Identification of high-risk individuals will facilitate early diagnosis, reduce overall costs, and also improve the current poor survival from lung cancer. The Liverpool Lung Project prospective cohort of 8,760 participants ages 45 to 79 years, recruited between 1998 and 2008, was followed annually through the hospital episode statistics until January 31, 2013. Cox proportional hazards models were used to identify risk predictors of lung cancer incidence. C-statistic was used to assess the discriminatory accuracy of the models. Models were internally validated using the bootstrap method. During mean follow-up of 8.7 years, 237 participants developed lung cancer. Age [hazard ratio (HR), 1.04; 95% confidence interval (CI), 1.02–1.06], male gender (HR, 1.48; 95% CI, 1.10–1.98), smoking duration (HR, 1.04; 95% CI, 1.03–1.05), chronic obstructive pulmonary disease (HR, 2.43; 95% CI, 1.79–3.30), prior diagnosis of malignant tumor (HR, 2.84; 95% CI, 2.08–3.89), and early onset of family history of lung cancer (HR, 1.68; 95% CI, 1.04–2.72) were associated with the incidence of lung cancer. The LLPi risk model had a good calibration (goodness-of-fit χ2 7.58, P = 0.371). The apparent C-statistic was 0.852 (95% CI, 0.831–0.873) and the optimism-corrected bootstrap resampling C-statistic was 0.849 (95% CI, 0.829–0.873). The LLPi risk model may assist in identifying individuals at high risk of developing lung cancer in population-based screening programs. Cancer Prev Res; 8(6); 570–5. ©2015 AACR.

Lung cancer is the leading cause of cancer-related death worldwide, with mortality rate exceeding that of prostate, breast, and colon cancer combined (1). It has one of the poorest survival outcomes of any cancer because a vast majority of patients are diagnosed at an advanced stage when surgical resection or other treatment options are less effective (2). The National Lung Screening Trial (NLST) showed that lung cancer screening with low-dose computed tomography (LDCT) reduced lung cancer mortality by 20% (3). In practice, the success of any lung cancer screening program will depend on the successful identification of individuals at high risk. Risk prediction models have been recognized as a method of identifying individuals at high risk of developing lung cancer (4, 5). The identification of individuals at high risk will facilitate early diagnosis, reduce overall costs, and also improve the current poor survival from lung cancer.

To date, most risk prediction models for lung cancer were developed in case–control studies (6–9). Case–control studies are proficient in studying dynamic populations where follow-up is difficult, are usually less expensive, and are less time consuming, but may be plagued by biases and cannot study the incidence of a disease (10–13). Cohort studies, although time consuming and expensive, offer the methodological advantage of direct calculation of incidence rate and can demonstrate the temporal sequence between exposure and outcome (14, 15).

We previously developed the LLP risk model from the LLP case–control study and validated it in three independent populations (16). Covariates such as prior history of respiratory diseases and prior diagnosis of malignant diseases have been reported as risk factors for lung cancer (16, 17). Clinical covariates in the previous case–control model were based on unvalidated self-reported questionnaire responses (18). However, in the newly developed model reported in this article, clinical covariates were confirmed in the Hospital Episode Statistics (HES) database. HES is the national statistical data warehouse for the National Health Service (NHS) that includes clinical and administrative information about the care provided to NHS patients who live or are treated in England (19). The aim of the current study was to use baseline data from the LLP population-based cohort to develop and validate a risk prediction model for lung cancer incident.

Study population

This study was performed as part of the Liverpool Lung Project. The objectives, methods, rationale, and study design have been described previously (18). In short, 8,760 randomly selected healthy subjects ages 45 to 79 years were recruited between 1998 and 2008 and followed annually for lung cancer and mortality outcomes through the Office for National Statistics (ONS), public health England (the North West Cancer Intelligence Service), and hospital case-note review until January 31, 2013.

Data collection and extraction of risk factors

A standardized questionnaire was used to collect self-reported information on demographic and socioeconomic economic characteristics, medical history, family history of cancer, history of tobacco consumption, and lifetime occupational history. Information on age, gender, smoking duration, marital status, education level, family history of lung cancer, prior history of other cancers, and prior history of nonmalignant lung disease such as asthma, chronic obstructive pulmonary disease (COPD), pneumonia, bronchitis, emphysema, tuberculosis, and exposure to asbestos was extracted from the questionnaire. Smoking duration was measured in years; ever smoker was defined as someone who had smoked at least 100 cigarettes in their lifetime and a current smoker was defined as a participant who reported smoking within 12 months of the date of the interview. Marital status was reported as single, married, living together, widowed, divorced/separated, or other. Education was classified as high school and below and greater than high school and recorded as “yes” or “no.” Family history of lung cancer included age at onset in a first-degree relative [none, early (<60 years), or late (>60 years)]. Previous history of cancer (except melanoma) was coded as “yes” or “no.” Information on prior history of nonmalignant lung diseases such as asthma, COPD, pneumonia, and tuberculosis was coded “yes” or “no,” and diagnosis was corroborated in the HES database prior to their recruitment into the study. Asbestos exposure was determined based on the documentation of participant's employment history in asbestos-related occupation or industry and was coded “yes” or “no.” The study protocol was approved by the Liverpool Research Ethic Committee, and all research participants provided written, informed consent in accordance with the Declaration of Helsinki.

Statistical analyses

Descriptive statistics were obtained and compared by using the χ2 test or the Fisher exact test for categorical variables and t tests for normally distributed variables, between participants who developed lung cancer (n = 237) and those who did not (n = 8,523) at the end of the follow-up period. The Cox proportional hazards model was used to develop a multivariable model for lung cancer risk and to compute hazard ratios (HR) and corresponding confidence intervals (CI) after testing the proportional hazards assumption with the unscaled and scaled Schoenfeld residuals (20). Follow-up duration was used as the time axis in the model. For participants who developed lung cancer, the follow-up time started from baseline visit, until the date of diagnosis of lung cancer. For participants without incident lung cancer, the follow-up duration was counted from the baseline visit until death, withdrawal, loss to follow-up, or January 31, 2013. Five percent of the 8,760 participants had missing data; we therefore performed complete case analysis because the characteristics of participants with missing data were largely similar to those of participants with complete observed data. Martingale residuals were used to determine the best function of all covariates (21). The multivariable model was built in two phases. First, all covariates with P ≤ 0.10 in the univariate analyses were considered for inclusion in the multivariable model. Second, backward selection procedure with (P < 0.05) was used to choose the covariates in the final multivariable model (22). Covariates eliminated were reentered in the final multivariable model, with adjustment for the remaining significant covariates to ensure that no omitted covariate significantly reduced the log likelihood χ2 of the model (23). The performance of the multivariable model was quantified by Harrell's concordance (C) statistics, which is analogous to the receiver-operating characteristics (ROC) curve for binary data with confidence interval estimates using the jackknife resampling method (24). A C-statistic can be interpreted as the probability that the model predicts a higher risk of lung cancer for those who actually developed lung cancer compared with those that did not develop lung cancer over the follow-up time (25). Model calibration was evaluated using the omnibus Grønnesby–Borgan goodness-of-fit test. Bootstrapping techniques were used for internal validation of the model (26, 27), and bootstrap samples were drawn 200 times with replacement. Regression models were created in each bootstrap sample and tested on the original sample to obtain stable estimates of the optimism of the model, i.e., how much the model performance was expected to decrease when applied in new datasets (28–30). All analyses were performed using STATA version 13.1 (StataCorp LP) and SAS statistical software version 9.3 (SAS Institute).

Overall, 8,760 participants ages 45 to 79 recruited between 1998 and 2008 were included in this study and 237 participants (2.7%) developed lung cancer during a mean follow-up of 8.7 years. Table 1 depicts the baseline characteristics of the study population stratified by incident lung cancer status. Lung cancer incidence was higher in older participants, men, and also in participants with longer smoking duration, participants with prior history of pneumonia, asthma, tuberculosis, COPD, emphysema, and bronchitis. Furthermore, the percentage of lung cancer incidence was higher in participants with a family history of lung cancer and prior diagnosis of malignant tumor.

Table 1.

Distribution of baseline characteristics stratified by lung cancer status

CharacteristicsIncident lung cancer (n = 237)No. of incident lung cancer (n = 8,523)P
Mean age (SD) 66.0 (7.2) 61.5 (8.5) <0.001 
Gender   0.002 
 Male 137 (57.8) 4,052 (47.5)  
 Female 100 (42.2) 4,471 (52.5)  
Mean smoking duration (SD) 39.9 (14.3) 18.9 (19.0) <0.001 
Prior diagnosis of pneumoniaa   0.066 
 No 186 (83.0) 6,971 (87.2)  
 Yes 38 (17.0) 1,022 (12.8)  
Prior diagnosis of asthmaa   0.011 
 No 173 (77.2) 6,687 (83.6)  
 Yes 51 (22.8) 1,310 (16.4)  
Prior diagnosis of tuberculosisa   0.674 
 No 216 (96.4) 7,745 (96.9)  
 Yes 8 (3.6) 246 (3.1)  
Prior diagnosis of COPDa   <0.001 
 No 90 (47.6) 5,196 (82.6)  
 Yes 99 (52.4) 1,094 (17.4)  
Emphysemaa   <0.001 
 No 203 (90.6) 7,718 (96.5)  
 Yes 21 (9.4) 277 (3.5)  
Bronchitisa   <0.001 
 No 135 (60.3) 5,890 (73.7)  
 Yes 89 (39.7) 2,103 (26.3)  
Marital statusa   0.017b 
 Single 17 (7.6) 725 (9.1)  
 Married 133 (59.4) 5,187 (65.0)  
 Living together 3 (1.3) 202 (2.5)  
 Widowed 44 (19.6) 958 (12.0)  
 Divorced/separated 27 (12.1) 897 (11.2)  
 Other 0 (0.0) 15 (0.2)  
Occupational exposure to asbestosa   0.99 
 No 213 (99.5) 7,877 (99.1)  
 Yes 1 (0.5) 68 (0.9)  
Prior diagnosis of malignant tumora   <0.001 
 No 154 (65.0) 7,193 (89.9)  
 Yes 70 (29.5) 805 (10.1)  
Education   <0.001 
 High school and below 201 (93.9) 6,728 (84.3)  
 Greater than high school 13 (6.1) 1,256 (15.7)  
Family history of lung cancer   0.055 
 No 179 (75.5) 6,940 (81.4)  
 Early onset (<60 years) 20 (8.4) 484 (5.7)  
 Late onset (≥60 years) 38 (16.0) 1,099 (12.9)  
CharacteristicsIncident lung cancer (n = 237)No. of incident lung cancer (n = 8,523)P
Mean age (SD) 66.0 (7.2) 61.5 (8.5) <0.001 
Gender   0.002 
 Male 137 (57.8) 4,052 (47.5)  
 Female 100 (42.2) 4,471 (52.5)  
Mean smoking duration (SD) 39.9 (14.3) 18.9 (19.0) <0.001 
Prior diagnosis of pneumoniaa   0.066 
 No 186 (83.0) 6,971 (87.2)  
 Yes 38 (17.0) 1,022 (12.8)  
Prior diagnosis of asthmaa   0.011 
 No 173 (77.2) 6,687 (83.6)  
 Yes 51 (22.8) 1,310 (16.4)  
Prior diagnosis of tuberculosisa   0.674 
 No 216 (96.4) 7,745 (96.9)  
 Yes 8 (3.6) 246 (3.1)  
Prior diagnosis of COPDa   <0.001 
 No 90 (47.6) 5,196 (82.6)  
 Yes 99 (52.4) 1,094 (17.4)  
Emphysemaa   <0.001 
 No 203 (90.6) 7,718 (96.5)  
 Yes 21 (9.4) 277 (3.5)  
Bronchitisa   <0.001 
 No 135 (60.3) 5,890 (73.7)  
 Yes 89 (39.7) 2,103 (26.3)  
Marital statusa   0.017b 
 Single 17 (7.6) 725 (9.1)  
 Married 133 (59.4) 5,187 (65.0)  
 Living together 3 (1.3) 202 (2.5)  
 Widowed 44 (19.6) 958 (12.0)  
 Divorced/separated 27 (12.1) 897 (11.2)  
 Other 0 (0.0) 15 (0.2)  
Occupational exposure to asbestosa   0.99 
 No 213 (99.5) 7,877 (99.1)  
 Yes 1 (0.5) 68 (0.9)  
Prior diagnosis of malignant tumora   <0.001 
 No 154 (65.0) 7,193 (89.9)  
 Yes 70 (29.5) 805 (10.1)  
Education   <0.001 
 High school and below 201 (93.9) 6,728 (84.3)  
 Greater than high school 13 (6.1) 1,256 (15.7)  
Family history of lung cancer   0.055 
 No 179 (75.5) 6,940 (81.4)  
 Early onset (<60 years) 20 (8.4) 484 (5.7)  
 Late onset (≥60 years) 38 (16.0) 1,099 (12.9)  

aNumbers do not add up to total due to missing data.

bP vales were derived from the Fisher exact test.

In univariate analysis, age, male gender, smoking duration, prior diagnosis of pneumonia, asthma, COPD, emphysema, bronchitis, prior diagnosis of malignant tumor, and family history of lung cancer were significantly associated with the risk of developing lung cancer. In addition, we also performed analyses that adjusted all potential confounders in Table 1 for smoking duration. There was a significant increase in risk with age both before (HR, 1.08; 95% CI, 1.06–1.09) and after adjusting for smoking duration (HR, 1.05; 95% CI, 1.03–1.07). Male gender was significantly associated with the incidence of lung cancer both before (HR, 1.55; 95% CI, 1.20–2.01) and after (HR, 1.51; 95% CI, 1.16–1.97) adjusting for smoking. Participants with a prior history of asthma had a significant increase in risk before adjustment (HR, 1.53; 95% CI, 1.12–2.09) but not after (HR, 1.25; 95% CI, 0.91–1.71). Participants with a prior history of COPD had a significant increase in risk both before (HR, 5.13; 95% CI, 3.85–6.84) and after adjustment (HR, 2.75; 95% CI, 2.03–3.72). Prior diagnosis of malignant tumor increased risk before (HR, 4.18; 95% CI, 3.15–5.55) and after adjustment (HR, 3.09; 95% CI, 2.33–4.11). Family history of lung cancer with early onset (<60 years) increased risk with marginal significantly before (HR, 1.56; 95% CI, 0.98–2.48) but not after adjustment (HR, 1.38; 95% CI, 0.87–2.20) for smoking. Education, a measure of socioeconomic status, was protective against lung cancer before (HR, 0.58; 95% CI, 0.43–0.79) for low education (high school and below) and (HR, 0.28; 95% CI, 0.16–0.49) for higher education (greater than high school), while only higher education was protective against lung cancer after adjusting for smoking (HR, 0.53; 95% CI, 0.30–0.95). No significant effect was observed on marital status, occupational exposure to asbestos, pneumonia, bronchitis, emphysema, and tuberculosis on lung cancer risk after adjustment for smoking duration. Ever smokers (HR, 24.1; 95% CI, 10.74–54.3) were at higher risk than never smokers. Fitting smoking duration as a continuous covariate and in 10- and 20-year intervals revealed a steady increase in lung cancer risk. There was also a steady increase in risk with increasing smoking pack-years and the average amount smoked, although in neither case was it as large as that with smoking duration. A significant dose–response effect was also observed for the number of cigarettes (P < 0.001), smoking duration (P < 0.001), and smoking pack-years (P < 0.001).

Table 2 presents the final multivariate Cox regression model. Age (HR, 1.04; 95% CI, 1.02–1.06), male gender (HR, 1.48; 95% CI, 1.10–1.98), smoking duration (HR, 1.04; 95% CI, 1.03–1.05), COPD (HR, 2.43; 95% CI, 1.79–3.30), prior diagnosis of malignant tumor (HR, 2.84; 95% CI, 2.07–3.89), and family history of early onset of lung cancer (HR, 1.68; 95% CI, 1.04–2.72) significantly increased the risk of developing lung cancer. The LLPi model for incident lung cancer had good discrimination C-statistic 0.852 (95% CI, 0.832–0.873; Fig. 1) and 0.849 (95% CI, 0.829–0.870) by internal validation with bootstrap resampling and correction for optimism. The Grønnesby–Borgan goodness-of-fit test demonstrated overall good calibration χ2 7.58, P = 0.371. A standard test of the proportional hazard (PH) assumption for the model for each covariate using scaled Schoenfeld residuals showed that the PH was not violated.

Figure 1.

Performance of the LLPi risk model: C-statistic.

Figure 1.

Performance of the LLPi risk model: C-statistic.

Close modal
Table 2.

Regression coefficients, HR (95% CI), and SE for covariates in the final model

Covariatesβ-Coefficient (SE)HR (95% CI)P*
Age, y 0.036 1.04 (1.02–1.06) <0.001 
Gender (men vs. women) 0.391 1.48 (1.10–1.98) 0.009 
Smoking duration 0.043 1.04 (1.03–1.05) <0.001 
COPD 0.890 2.43 (1.79–3.30) <0.001 
Prior diagnosis of malignant tumor 1.044 2.84 (2.08–3.89) <0.001 
Family history of lung cancer 
 None Reference Reference  
 Early onset (<60 years) 0.521 1.68 (1.04–2.72) 0.034 
 Late onset (≥60 years) 0.071 1.05 (0.72–1.59) 0.722 
Covariatesβ-Coefficient (SE)HR (95% CI)P*
Age, y 0.036 1.04 (1.02–1.06) <0.001 
Gender (men vs. women) 0.391 1.48 (1.10–1.98) 0.009 
Smoking duration 0.043 1.04 (1.03–1.05) <0.001 
COPD 0.890 2.43 (1.79–3.30) <0.001 
Prior diagnosis of malignant tumor 1.044 2.84 (2.08–3.89) <0.001 
Family history of lung cancer 
 None Reference Reference  
 Early onset (<60 years) 0.521 1.68 (1.04–2.72) 0.034 
 Late onset (≥60 years) 0.071 1.05 (0.72–1.59) 0.722 

Based on the Cox model, the probability of developing lung cancer at an average of 8.7 years of follow-up is defined by the formula:

|$\hat P = 1 - S_0 \left( t \right)^{\exp \left( {\sum\nolimits_{i = 1}^p {\beta _i X_i - \sum\nolimits_{i = 1}^p {\beta _i \bar X_i } } } \right)}$|

where |$S_0 \left( t \right)$| is the baseline survival at mean value 8.7 years, |$\beta _i$| is the estimated regression coefficient, |$X_i$| is the value of the covariate, |$\bar X_i$| is the corresponding mean for continuous covariates or proportion for categorical covariates, and P is the number of covariates.

Using the above equation, the probability of developing lung cancer during the mean follow-up period of 8.7 years was calculated. This can be illustrated by considering a man diagnosed with lung cancer at the age of 50, with 30 years smoking history and a known history of COPD and no other risk factors. The estimated risk based on the LLPi model is

|$\sum\nolimits_{i = 1}^p {\beta _i X_i }$| = 0.036(50) + 0.391(1) + 0.043(30) + 0.890(1) + 1.044(0) + 0.521(0) + 0.071(0) = 4.371.

|$\sum\nolimits_{i = 1}^p {\beta _i \bar X_i }$| = 0.036(61.65) + 0.391(0.478) + 0.043(19.42) + 0.890(0.185) + 1.044(0.106) + 0.521(0.058) + 0.071(0.129) = 3.556.

|$\hat P = 1 - S_0 \left( t \right)^{\exp \left( {\sum\nolimits_{i = 1}^p {\beta _i X_i - \sum\nolimits_{i = 1}^p {\beta _i \bar X_i } } } \right)}$|

= 1− 0.9728386exp(4.371–3.556) = 0.060 = 6.0%.

*Mutually adjusted baseline survival probability at 8.7 years |$S_0 \left( t \right)$| = 0.9728386.

Using the equation under Table 2, the absolute risk of lung cancer within a mean follow-up period of 8.7 years was calculated. The LLPi risk model was used to estimate the absolute risk of developing lung cancer for three hypothetical individuals with diverse risk profile. First, the absolute risk of a woman ages 65 with 37 years smoking history and a history of COPD diagnosis with a late-onset family history of lung cancer (ages >60 at diagnosis) and no other risk factors is calculated as |$\hat P = 1 - S_0 \left( t \right)^{\exp \left( {\sum\nolimits_{i = 1}^p {\beta _i X_i - \sum\nolimits_{i = 1}^p {\beta _i \bar X_i } } } \right)}$| = 9.9% (see details of the formula under Table 2). Second, the absolute risk of a man ages 67 without a smoking history but with an early-onset family history of lung cancer (ages <60 at diagnosis) and a prior diagnosis of cancer is 6%. Third, the absolute risk of a man ages 73 who has smoked for 59 years and also had an early-onset (ages <60 at diagnosis) family history of lung cancer is 29%.

In this study, we developed and internally validated the LLPi risk model for incident lung cancer in the LLP population cohort. Age, male gender, smoking duration, prior history of COPD, prior diagnosis of malignant tumor, and family history of early onset of lung cancer (< 60 years) were significant risk factors. The C-statistic of the LLPi risk model and the bias-corrected bootstrap resampling were very similar: 0.852 (95% CI, 0.832–0.873) and 0.849 (95% CI, 0.829–0.870), respectively. The average difference known as the degree of optimism was 0.003, which indicates that the LLPi can discriminate well between patients with lung cancer and population controls. A model such as the LLPi with a high discrimination and good calibration is expedient for counselling and population-based screening programs.

To date, several risk prediction models have been developed in population-based cohort data with variable discrimination presented as ROC AUC or C-statistic (Table 3). Bach and colleagues used prospective cohort data for smokers in the Carotene and Retinol Efficacy Trial (CARET) to develop their model that included age, gender, smoking history, and exposure to asbestos as predictors (31). The model was externally validated in the Alpha-Tocopherol, Beta-Carotene Cancer Prevention (ATBC) study control arm with a C-statistic of 0.69 (32, 33). The Prostate, Lung, Colorectal and Ovarian (PLCO) cancer screening used prospective data to build two risk prediction models for the general population (model 1) and a subcohort of ever-smokers (model 2; ref. 34). The bootstrap optimism-corrected ROC AUC were 0.857 and 0.805, respectively. External validation of the models was performed in the PLCO intervention arm with AUC ROC of 0.841 and 0.784 for models 1 and 2, respectively (34). Park and colleagues developed a risk model in a large population-based cohort study of Korean men with a C-statistic of 0.864 (95% CI, 0.860–0.868; ref. 35). External validation of their model produced a C-statistic of 0.871 (95% CI, 0.867–0.876; ref. 35). Hoggart and colleagues used prospective data from the European Prospective Investigation into Cancer (EPIC) to build a predictive model for lung cancer (36). Using smoking information alone, their model had an AUC ROC of 0.843 (95% CI, 0.810–0.875). External validation of the Bach model using the same data gave an AUC ROC of 0.775 (95% CI, 0.737–0.813; ref. 36). In another study, Tammemagi and colleagues modified the PLCO models 1 and 2 to ensure applicability to NLST (37). The newly developed model PLCOM2012 utilized data from PLCO control groups of individuals who had ever smoked and validated the model in the PLCO intervention group of smokers, NLST participants, and in the PLCO intervention group stratified according to whether or not they met the NLST criteria. The AUC of their model was 0.803 in the developmental dataset and 0.797 in the validation dataset (37).

Table 3.

Risk prediction model developed in the prospective population-based cohort

Study characteristicsBach (31)PLCO (34)PLCOM2012 (37)EPIC (36)Park (35)
Population for modeling 14,254 heavy current or former smokers and 3,918 asbestos-exposed current or former smokers. Ages 45 to 69 at baseline. Population origin: United States. 70,962 cancer-free population-based individuals. Ages 55 to 74 at recruitment. Population origin: United States. 36,286 ever smoked population. Ages 55 to 74 at recruitment. Population origin: United States 169,035 former and current smokers from the general population (a random sample of 90% of each of the cases and controls as the training set, and the remaining 10% as the validation set). Population origin: Europe. 1,309,144 men who underwent health examination. Ages 30 to 80 at baseline. Population origin: South Korea. 
Risk factors included in model Age, sex, asbestos exposure, and smoking duration, duration of abstinence (if applicable), and average amount per day (while smoking). Age, socioeconomic status (education), BMI, family history of lung cancer, COPD, recent chest X-ray and smoking status, pack-year, duration and time since quitting smoking (if applicable). Age, race/ethnic group, education, BMI, COPD, personal history of cancer, family history of lung cancer, and smoking-related factors (status, intensity, duration, and quit time). Age, smoking intensity (measured by average number of cigarettes smoked per day), age started smoking, and smoking duration. Age, smoking exposure (status and intensity), age at smoking initiation, BMI, physical activity, and fasting glucose level. 
Discriminatory power in modeling population 0.72 0.86 (for all subjects); 0.81 (for ever-smokers) 0.80 0.84 0.86 
Discriminatory power in populations for external validation 0.69 (32); 0.66 (33) 0.84 (for all subjects); 0.78 (for ever-smokers only) 0.80 0.78 0.87 
Strength First lung cancer risk model. Use of spline function in modeling; high discriminatory power. Major classic risk factor included; high discriminatory power. Large population; high discriminatory power. Very large population; high discriminatory power. 
Weakness Moderate discriminatory power; smoker population only. Healthy volunteer effect may limit external generalization Smoker population only. Based on only age and smoking-related factors. Men only; missing some other classic risk factors. 
Study characteristicsBach (31)PLCO (34)PLCOM2012 (37)EPIC (36)Park (35)
Population for modeling 14,254 heavy current or former smokers and 3,918 asbestos-exposed current or former smokers. Ages 45 to 69 at baseline. Population origin: United States. 70,962 cancer-free population-based individuals. Ages 55 to 74 at recruitment. Population origin: United States. 36,286 ever smoked population. Ages 55 to 74 at recruitment. Population origin: United States 169,035 former and current smokers from the general population (a random sample of 90% of each of the cases and controls as the training set, and the remaining 10% as the validation set). Population origin: Europe. 1,309,144 men who underwent health examination. Ages 30 to 80 at baseline. Population origin: South Korea. 
Risk factors included in model Age, sex, asbestos exposure, and smoking duration, duration of abstinence (if applicable), and average amount per day (while smoking). Age, socioeconomic status (education), BMI, family history of lung cancer, COPD, recent chest X-ray and smoking status, pack-year, duration and time since quitting smoking (if applicable). Age, race/ethnic group, education, BMI, COPD, personal history of cancer, family history of lung cancer, and smoking-related factors (status, intensity, duration, and quit time). Age, smoking intensity (measured by average number of cigarettes smoked per day), age started smoking, and smoking duration. Age, smoking exposure (status and intensity), age at smoking initiation, BMI, physical activity, and fasting glucose level. 
Discriminatory power in modeling population 0.72 0.86 (for all subjects); 0.81 (for ever-smokers) 0.80 0.84 0.86 
Discriminatory power in populations for external validation 0.69 (32); 0.66 (33) 0.84 (for all subjects); 0.78 (for ever-smokers only) 0.80 0.78 0.87 
Strength First lung cancer risk model. Use of spline function in modeling; high discriminatory power. Major classic risk factor included; high discriminatory power. Large population; high discriminatory power. Very large population; high discriminatory power. 
Weakness Moderate discriminatory power; smoker population only. Healthy volunteer effect may limit external generalization Smoker population only. Based on only age and smoking-related factors. Men only; missing some other classic risk factors. 

Abbreviation: BMI, body mass index.

Although the LLPi had a high C-statistic with magnitude comparable to the aforementioned models, its interpretation is not directly comparable with these models because the differences in distributions of covariates can affect performance statistics (38). However, the LLPi model will be more directly applicable for use in primary care setting because it included readily available, strong, and plausible covariates that have been implicated in the etiology of lung cancer from our own and numerous other case–control and cohort studies.

Previous lung diseases, including COPD (emphysema, bronchitis), tuberculosis, pneumonia (39), and asthma (40), have been reported as risk factors for lung cancer. In our study, we also examined the association among COPD, emphysema, bronchitis, pneumonia, tuberculosis, and asthma on lung cancer. COPD was the only significant covariate in our final multivariable model. The association between COPD and lung cancer we found in our study was consistent with the result from two earlier meta-analyses (17, 39). In contrast, emphysema, bronchitis, pneumonia, asthma, and tuberculosis were not significantly associated with lung cancer in our final multivariate model. A plausible explanation for this observation is the small numbers of participants with these respiratory conditions in our study compared with the pooled analysis of 16, 17, and 39 studies for asthma, and other respiratory diseases in the aforementioned meta-analysis, respectively. In addition, occupational exposure to asbestos was a risk factor for lung cancer in the LLP risk model (8). However, in this present study, occupational exposure to asbestos was not a risk factor for lung cancer. This observation may be attributed to the fact that only 1 of the 237 individuals who developed lung cancer was exposed to occupational asbestos. Twenty-three participants had missing data on occupational exposure to asbestos, and 213 participants who also developed lung cancer were not exposed to asbestos (Table 1). Thus, occupational exposure to asbestos cannot be included in the model because our study population does not have adequate sample of such individuals. Another plausible explanation for this observation is reporting bias and/or misclassification as there is no means of validating the asbestos exposure other than as reported by participants. This is not the case with the medical conditions that may be validated using HES.

Strengths of our study include its prospective design, the large number of participants, long follow-up period, the population-based setting, and that detailed information on the main risk factors (such as smoking and family history of lung cancer) was ascertained by closely supervised trained interviewers, using standardized questionnaires. In addition, information bias was prevented by complete documentation of lung cancer incidence (by the ONS, the North West Cancer Intelligence Service, and hospital case-note review) and comorbidity data that were corroborated using the HES database. Furthermore, unlike in case–control studies, we were able to easily compute the absolute risk estimates for each combination of risk factors from Cox regression.

Although the LLPi model demonstrated good discrimination and calibration, more work is needed to test the applicability of the model in diverse populations, including those from diverse geographic regions. Because of marked geographical differences in incidence rates, evaluation of the LLPi risk model in high- and low-risk areas is necessary. Advancement in high-throughput methodologies and their application in molecular and genetic epidemiological studies have expanded the potential for “omic”-based risk prediction (41). Genome-wide association studies have identified inherited susceptibility patterns for lung cancer at different loci (42–44), and several methylation (45–47) and microRNA biomarkers (48–51) associated with lung cancer have been identified. It is anticipated that many more DNA polymorphism and biomarkers will be developed. Due to fewer opportunities to access biomaterials from large prospective cohort studies, most current biomarkers are used mainly for diagnosis, but their value in risk prediction has not been widely explored (41). We therefore recommend future studies to explore the contribution of genetic polymorphism, methylation, and microRNA in combination with clinical and epidemiological factors on lung cancer risk.

In conclusion, we have developed and internally validated the LLPi risk model based on readily available, strong, and plausible covariates that have been implicated in the etiology of lung cancer from our own and numerous other case–control and cohort studies. The application of the LLPi risk model in identifying individuals at high risk of developing lung cancer in population-based screening programs needs further study.

No potential conflicts of interest were disclosed.

Conception and design: M.W. Marcus, Y. Chen, O.Y. Raji, J.K. Field

Development of methodology: M.W. Marcus, Y. Chen, O.Y. Raji, S.W. Duffy, J.K. Field

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

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M.W. Marcus, O.Y. Raji

Writing, review, and/or revision of the manuscript: M.W. Marcus, O.Y. Raji, S.W. Duffy, J.K. Field

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M.W. Marcus, J.K. Field

Study supervision: M.W. Marcus, J.K. Field

The Liverpool Lung project was principally funded by the Roy Castle Lung Cancer Foundation, UK. M.W. Marcus is funded by the European Community's Seventh Framework Programme (FP7/2007-2013) under grant agreement no. HEALTH-F2-2010-258677 (CURELUNG project) and grant agreement no. 258868 (LCAOS project).

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.

1.
Dela Cruz
CS
,
Tanoue
LT
,
Matthay
RA
. 
Lung cancer: epidemiology, etiology, and prevention
.
Clin Chest Med
2011
;
32
:
605
44
.
2.
Qi
WX
,
Wang
Q
,
Jiang
YL
,
Sun
YJ
,
Tang
LN
,
He
AN
, et al
Overall survival benefits for combining targeted therapy as second-line treatment for advanced non-small-cell-lung cancer: a meta-analysis of published data
.
PLoS One
2013
;
8
:
e55637
.
3.
National Lung Screening Trial Research T
Aberle
DR
,
Adams
AM
,
Berg
CD
,
Black
WC
,
Clapp
JD
, et al
Reduced lung-cancer mortality with low-dose computed tomographic screening
.
N Engl J Med
2011
;
365
:
395
409
.
4.
Cassidy
A
,
Duffy
SW
,
Myles
JP
,
Liloglou
T
,
Field
JK
. 
Lung cancer risk prediction: a tool for early detection
.
Int J Cancer
2007
;
120
:
1
6
.
5.
Field
JK
. 
Lung cancer risk models come of age
.
Cancer Prev Res
2008
;
1
:
226
8
.
6.
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
.
7.
Etzel
CJ
,
Kachroo
S
,
Liu
M
,
D'Amelio
A
,
Dong
Q
,
Cote
ML
, et al
Development and validation of a lung cancer risk prediction model for African-Americans
.
Cancer Prev Res
2008
;
1
:
255
65
.
8.
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
.
9.
Spitz
MR
,
Etzel
CJ
,
Dong
Q
,
Amos
CI
,
Wei
Q
,
Wu
X
, et al
An expanded risk prediction model for lung cancer
.
Cancer Prev Res
2008
;
1
:
250
4
.
10.
Miettinen
OS
. 
Matching and design efficiency in retrospective studies
.
Am J Epidemiol
1970
;
91
:
111
8
.
11.
Breslow
NE
,
Day
NE
. 
Statistical methods in cancer research. Volume I—The analysis of case–control studies
.
IARC Sci Publ
1980
:
5
338
.
12.
Schulz
KF
,
Grimes
DA
. 
Case–control studies: research in reverse
.
Lancet
2002
;
359
:
431
4
.
13.
Vandenbroucke
JP
,
Pearce
N
. 
Case–control studies: basic concepts
.
Int J Epidemiol
2012
;
41
:
1480
9
.
14.
Grimes
DA
,
Schulz
KF
. 
Cohort studies: marching towards outcomes
.
Lancet
2002
;
359
:
341
5
.
15.
Mann
CJ
. 
Observational research methods. Research design II: cohort, cross sectional, and case-control studies
.
Emerg Med J
2003
;
20
:
54
60
.
16.
Raji
OY
,
Duffy
SW
,
Agbaje
OF
,
Baker
SG
,
Christiani
DC
,
Cassidy
A
, et al
Predictive accuracy of the Liverpool Lung Project risk model for stratifying patients for computed tomography screening for lung cancer: a case–control and cohort validation study
.
Ann Intern Med
2012
;
157
:
242
50
.
17.
Brenner
DR
,
McLaughlin
JR
,
Hung
RJ
. 
Previous lung diseases and lung cancer risk: a systematic review and meta-analysis
.
PloS one
2011
;
6
:
e17479
.
18.
Field
JK
,
Smith
DL
,
Duffy
S
,
Cassidy
A
. 
The Liverpool Lung Project research protocol
.
Int J oncol
2005
;
27
:
1633
45
.
19.
Clarke
LC
,
Fraser
SG
. 
Hospital Episode Statistics and trends in ophthalmic surgery 1998–2004
.
BMC Ophthalmol
2006
;
4
:
37
.
20.
Grambsch
PM
,
Therneau
TM
. 
Proportional hazards tests and diagnostics based on weighted residuals
.
Biometrika
1994
;
81
:
515
526
.
21.
Therneau
TM
,
Grambsch
PM
. 
Martingale–based residuals for survival models
.
Biometrika
1990
;
77
:
147
60
.
22.
Sun
GW
,
Shook
TL
,
Kay
GL
. 
Inappropriate use of bivariable analysis to screen risk factors for use in multivariable analysis
.
J Clin Epidemiol
1996
;
49
:
907
16
.
23.
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
.
24.
Harrell
FE
,
Califf
RM
,
Prvor
DB
,
Lee
KL
,
Rosati
RA
. 
Evaluating the yield of medical tests
.
JAMA
1982
;
247
;
2543
6
.
25.
Hanley
JA
,
McNeil
BJ
. 
The meaning and use of the area under a receiver operating characteristic (ROC) curve
.
Radiology
1982
;
143
:
29
36
.
26.
Efron
B
. 
Bootstrap methods: another look at the jackknife
.
Ann. Statist
1979
;
7
:
1
26
.
27.
Steyerberg
EW
,
Harrell
FE
 Jr
,
Borsboom
GJ
,
Eijkemans
MJ
,
Vergouwe
Y
,
Habbema
JD
. 
Internal validation of predictive models: efficiency of some procedures for logistic regression analysis
.
J Clin Epidemiol
2001
;
54
:
774
81
.
28.
Harrell
FE
 Jr
,
Lee
KL
,
Mark
DB
. 
Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors
.
Stat Med
1996
;
15
:
361
87
.
29.
Loughin
TM
,
Koehler
KJ
. 
Bootstrapping regression parameters in multivariate survival analysis
.
Lifetime Data Anal
1997
;
3
:
157
77
.
30.
Schumacher
M
,
Hollander
N
,
Sauerbrei
W
. 
Resampling and cross-validation techniques: a tool to reduce bias caused by model building
?
Stat Med
1997
;
16
:
2813
27
.
31.
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
.
32.
Cronin
KA
,
Gail
MH
,
Zou
Z
,
Bach
PB
,
Virtamo
J
,
Albanes
D
. 
Validation of a model of lung cancer risk prediction among smokers
.
J Natl Cancer Inst
2006
;
98
:
637
40
.
33.
D'Amello
AM
 Jr
,
Cassidy
A
,
Asomaning
K
,
Raji
OY
,
Duffy
SW
,
Field
JK
, et al
Comparison of discriminatory power and accuracy of three lung cancer risk models
.
Br J Cancer
2010
;
103
:
423
9
.
34.
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
.
35.
Park
S
,
Nam
B-H
,
Yang
H-R
,
Lee
JA
,
Lim
H
,
Han
TJ
, et al
Individualized risk prediction model for lung cancer in Koran men
.
Plos One
2013
;
8
:
e54823
.
36.
Hoggart
C
,
Brennan
P
,
Tjonneland
A
,
Vogel
U
,
Overvad
K
,
Ostergaard
JN
, et al
A risk model for lung cancer incidence
.
Cancer Prev Res
2012
;
5
:
834
46
.
37.
Tammemagi
MC
,
Katki
HA
,
Hocking
WG
,
Church
TR
,
Caporaso
N
,
Kvale
PA
, et al
Selection criteria for lung-cancer screening
.
N Engl J Med
2013
;
368
:
728
36
.
38.
Whittemore
AS
. 
Evaluating health risk models
.
Stat Med
2010
;
29
:
2438
52
.
39.
Brenner
DR
,
Boffetta
P
,
Duell
EJ
,
Bickeboller
H
,
Rosenberger
A
,
McCormack
V
, et al
Previous lung diseases and lung cancer risk: a pooled analysis from the International Lung Cancer Consortium
.
Am J Epidemiol
2012
;
176
:
573
85
.
40.
Rosenberger
A
,
Bickeboller
H
,
McCormack
V
,
Brenner
DR
,
Duell
EJ
,
Tjonneland
A
, et al
Asthma and lung cancer risk: a systematic investigation by the International Lung Cancer Consortium
.
Carcinogenesis
2012
;
33
:
587
97
.
41.
Field
JK
,
Chen
Y
,
Marcus
MW
,
McRonald
FE
,
Raji
OY
,
Duffy
SW
. 
The contribution of risk prediction models to early detection of lung cancer
.
J Surg Oncol
2013
;
108
:
304
11
.
42.
Hung
RJ
,
McKay
JD
,
Gaborieau
V
,
Boffetta
P
,
Hashibe
M
,
Zaridze
D
, et al
A susceptibility locus for lung cancer maps to nicotinic acetylcholine receptor subunit genes on 15q25
.
Nature
2008
;
452
:
633
7
.
43.
Amos
CI
,
Wu
X
,
Broderick
P
,
Gorlov
IP
,
Gu
J
,
Eisen
T
, et al
Genome-wide association scan of tag SNPs identifies a susceptibility locus for lung cancer at 15q25.1
.
Nat Genet
2008
;
40
:
616
22
.
44.
Thorgeirsson
TE
,
Geller
F
,
Sulem
P
,
Rafnar
T
,
Wiste
A
,
Magnusson
KP
, et al
A variant associated with nicotine dependence, lung cancer and peripheral arterial disease
.
Nature
2008
;
452
:
638
42
.
45.
Schmidt
B
,
Liebenberg
V
,
Dietrich
D
,
Schlegel
T
,
Kneip
C
,
Seegebarth
A
, et al
SHOX2 DNA methylation is a biomarker for the diagnosis of lung cancer based on bronchial aspirates
.
BMC cancer
2010
;
10
:
600
.
46.
Nikolaidis
G
,
Raji
OY
,
Markopoulou
S
,
Gosney
JR
,
Bryan
J
,
Warburton
C
, et al
DNA methylation biomarkers offer improved diagnostic efficiency in lung cancer
.
Cancer Res
2012
;
72
:
5692
701
.
47.
Sandoval
J
,
Mendez-Gonzalez
J
,
Nadal
E
,
Chen
G
,
Carmona
FJ
,
Sayols
S
, et al
A prognostic DNA methylation signature for stage I non-small-cell lung cancer
.
J Clin Oncol
2013
;
31
:
4140
7
.
48.
Lin
PY
,
Yu
SL
,
Yang
PC
. 
MicroRNA in lung cancer
.
B J Cancer
2010
;
103
:
1144
8
.
49.
Zheng
D
,
Haddadin
S
,
Wang
Y
,
Gu
LQ
,
Perry
MC
,
Freter
CE
, et al
Plasma microRNAs as novel biomarkers for early detection of lung cancer
.
Int J Clin Exp Pathol
2011
;
4
:
575
86
.
50.
Hennessey
PT
,
Sanford
T
,
Choudhary
A
,
Mydlarz
WW
,
Brown
D
,
Adai
AT
, et al
Serum microRNA biomarkers for detection of non-small cell lung cancer
.
PloS one
2012
;
7
:
e32307
.
51.
Bediaga
NG
,
Davies
MP
,
Acha-Sagredo
A
,
Hyde
R
,
Raji
OY
,
Page
R
, et al
A microRNA-based prediction algorithm for diagnosis of non-small lung cell carcinoma in minimal biopsy material
.
B J Cancer
2013
;
109
:
2404
11
.