year absolute risks of developing lung cancer. The model has the potential to identify high-risk individuals by focusing on information that can be readily obtained in the primary care setting. The LLP risk model also appears to discriminate between high and low risk, although it will require rigorous validation in separate populations. As well as accounting for the three most important risk factors for lung cancer: age, sex and smoking, the LLP risk model incorporates other important disease risk factors such as family history of lung cancer, occupational exposure to asbestos, prior diagnosis of pneumonia and prior diagnosis of a malignant tumour other than lung cancer.
Similar to previously well-conducted cohort studies (Doll and Peto, 1978
; Flanders et al, 2003
), we have identified the importance of duration of cigarette smoking beyond the absolute amount of tobacco smoked. Indeed, contribution of smoking duration to the model did not change irrespective of whether 10- or 20-year categories were added. Therefore, the broadest category of smoking duration (20 years) was chosen for simplicity. Of the risk factors included in the LLP risk model, a prior diagnosis of malignant tumour is particularly interesting. First, it is important to emphasise that lung cancer cases and controls diagnosed with a malignant tumour (except melanoma) within 5 years of recruitment were excluded from the study. In addition, there were no significant effect modifications between history of previous malignancy and other risk factors in terms of their effect on lung cancer risk. Having had a previous malignancy was associated with a twofold increase in lung cancer risk. This is not unprecedented (Kabat, 1993
; Mery et al, 2004
). Previous studies have reported an increased risk of lung cancer among women receiving radiotherapy for breast cancer (Neugut et al, 1993
), possibly related to an interaction between radiotherapy and cigarette smoking (Prochazka et al, 2002
Sensitivity and specificity of the LLP risk model compares favourably with previous lung cancer absolute risk models developed by Bach et al (2003)
and Spitz et al (2007)
. The Bach model is based on a person's age, sex and smoking history, but it is predictive only for individuals between the age of 50 and 75, who smoked 10–60 cigarettes per day for 25–55 years. The Spitz model, like the LLP risk model, expands this concept by incorporating a panel of epidemiological risk factors to more accurately predict an individual's absolute risk of developing lung cancer. One limitation, however, is that cases and controls were frequency matched based on smoking status, perhaps affecting the importance of smoking as a risk factor. We believe that the LLP risk model's simplicity makes it more directly applicable for use in the primary care setting. Indeed, we are about to embark on a feasibility study in a large medical practice, which utilises the LLP risk model as the first stage in an early detection strategy, highlighting the potential importance and clinical relevance of our model.
An obvious strength of this study is that detailed information on the main risk factors such as active smoking, family history of lung cancer and occupational exposure was ascertained by closely supervised, trained interviewers using standardised questionnaires. No proxy interviews were performed. All cases had histologically or cytologically confirmed primary lung tumours. In common with all risk prediction models, the LLP risk model has several limitations. First, the absolute risks estimated for each combination of risk factors are based on relative risks derived from a case–control study. In this study, cases were individuals with newly diagnosed lung cancer identified by surveillance of all the hospitals responsible for the treatment of lung cancer in Liverpool and, after matching for age and sex, controls were selected from population lists that were essentially complete. Thus, cases and controls were both drawn from the same underlying population. However, the refusal rates were high, 42% among cases and 39% among controls. Given that these data did not permit description of smoking patterns of cases and controls refusing to participate, one cannot exclude that the lifestyle characteristics of nonparticipants may differ from the participants leading to an under- or overestimation of the true relative risks. There is also the potential that recall and other information biases could influence our results, as cases and controls were asked to report their lifestyle habits and behaviours for many years prior to interview. For these reasons, it is important that independent validation data be obtained to assess the relative risk features and absolute risk projections from this model. The limitations notwithstanding, this study can inform the debate about the best approach to select individuals at high risk of lung cancer for surveillance or prevention programmes.
Numerous nonrandomised studies have demonstrated that lung cancer can be diagnosed at a significantly earlier stage with CT screening than in current clinical practice (Humphrey et al, 2004
). Strategies combining smoking history as defined by pack-years have, for the most part, been used as an approach to more efficiently conduct screening in high-risk smoking cohorts (van Klaveren et al, 2002
), thereby excluding individuals at significantly elevated risk of lung cancer not adequately reflected using these factors. For example, a 60-year-old male with long smoking history would be included while a similar aged never smoker would not. Using the LLP risk model, the 5-year estimated risk for a 60-year-old male with 42 years of smoking and a family history in an affected relative aged 60 years or over is 3.73% (95% CI, 1.85–7.38) while the risk for a 60-year-old male with no smoking history, a family history with an affected relative aged under 60 years, a prior diagnosis of cancer, a prior diagnosis of pneumonia and exposure to asbestos is 3.52% (95% CI, 1.90–6.45). Both these individuals have almost identical risk estimates (a one in 28 and one in 27 chance, respectively of lung cancer in the next 5 years) even though one has never smoked. The LLP risk model potentially provides a means to identify subgroups of both the smoking and nonsmoking populations that may benefit most from prevention or surveillance.
Although some of the absolute risk estimates may seem rather high, they are consistent with population incidence. In the penultimate example in , a 77-year-old male, never smoker but with an early-onset family history of lung cancer and occupational exposure to asbestos, has an estimated 5-year risk of 3.17%. The general population risk in England and Wales for males aged 75–79 is approximately 2.5%. The modelled risk is slightly higher due to the two risk factors. It is not four times higher as one might anticipate from the odds ratios for early-onset family history and exposure to asbestos, because the general population on average has some exposure to the risk factors in the model. The two most extreme examples in have very high-predicted risks of around 30%, mainly but not entirely due to smoking. It should be noted that individuals with this level of risk are very rare; that is not more than 0.7% of cases and 0.3% of controls. This is reflected in the wide area of uncertainty in the confidence intervals, in both examples, showing a range of approximately 15–50% risk.
Although the results suggest that the LLP risk model may be useful for predicting risk, more work is needed to test the applicability of the model in diverse populations, including those from diverse geographic regions. It is clear that some populations, for example, African Americans, have risk factors other than those in our model (Abidoye et al, 2007
). Marked geographic differences in incidence rates necessitate separate evaluation of the LLP risk model in high- and low-risk areas. Moreover, developing separate models for men and women may allow for the inclusion of distinctive predictors and/or account for their variable distribution, thereby increasing predictive ability.
Although several issues concerning lung cancer risk prediction have been highlighted, we believe that its application as the first stage in an early detection strategy is a logical evolution in patient care. The results presented in this paper suggest that the LLP risk model could predict approximately two-thirds of lung cancer within 5-years, screening only 30% of the population. If resources were limited or the intervention carried such adverse effects as to require a very high-risk population to have a strong benefit–harm balance, a subset of 10% of the population could be identified in which 34% of the cases would arise. The effect of restricting screening to a subpopulation of high-risk individuals will markedly reduce the cost of screening programmes at the expense of missing a proportion of lung cancers in individuals below the cutoff. This ‘high-risk strategy' aims to help individuals with the greatest need of, and the potential to benefit from early detection. Such stratification on the basis of efficiency implies the difficult decision of where to place the cutoff, but the issue must be addressed nonetheless. It is likely that identifiable genetic susceptibility will, in the future, constitute an important factor in the selection of a more tightly defined risk group. In the meantime, it is appropriate to use a risk prediction model such as ours to identify a high-risk group for CT screening, so long as the results are not used to make inferences about a screening strategy in the general population (Baker et al, 2004
). If confirmed in validation studies, the LLP risk model could provide individuals and healthcare professionals with an easily obtained estimate of lung cancer risk to guide discussions and decisions regarding prevention and surveillance.