Computed Tomography Screening and Lung Cancer Outcomes FREE

Lung cancer accounts for 20% of cancer deaths in Italy and 25% of cancer deaths in the United States, and 6% of all deaths in both countries.^1- 2 Screening individuals at high risk for lung cancer might reduce these statistics based on the premise that most cases of lung cancer that will cause death can be detected through routine screening while they are still localized and potentially curable. However, prior randomized studies of lung cancer screening with chest x-ray have not supported this premise. Rather, chest x-ray was effective at identifying many additional small tumors in the lung that could be removed, but their discovery and removal did not reduce the likelihood that individuals would be diagnosed with new cases of advanced lung cancer, or would die of lung cancer.^3- 7 These findings led to speculation that the additional small cancers that were being found through screening may be indolent, relative to lung cancer that is typically encountered in a clinical setting.^8- 9

Today there is renewed enthusiasm for lung cancer screening with computed tomography (CT) because it is more sensitive for the detection of very small nodules.^10- 11 Yet, CT screening depends on the same unproven premise as chest x-ray screening.

We studied the effect of CT screening on individuals enrolled in 1 of 3 single-arm studies of screening by comparing the frequency of lung cancer cases, lung cancer resection, advanced lung cancer cases, and deaths from lung cancer occurring in these studies with what would have occurred in the absence of screening, as determined from a set of validated prediction models.^12- 15 We aimed to produce preliminary estimates of the impact of widespread CT screening on these lung cancer outcomes.

Individuals were enrolled in 1 of 3 studies conducted at the Istituto Tumori in Milan, Italy, the Mayo Clinic in Rochester, Minn, and the Moffitt Cancer Center in Tampa, Fla.^16- 20 Each study recruited individuals with a smoking history and no prior history or symptoms suggestive of lung cancer through advertisements, direct mail, and local physician outreach. The Istituto Tumori study is ongoing, the Mayo Clinic study offered an initial and 3 subsequent annual CT scans, and the Moffitt study offered an initial and 4 subsequent annual CT scans. In these studies, when noncalcified nodules were detected, they were evaluated using a systematic approach of follow-up imaging and biopsy.^16- 17

The frequency of lung cancer events occurring among the study participants was compared with benchmarks that were determined using 2 models developed to estimate an individual's level of risk of either being diagnosed with lung cancer or dying of lung cancer. Both models have been extensively described and validated in prior studies.^12- 15 The models were designed to apply to high-risk individuals, whose risk factor values fall within the following ranges: age 50 to 80 years, smoked for between 25 and 60 years, averaged between 10 and 60 cigarettes per day, and if quit, had quit within the past 20 years. These limits led us to exclude 58 of the 1035 individuals from the Istituto Tumori study, 81 of 1520 individuals from the Mayo Clinic study, and 321 of 1151 individuals from the Moffitt study.

The probabilities of advanced lung cancer diagnosis and lung cancer surgery were based on the estimated probability of lung cancer diagnosis from our model multiplied by the conditional probability of these events among age- and sex-matched individuals diagnosed with lung cancer in the National Cancer Institute's Surveillance, Epidemiology, and End Results cancer registry database for 1993 through 1998. Estimates for the Moffitt study were also adjusted to compensate for an enriched prevalence of obstructive lung disease (due to a requirement that participants have obstructive lung disease) in the first cohort (enrolled between December 11, 1998, and January 4, 2001; n = 376), and subsequent dilution of the prevalence of obstructive lung disease (resulting from late entry of individuals without obstructive lung disease who were not eligible for the first cohort) in the second cohort (enrolled between January 5, 2001, and August 13, 2003; n = 458). The estimated risk of each individual was multiplied by the aggregate change in the population risk resulting from the alteration in prevalence of obstructive lung disease based on the following prevalences for obstructive lung disease: 97% for cohort 1 and 25% for cohort 2. The expected prevalence in the absence of the additional entry criterion was 46%, as estimated for the 163 individuals who were newly recruited to the second cohort. The relative risk (RR) estimates for the impact of obstructive lung disease on lung cancer incidence (2.80) and lung cancer death (2.02) were taken from the literature, the latter being an average of the additional risks conferred by differing levels of obstruction, weighted by the proportion of individuals with those differing levels of obstruction in the first cohort.^21- 22 The end result of these adjustments was a small increase in the expected number of events.

We evaluated the number of individuals diagnosed with lung cancer (first diagnoses of non−small cell or small cell lung cancer), the number of first surgical resections of lung cancer, the number of individuals whose first lung cancer diagnosis was of advanced stage (stage III or IV non–small cell lung cancer or any stage of small cell lung cancer), and the number of individuals who died from lung cancer. For the Istituto Tumori study, dates and cause of death were ascertained through medical record review and the Lombardy Region cancer registry and death records, and clinician investigators coded cause of death. For the Mayo Clinic and Moffitt studies, dates and cause of death were coded according to the National Death Index assignment of cause of death to parallel other studies of lung cancer screening.⁸ Individuals whose vital status could not be reliably checked through public data sources were eliminated from the mortality analyses, as was the case for 1 individual enrolled in the Istituto Tumori study who resided outside of the Lombardy region of Italy, and 7 and 22 individuals enrolled in the Mayo Clinic and Moffitt studies, respectively, who did not provide Social Security numbers at the time of study enrollment.

The date of the first CT screen is the baseline for all of our analyses. The timing of lung cancer diagnoses corresponds to either the date of lung cancer in the diagnostic pathology report or the date of death for individuals who died from lung cancer without a known antecedent diagnosis. This approach to ascertaining the date of cancer diagnosis parallels the manner in which epidemiological and clinical data are collected, and thus is appropriate for the comparisons presented herein. An alternative would be to back-date the date of diagnosis to the date of the first scan on which the cancer was theoretically visible (eg, prevalence or incidence scan). Follow-up times for diagnostic and treatment end points for each individual extend to the point that the individual was diagnosed with lung cancer, or underwent surgery for lung cancer, or alternatively was lost to follow-up or died. Few individuals were lost to follow-up during the first years of the studies (89% follow-up through 2 years, 71% follow-up through 3 years), but the median follow-up was 3.9 years and only 33% were followed up past 4 years.

There is a potential that our results are biased if the probability that an individual is lost to follow-up is correlated with the likelihood that he/she experienced one of the study events. For instance, if individuals with symptoms of late-stage cancer are more likely to remain in clinical follow-up than individuals who feel healthy, such a bias would produce results that downplay the effect of CT screening on preventing advanced cancers, while the alternative would exaggerate the effect of CT screening on preventing advanced diagnosis. Given this potential source of bias, additional analyses are presented that reflect how many events would have been estimated had all individuals been followed up to the time of their observed event, or otherwise followed up to the maximum time possible, which was set to be the last date of observed follow-up for any individual in the study. These alternative analyses increase the number of predicted events without increasing the number of observed events, so they represent the most conservative estimate of the extent to which CT screening may increase the likelihood of lung cancer diagnosis, and lung cancer surgery, as well as the most optimistic projection of the extent to which CT may reduce the frequency of advanced cancer diagnosis.

For the Istituto Tumori study, the censoring date was the date of death from a cause other than lung cancer, or the date of last confirmation of vital status; for the Moffitt and Mayo Clinic studies, the censoring date was the earlier of either the date of death from a cause other than lung cancer or Decemer 31, 2004, the most recent date through which National Death Index data linked to death certificate information were complete. Our original analytic plan was to examine all deaths from lung cancer beginning at the commencement of screening but there was strong evidence for a healthy volunteer bias in the 3 studies. There was only 1 death from lung cancer occurring in the first year across all studies (compared with 9.5 deaths anticipated in the absence of screening). Because it is plausible that there was a large reduction in the first year due to the requirement that all study individuals be asymptomatic at the time of enrollment in the studies, and not plausible that CT screening reduced deaths immediately, the mortality end point was evaluated commencing 1 year after the first CT scan through death due to lung cancer or a censoring event. How the results would have appeared had the first year's events been considered is also reported. The date of tissue diagnosis served as the start time for the survival analyses.

The numbers of observed O events were compared with the numbers expected E at the end of available follow-up time using the formula (O-E)²/E, which produces a χ² statistic with 1 degree of freedom. The results are indistinguishable from those obtained from a more computationally intensive approach to calculate the variance of E.¹³ The 95% confidence intervals (CIs) assume a Poisson distribution. The survival estimates are based on the Kaplan-Meier method. All P values are 2-sided. P<.05 was considered statistically significant. Analyses were performed using Stata version 9.0 (StataCorp, College Station, Tex).

The study was approved by the institutional review boards of the 3 study sites where the screening was performed and written informed consent was obtained. The institutional review board at Memorial Sloan-Kettering Cancer Center acknowledged the authority of these institutions to authorize this research. Only the investigators at Memorial Sloan-Kettering Cancer Center had access to data from all 3 sites, which had been deidentified.

The Mayo Clinic study contributed both the largest number of individuals and the greatest amount of follow-up time, followed by the Istituto Tumori study (Table 1). The lung cancer mortality rate overall was 3.5 lung cancer deaths per 1000 person-years; the 5.6 lung cancer deaths per 1000 person-years in the Moffitt study is most likely due to the additional prevalence of obstructive lung disease. The lung cancer mortality rates in the Istituto Tumori and Mayo Clinic cohorts (2.7 and 3.1 per 1000 person-years, respectively) are similar to the rates in prior well-known studies of lung cancer screening with chest x-ray (3.0 and 3.2 per 1000 person-years in the 2 arms of the original Mayo Clinic study; 2.6 and 3.6 per 1000 person-years in the 2 arms of a study of chest x-ray screening performed in Czechoslovakia; and 2.7 per 1000 person-years in the Memorial Sloan-Kettering study).^7- 9,23

Individuals in all 3 studies were diagnosed with lung cancer in far greater numbers than would have occurred in the absence of screening (Figure 1A, B, C and Table 2). For the Istituto Tumori study, the observed frequency was 36 compared with an expected frequency of 11.3; Mayo Clinic, 66 vs 19.5; and Moffitt, 42 vs 13.7 (P<.001 for each). The number of lung cancer surgeries performed exceeded the number expected in the absence of screening to an even greater extent for the Istituto Tumori study (33 vs 2.7); Mayo Clinic (48 vs 4.8); and Moffitt (28 vs 3.4) (P<.001 for each; Figure 1D, E, F and Table 2). Combining data from the 3 studies, there were 144 cases of lung cancer diagnosed, whereas 44.5 cases were expected, which resulted in an RR of lung cancer diagnosis of 3.2 (95% CI, 2.7-3.8; P<.001; Figure 2A). A total of 109 lung cancer surgeries were performed compared with 10.9 cases expected, which resulted in an RR of lung cancer surgery of 10.0 (95% CI, 8.2-11.9; P<.001; Figure 2B). If follow-up time is included for those patients who were lost to follow-up for unknown reasons, the number of observed lung cancer cases is 2.3 times the number predicted (144 vs 62.2 predicted; P<.001) and the number of lung cancer surgeries performed is 7.2 times the number predicted (109 vs 15.2 predicted; P<.001).

Early detection via CT screening did not appear to reduce the risk of advanced lung cancer diagnoses (Figure 1G, H, I, Figure 2C, and Table 2). Combined, there were 42 cases of advanced lung cancer, while the model predicted 33.4 (P = .14). If we include the follow-up times of patients lost to follow-up for unknown reasons, providing an upper bound of possible benefit, the number of observed cases of advanced lung cancer is less than the number predicted by 18% (42 observed vs 51.2 predicted; P = .18).

There was no evidence that CT screening reduced the risk of death due to lung cancer in any of the studies individually or combined (Figure 1J, K, L, Figure 2D, and Table 2). Loss to follow-up is unlikely to have biased this outcome, which was assessed using a standardized follow-up time. Combined, there were 38 deaths due to lung cancer after the first year of screening and the model predicted 38.8 deaths (RR, 1.0; 95% CI, 0.7-1.3; P = .90). If we had included the first year of mortality data in this analysis (Table 2), CT screening would have appeared to reduce lung cancer mortality by 20%, although this reduction would not have been statistically significant (RR, 0.8; 95% CI, 0.55-1.06; P = .18).

Shown in Table 3 are the lung cancers detected during the course of the study, the number of individuals who died from lung cancer after 1 year or more of study participation who had each type of lung cancer diagnosis, and the 2-year probability of overall survival for those individuals diagnosed with lung cancer. As in other studies of CT screening, the preponderance of lung cancers (96 [67%] of 144) were of early stage (ie, stage I or stage II), and the outcomes for these individuals were quite good—only 12 (13%) of the individuals with early stage non–small cell lung cancer died from lung cancer during the study. Instead, the majority of individuals who died from lung cancer, despite participation in annual screening, did not have their cancer detected when it was in an early stage and likely to be curable: 13 (34%) of those who died from lung cancer were initially diagnosed with stage III or stage IV non–small cell lung cancer, 7 (18%) were diagnosed with small cell lung cancer, and 6 (16%) had no documented diagnosis of lung cancer prior to their death from lung cancer.

Reported in Table 3 are also the 2-year overall survival probabilities for each group of individuals by initial diagnosis. For instance, individuals diagnosed with stage I to stage II and stage III to stage IV non–small cell lung cancer had 2-year overall survival rates of 90% and 47%, respectively, superior to the published survival rates of 86% (stage IA) to 56% (stage IIB) in the first case, and 40% (stage IIIA) to 5% (stage IV) in the second case.²⁴ The 4-year lung cancer−specific survival of the 81 individuals in our studies diagnosed with clinical stage I lung cancer and undergoing surgery was 94% (95% CI, 85%-97%), matching outcomes reported in another recent study of CT screening.²⁵

Previous studies of chest x-ray screening for lung cancer produced 2 intriguing results. Screening increased the rate of detection of small resectable lung cancers and thus the frequency of lung surgery. However, screening did not reduce the risk of either advanced lung cancer diagnosis or death from lung cancer.⁷ In our study of CT screening, we observed a similar pattern. When individuals are screened for lung cancer with CT, the likelihood that they are diagnosed with lung cancer is increased more than 3-fold, and the likelihood that they undergo a thoracic resection for lung cancer is increased 10-fold. However, as for chest x-ray screening, there appears to be neither a meaningful reduction in the number of advanced cancers being diagnosed nor a reduction in the number of individuals who die of lung cancer. These findings, because they are thematically consistent with the findings of several randomized studies of lung cancer screening with chest x-ray, should raise doubts about the premise underpinning CT screening for lung cancer, and also raise concerns about its potential harms if pursued on a wide scale.

To generate our findings, we used a comparative approach similar to that used in evaluations of other cancer screening tests. Cytological cervical cancer screening was established through comparisons of the incidence and mortality rates between populations in which cervical cytology screening was common compared with populations in which it was uncommon.^26- 28 Colonoscopy screening for colon cancer is widely recommended based on comparisons of colon cancer rates among individuals screened with colonoscopy compared with those not screened.^29- 31 Screening children for neuroblastoma is not recommended due to the absence of a difference in advanced cases and deaths due to neuroblastoma between screened and unscreened populations.^32- 33

There are 2 differences between our analyses and these other comparative studies. First, these other studies were large and definitive, while ours is small and preliminary. Second, rates of cancer events were compared between unscreened and screened populations in these other studies, while in our analyses the comparator was generated by prediction models that used individual risk factor information to estimate individual-level risk. So the validity of our comparisons hinges on our models, which in prior analyses have been shown to accurately predict the frequency of lung cancer events occurring among individuals enrolled in other screening and prevention studies.^12- 15

A possible explanation for the association of screening and an increase in the frequency of lung cancer diagnosis more than 3-fold is that many of the early lung cancers found through screening would not have progressed rapidly to a point of clinical detection. As such, if left unattended, they would have been unlikely to account for a meaningful share of all deaths occurring from lung cancer among the screened individuals.^34- 36 An alternative perspective is that the additional cancers found at screening would have progressed if left untreated to cause clinically significant disease, and ultimately death in a substantial proportion of cases.^37- 39 If there were a reduction in lung cancer mortality observed over a longer period of follow-up, such a finding would support this latter explanation.

Our data more strongly support the former explanation because of 2 findings. (1) The number of excess cases diagnosed each year continues to exceed the number predicted. An initial rise followed by a return to the baseline incidence rate would be more consistent with a 1-time discovery of a large number of cancers that would have soon appeared sporadically had screening not been performed. This persistent elevation of lung cancer incidence suggests that there is a pool of small nodules being found by CT screening that either have reduced metastatic potential, or have a much slower growth rate and longer natural history than that of the clinically detected lung cancers on which current knowledge of this disease is based. (2) There was no decline in the number of advanced cancers being detected in the studies. Had there been, this too would have constituted evidence that the additional cancers found through screening would have soon progressed had they not been detected. Even with our most conservative estimate, 9 fewer cases of advanced lung cancer than anticipated were observed over the entire study period across all 3 studies, which is less than one tenth the number of excess cases of lung cancer detected by screening, totaling 99.

Our finding of a 10-fold increase in lung cancer surgeries resulting from screening underscores one of the potential public health consequences of CT screening. If the majority of excess early cancers found through screening are unlikely to progress rapidly to a point where they cause clinically significant disease or death, then the thoracic surgeries performed to remove them may be insufficiently beneficial to justify the resulting morbidities. Despite some studies that have demonstrated excellent outcomes when lung cancer resections are performed in high-volume hospitals by thoracic surgeons, excellent outcomes are not uniform.^38- 39 Rather, the postoperative mortality rate following resection of lung cancer in the United States averages 5%, and the frequency of serious complications ranges from 20% to 44%.^40- 43 We did not examine the additional morbidities that result from diagnostic procedures or biopsies performed in response to an abnormal CT result, but other investigators have reported that up to 12% of people who are screened for lung cancer with CT undergo invasive biopsies that ultimately reveal 1 or more benign processes.⁴⁴ Thus, these biopsies and the other diagnostic procedures that are performed in response to findings on a screening CT constitute another potential downstream harm that could result from widespread CT screening.

Our findings that CT screening is not associated with a reduction in the chance that a person will develop advanced lung cancer or die from lung cancer are important negative results that should influence how screening is viewed up until that time when more rigorous data are available from randomized trials. Our findings also emphasize the potentially confusing nature of survival analyses in screening studies. The individuals in our study with early lung cancer had excellent lung cancer−specific survival, equivalent to that reported by the International Early Lung Cancer Action Project.²⁵ However, as our study illustrates, excellent survival of a few individuals does not necessarily equate to a benefit overall.

The mechanism responsible for the disconnect between the excellent lung cancer−specific survival observed among the few individuals with early stage cancer found by screening and the unchanged lung cancer mortality seen in the group as a whole is illustrated in Table 3. Few of the individuals who died from lung cancer in our study also belonged to the group who had their lung cancers discovered at an early stage. Instead, despite routine screening, most of the lung cancers that were ultimately fatal were not detected until an advanced stage, or until they caused death. So, although excellent survival of individuals with early stage lung cancer is a necessary condition for CT screening to be beneficial, it is not a sufficient condition. Computed tomography screening must also intercept at an early stage those cancers that will later progress to cause clinical disease and death, and in our study, CT screening did not intercept these cancers. Had it, then the number of deaths from lung cancer would have been lower than the number that we expected. In other words, our results raise further doubts about the premise of lung cancer screening, suggesting that it may be difficult to detect at an early stage a meaningful proportion of the lung cancers that cause clinically significant disease and death, even when using a sensitive technology such as CT.

These findings must be viewed in consideration of our study's limitations. A larger sample size may have allowed us to detect a benefit of screening; our 95% CIs actually allow for a reduction in lung cancer mortality as large as 30%, which would constitute a potentially important public health benefit. Also, a longer period of follow-up, or a longer period of screening, may have allowed us to detect a benefit of screening. The design of the National Lung Screening Trial of CT is germane to this latter concern because the duration of screening and length of follow-up roughly match those in our study. In the National Lung Screening Trial, individuals will be screened 3 times (at baseline and at 2 annual follow-up appointments). All 3 of our studies included at least this many screening evaluations. The National Lung Screening Trial is then powered to detect a 50% reduction in lung cancer mortality within about 2 years of follow-up, and a 20% reduction in mortality within 6 years of the commencement of screening.⁴⁵ In our study, the median amount of follow-up from the initial CT evaluation to the mortality end point was nearly 5 years.

Despite the paucity of evidence supporting lung cancer screening, and no clear delineation of the harms that may result from excess diagnoses, additional diagnostic procedures, and additional treatment, screening is being offered widely, and claims that screening saves lives and should be available to all are widespread.^25,46- 47 Legislation has also been introduced that would require Medicare to cover lung cancer screening.⁴⁸ A more prudent course would be to await the findings of the National Lung Screening Trial and several trials that are being conducted and planned in Europe. It would also be wise to explore other approaches to lung cancer prevention and early detection based on modalities other than regular imaging. Until then, CT screening for lung cancer should be considered an experimental procedure, based on an uncorroborated premise.