Factors Associated with Human Small Aggressive Non–Small Cell Lung Cancer

Oncologist and radiologist members of the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial (PLCO) Lung Subcommittee (1), who are involved in the diagnosis and treatment of lung cancer patients, had reported observing a small but consistent proportion of non–small cell lung cancers (NSCLC) that spread to distant lymph nodes or metastasize while the primary tumor was still relatively small, as opposed to those that are relatively large before they progress. Subsequent discussions with colleagues in the International Early Lung Cancer Action Program (2) indicated that they were similarly observing a small but consistent number of these small aggressive NSCLC (SA-NSCLC). A review of the literature found that, although the genetic alterations that accompany aggressive lung cancers have been studied (3-8), very little is known about who develops SA-NSCLC and factors associated with SA-NSCLC. Yet, SA-NSCLCs are an important subset of lung cancers. The prognosis for individuals with these SA-NSCLC is grave as disseminated cancers are no longer amenable to surgery with curative intent. The outlooks for individuals destined to develop SA-NSCLC might be improved if diagnoses were obtained early in the course of disease. Their small size presents difficult clinical and screening challenges, which might benefit from targeted approaches applied to high-risk populations. The natural history of SA-NSCLC, progressing before becoming large, may involve unique biological pathways and developing an understanding of these pathways may provide new insights into lung carcinogenesis. Hence, a novel exploratory study was conducted to identify factors associated with SA-NSCLC versus non–SA-NSCLC, that is, how do individuals with SA-NSCLC differ from those with cancers that do not spread until they are larger?

This study uses a case-case design nested in the National Cancer Institute's PLCO cohort. All eligible NSCLCs occurring in the PLCO as of January 31, 2005 were included. The PLCO study has been described in detail (1, 9). In brief, the PLCO is a randomized screening trial evaluating the effect of selected periodic screening interventions on prostate, lung, colorectal, and ovarian cancer mortalities. The trial began enrolling participants in 1993 and completed recruitment of 154,938 subjects in 2001. Subjects at study entry were men and women ages 55 to 74 years who were free of the cancers under study. The screening intervention for lung cancer consisted of a posteroanterior chest radiograph at baseline and then annually for 3 years. Controls received regular care as recommended by their physician. The PLCO received Institutional Review Board approval and informed consent was obtained from each study participant.

Disease status was classified according to American Joint Committee on Cancer tumor-node-metastasis staging (4th edition; ref. 10). The study focus was to identify factors associated with biological progression to aggressive spread while the primary tumor was still relatively small. SA-NSCLC cases were lung cancer patients with T₁ tumors (<3 cm) that had spread to distant regional lymph nodes (N₃), had metastasized (M₁), or both. The non–SA-NSCLC cases, for the purposes of this report referred to as “controls,” included individuals who had tumors that had grown larger than T₁ and excluded individuals with locally aggressive disease (T₄), distant nodal disease (N₃), or metastasis (M₁); the primary control group includes T₂ to T₃, N₀ to N₂, and M₀ (Table 1 cells A + B). Alternative control groups were evaluated to assess the robustness of study associations. They differ from the primary controls by inclusion/exclusion of T₁ and N₁ to N₂ disease. Control 2 include T₂ to T₃, N₀, and M₀ (Table 1 cell A); control 3 include T₁ to T₃, N₀, and M₀ (Table 1 cells A + C); and control 4 include T₁ to T₃, N₀ to N₂, and M₀ (Table 1 cells A + B + C + D). The primary control group was selected a priori and without knowledge of the results of analyses using alternative control groups.

Predictor variables evaluated for associations with SA-NSCLC are listed in Table 2. T staging used computed tomography for 91.7% of cases and 93.0% of controls, with the remainder being staged by chest radiography or alternative procedures (P = 0.76). Predictor variable data came from an epidemiologic questionnaire that was administered to participants at study baseline. Regular ibuprofen use was estimated from data from the baseline questionnaire, which asked: “During the last 12 months, have you regularly used ibuprofen-containing products, such as Advil, Nuprin, or Motrin?” Average frequency of use during this period was measured in number of pills of ibuprofen-containing products taken per seven times ranging from less than two per month to more than or equal to two per day. Female hormone use refers to the ever use of estrogen or progesterone (tablets, pills, or creams) for menopause.

Distributions of variables were evaluated by contingency table analyses and differences in distributions were evaluated by Fisher's exact test, χ² test for multilevel variables, and t test for continuous variables. Univariate and multivariate associations were evaluated by logistic regression (11). Variables approaching significance (P < 0.15) in univariate analysis were evaluated in multivariable models. Smoking variables, family history, and histology were evaluated in multivariable models regardless of P values based on a priori reasoning (12). Histology was considered a potential confounder and was forced into logistic regression models. Backward selection was used to develop parsimonious models. All possible two-way interactions for main effects variables in multivariable models were evaluated and the likelihood ratio test (LRT) was used to estimate the significance of interaction terms. Because the number of SA-NSCLC was small, predictor variable inclusion in final multivariable models was restricted to those for which data were present for at least 95% of SA-NSCLC cases and missing data did not differ significantly by SA-NSCLC/non–SA-NSCLC status. Regression diagnostics, model fit, and interaction between covariates in the final model were evaluated. α-Error was set at 0.05 and all reported P values are two sided. Stata 9.2 software (Stata Corp.) was used to compute statistics.

The distributions of SA-NSCLC/non–SA-NSCLC by tumor-node-metastasis staging categories for cases and primary and alternative controls are presented in Table 1. The distributions and associations of selected study variables by case-control status are presented in Table 3.

In univariate analysis, SA-NSCLC was associated with younger age at diagnosis, female gender, and adenocarcinoma versus other NSCLC and was inversely associated with duration smoked squared, history of heart disease/infarction, and regular ibuprofen use (Table 3).

In multivariate analysis, SA-NSCLC was significantly associated with younger age, female gender, family history of lung cancer, and gender*family history of lung cancer interaction (LRT interaction, P = 0.02) and was inversely associated with ibuprofen use (Table 4, model 1). The model Hosmer-Lemeshow goodness-of-fit test P value was 0.68, indicating that the null hypothesis of goodness of fit was not rejected. When models were prepared using alternative controls (described in Table 1), the directions and magnitudes of odds ratios (OR) for independent variables were consistent with those obtained in the primary case-control analysis (Table 4; data not shown). These observations indicate that the presented findings are not the result of an idiosyncratic comparison group. Adjusted for age, ibuprofen use, and histology, the gender-SA-NSCLC (female versus male) OR in those with a family history of lung cancer was 11.76 [95% confidence interval (95% CI), 2.00-69.22; P = 0.006] and for those without a family history was 1.86 (95% CI, 0.88-3.96; P = 0.11). Although histology was associated with SA-NSCLC (Table 3) and with gender (OR_{adenocarcinoma∼female}, 1.58; 95% CI, 1.14-2.18), adjustment for histology did not explain away the strong association between SA-NSCLC and gender. Conversely, adjusted for age, ibuprofen, and histology, the SA-NSCLC-family history OR was 3.81 (95% CI, 1.37-10.59; P = 0.01) in women and 0.54 (95% CI, 0.12-2.50; P = 0.43) in men.

When considering women only, 64.3% of 28 cases and 57.0% of 107 controls had a history of female hormone use (OR, 1.36; 95% CI, 0.57-3.22; P = 0.49) and some or all ovarian tissue had been removed in 10.7% of 28 cases and 18.3% of 104 controls (OR, 0.54; 95% CI, 0.15-1.96; P = 0.22). These data are consistent with the hypothesis that female hormones might be involved in the development of SA-NSCLC but are based on limited numbers.

In multivariable analysis, ibuprofen use was an important independent protective factor (Table 4). Adjusted for covariates, compared with nonusers, the SA-NSCLC-ibuprofen OR was 0.41 (95% CI, 0.15-1.12; P = 0.08) for regular users of less than one per day and 0.34 (95% CI, 0.07-1.59; P = 0.17) for users of more than one or one pill per day. The inverse association between SA-NSCLC and ibuprofen was similar in current and former smokers but seemed to be minimal in never smokers: age- and histology-adjusted ORs were 0.31 (95% CI, 0.07-1.47; P = 0.14), 0.35 (95% CI, 0.10-1.25; P = 0.10), and 0.84 (95% CI, 0.07-10.60; P = 0.89), respectively. However, the number of never smokers was limited and the smoking status*ibuprofen interaction did not approach significance (LRT, P = 0.38). Thus, it cannot be concluded that a SA-NSCLC-ibuprofen association is absent in never smokers. The SA-NSCLC-ibuprofen association did not differ by gender (interaction LRT, P = 0.64) or histology (interaction LRT, P = 0.20).

In this study, small cancers that have spread to distal sites or have metastasized are contrasted with larger cancers that have not. The study associations might be biased if lung cancer detection was differentially moved forward or backward to an uninformative period in lung carcinogenesis, that is, to early stage (T₁ before distant spread or metastasis had a chance to occur in cancers that would have become SA-NSCLC) or advanced stage (when delayed detection allowed SA-NSCLC to have grown locally to become T₂ or larger cancers and thus no longer would be classified as SA-NSCLC). Differential diagnosis times along the disease progression pathway leading to uninformative early or late diagnosis could have biased study findings. Because it is not possible to identify the size and progression transition times and sequences, the effect of differential diagnosis time bias cannot be evaluated directly. Several steps were taken to attempt to assess the potential effect of differential diagnosis time bias. Models were adjusted for factors thought to be associated with stage at diagnosis, that is, socioeconomic status estimated by education and use/access to health care estimated by fecal occult blood test and chest radiograph(s) in the 3 years before study entry. The study models were adjusted for time from entry into the PLCO to lung cancer diagnosis, stratified by PLCO screen and control arms (primarily asymptomatic versus symptomatic diagnoses), and adjusted for time from last screening to diagnosis in the screening arm only (Table 4, models 2-5). In all of these analyses, the effect estimates for the presented variables were consistent in direction and magnitude with model 1, suggesting that study findings are not the spurious results of differential diagnosis time bias.

This study found that SA-NSCLC was inversely associated with ibuprofen use and associated with younger age and female gender, and the latter association was significantly more pronounced in those with a family history of lung cancer—a 12-fold increased risk in women compared with men. The association between SA-NSCLC and family history of lung cancer was present in women but not in men. A similar pattern of cancer risk that incorporates associations with family history and earlier age of onset has been observed in inherited genetic susceptibilities to cancers due to germ-line mutations [e.g., BRCA1 and BRCA2 (13) and p53 (14)]. The findings of the current study are different but analogous. The current study evaluated risk of having aggressive versus less aggressive cancer rather than risk of developing cancer. Nevertheless, the study findings do suggest the possibility that inherited genetic factor(s) interacting with female factor(s) might be associated with early progression of lung cancer.

In this study population, SA-NSCLC was significantly associated with adenocarcinoma and showed a trend to association with lighter smoking history, both of which were associated with female gender. It may be possible that the female-SA-NSCLC association is explained by the association between female gender and adenocarcinoma and smoking history. However, in all models evaluated that included variables for histology and smoking exposure, the gender-SA-NSCLC effect remained and seemed to be the most important of the three predictors. It is possible that misclassification of smoking history may have led to underestimation of the effect of smoking. However, detailed smoking history was collected at baseline and the mean duration between smoking data collection and lung cancer diagnosis was 2.98 years. Given that the study subjects had a mean pack-years smoking history of 45.3, it is not expected that misclassification of smoking history due to interim smoking between data collection and lung cancer diagnosis is substantial. Substrata analyses by smoking status and histology were hampered by limited sample size.

Epidemiologic and biological studies indicate that lung carcinogenesis in women differs from that in men (15). Some (16-21), but not all studies (22, 23), have found that, for given level of smoking or for nonsmoking, women are at greater risk of lung cancer (mortality or incidence). Mechanisms implicated in the gender-associated increased risk associated with smoking include increased activation of tobacco procarcinogens by CYP1A1 enzyme (24), elevated levels of polycyclic aromatic hydrocarbon-DNA adducts per level of smoke exposure (24-26), reduced detoxification of tobacco carcinogens by glutathione S-transferase M₁ (27), lower capacity to repair tobacco carcinogen-induced DNA damage (28), greater sensitivity to the genotoxic effects of the tobacco carcinogen 4-(methylnitrosamino)-I-(3-pyridyl)-1-butanone (29), and greater lung tissue expression of gastrin-releasing peptide receptor (30). Estrogen receptor, particularly estrogen receptor-β, expression occurs in lung normal and tumor tissues and estrogens and antiestrogens have been shown to stimulate and inhibit lung cancer cell proliferation, respectively (31-34). Some data suggest that early menopause reduces risk and estrogen therapy increases risk of adenocarcinoma in women (35). Some of these pathways that seem to increase risk of lung cancer development may also increase the probability of mutations involved in cancer progression and thus could explain some of the SA-NSCLC-gender association. An integrated understanding of how these aforementioned factors explain gender differences in lung carcinogenesis awaits further research, as does insight into the mechanisms explaining the SA-NSCLC-gender association.

Ibuprofen use was inversely associated with SA-NSCLC. Ibuprofen and aspirin are nonsteroidal anti-inflammatory drugs (NSAID) that are cyclooxygenase (COX) inhibitors. Elevated expression of COX-2 has been associated with many cancers and has been observed in animal (36) and human (37-40) lung tumors and precursor lesions (37, 38, 41). NSAIDs reduce COX activity, cell proliferation, and/or tumorigenicity in animal lung cancer models (42, 43) and human lung cancer cells (44). A meta-analysis of 14 epidemiologic studies investigating NSAIDs and lung cancer found a significant protective association (45). These studies relate NSAIDs to lung cancer development. In contrast, the current study describes associations with lung cancer progression. NSAIDs have been associated with reduction in tumor progression (46-48). Hida et al. (37) and Takahashi et al. (40) found that lung cancer cells in lymph node metastases had markedly greater COX-2 expression than in the corresponding primary tumor and that the metastatic capability of lung cancer cells was associated with COX-2 expression. Their conclusion that increased COX-2 expression may be associated with an invasive/metastatic phenotype is consistent with the findings of this study—that ibuprofen might be protective against early lung cancer progression.

Aspirin was found to have a weak, nonsignificant protective effect. Similarly, studies have found that ibuprofen was more protective than aspirin against developing breast and lung cancers (49). The findings of this study suggest that further evaluation of ibuprofen as a protective against cancer progression is warranted.

Control selection is an important determinant of validity in epidemiologic studies. We initially reasoned that patients with nonaggressive T₁ tumors should be excluded from the control group because they include potential SA-NSCLC cases that had not developed N₃ or M₁ because of fortuitous early diagnosis. If they had been left and watched, they may have developed N₃ or M₁ or both before growing into T₂ tumors and thus would have been classified as SA-NSCLC. Individuals with T₂ or T₃ disease have tumors that have presumably undergone more cell replications than T₁ cancers, and if they have not yet shown N₃/M₁ dissemination, they definitely do not have the SA-NSCLC natural history. Thus, a priori, the primary control group for this study consisted of individuals with T₂ or T₃ disease with no or only local nodal involvement and no metastases. It is important to note that all other cases have not “declared” their natural history, and so it is unclear to which group they belong and to what extent they are differentially excluded with respect to any given risk factor. To the extent possible, we examined the sensitivity of our analyses to the assumption that the exclusion is not differential. One way of examining this potential bias was to evaluate associations found to be important in case-control comparisons for consistency using alternative control groups. Consistency of study findings was shown when three alternative control groups described in Table 1 were evaluated (data not shown).

This study contrasts small progressed cases with larger nonprogressed cases. Those cases that happened to be diagnosed in these categories can be distinguished. However, data and methods are not available to identify what course of natural history small localized (“early”) cancers would have taken if left and watched or diagnosed at a later time and what course large progressed (“late”) cancers took in their development. Thus, differential diagnosis time may be a potential source of bias. To affect the current study, diagnosis times would have to have been differential by age, ibuprofen use, and sex-family history combinations and transition interval distribution would have to differ by natural history. Because it is not possible to evaluate differential diagnosis time directly, several surrogate variables thought to be associated with stage and diagnosis time were included in models. Adjustment for education and occurrence of fecal occult blood test or chest radiograph in the 3 years before study entry did not alter study conclusions (data not shown) nor did inclusion of time from study entry to lung cancer diagnosis, study arm, or time from screening to diagnosis (Table 4). However, none of these variables directly adjusts for differential diagnosis time, so the possibility of bias cannot be completely dismissed. Study designs and analytic methods need to be developed to advance the study of disease progression.

The study found interesting novel associations. However, some associations may be due to unknown confounders and may not be causal. Revelation of confounded or noncausal associations will require further detailed study. Forty-seven associations were potentially considered for multivariate analysis, and with type I error of 0.05, approximately two associations could have been found to be significant due to chance. The PLCO study, which serves as the source of the current study, consists of a sample that is self-selected and has a higher socioeconomic status than the general population. The age at entry into the PLCO was 55 to 74 years. Thus, external generalizability to the population at large is not assured. However, the associations reported seem to be biological and are not expected to be strongly modified by these factors. Study strengths include the case-case design nested within a very large, relatively representative, and well-defined cohort: data collection was prospective, eliminating recall bias, and sampling cases and controls arising within the same cohort minimizes selection bias that might result from alternative sampling methods.

It is likely that chest radiography is not the optimal lung cancer screening method, in particular, when compared with more sensitive imaging methods, such as low-dose spiral helical computed tomography, which is currently being evaluated in clinical trials (50). Chest radiographs are thought to be more efficient at detecting relatively large peripheral tumors, which are predominantly adenocarcinoma, compared with centrally located tumors, which are more frequently squamous cell carcinoma. It remains to be determined in larger studies using more sensitive imaging techniques whether the current study associations are histology specific and what the lower primary tumor size limits are for detection of SA-NSCLC.

This study was exploratory and it is mandatory that observed associations be validated in different populations. Given that associations are found to be causal, the findings of this study suggest additional research, in particular, identification of the gene(s), gene alterations, and female factors that are associated with SA-NSCLC. Knowledge of the gene-gender interaction and biology of SA-NSCLC might lead to a better understanding of carcinogenesis and cancer progression and identification of high-risk populations that might benefit from screening or increased clinical monitoring and may lead to effective chemoprevention and improved therapeutics.

We thank Dr. Anthony Miller for his numerous valuable comments and the stimulating discussions he provoked in the development of this report.