Low-dose computed tomography for lung cancer screening in high-risk populations: a systematic review and economic evaluation

Smoking

The main cause of lung cancer is smoking (mainly cigarettes). It is estimated that 85% of lung cancer is attributable to smoking. 2 Smoking cigars, pipes or waterpipes is also associated with lung cancer, as well as other health problems. 3,4

Smoking cessation leads to reduced risk of lung cancer, but the risk still remains higher than for those who have never smoked if there is a significant smoking history. 5 Generally, the risk of lung cancer increases with the number of cigarettes smoked and the duration of smoking, with the pack-year a commonly used measure for smoking history. One pack-year is the equivalent of smoking one pack (20 cigarettes) per day for 1 year (around 7300 cigarettes). Duration of smoking appears to be a more important factor than smoking intensity (the number of cigarettes per day), such that it is worse to smoke one pack per day for 40 years than two packs per day for 20 years. 6

Smoking rates have declined substantially and steadily over the previous 40 years, but there remains a significant proportion of adults who continue to smoke and who have quit smoking but remain at high risk of lung cancer. Since 2005, at least half of the people who were ever regular cigarette smokers have quit. 7

Cigarette smoking rates vary geographically within the UK. 8 Compared with other European Union (EU) countries in 2014, the UK had the second lowest rate of current smokers (17.3%), higher only than Sweden (16.7%). The UK also had a gender gap in the smoking rate of 3.1%, which was significantly below the EU average (9.2%). 9

Smoking is also strongly correlated with income and socioeconomic status (SES). Individuals with a gross personal income of < £20,000 are twice as likely to be current smokers compared with individuals with a gross personal income of ≥ £40,000 (22% vs. 11%). Employment is also significant: 36% of unemployed adults smoke compared with 20% of employed adults, and 15% of economically inactive adults (e.g. pensioners and students). 7

Efforts in the UK to reduce smoking include direct taxation of smoking (tobacco duty),10 smoking cessation services and campaigns (e.g. Smokefree,11 Stoptober12), standardised (plain) packaging (compulsory from May 2017)13 and restrictions on advertising. 14

Passive smoking

Passive smoking is when a person does not smoke themselves, but spends time in an environment where smoking occurs. Passive smoking causes lung cancer and other diseases, including respiratory infections and asthma. For example, women who never smoked but were exposed to second-hand smoke from their spouse face a 27% higher risk of lung cancer than similar women not exposed to second-hand spousal smoke. 15

Environmental risk factors

In addition to smoking, there are other environmental risk factors that, following exposure, have demonstrated an increased risk of lung cancer. These risk factors include asbestos, radon gas and air pollution. When these environmental risk factors are combined with smoking, the overall risk of lung cancer is exacerbated. 16

Lung cancer cell type

Lung cancer can be divided into two broad categories: (1) small cell lung cancer (SCLC) and (2) non-small cell lung cancer (NSCLC). SCLC is a highly malignant tumour accounting for 11% of all lung cancer cases in the UK, and is derived from cells exhibiting neuroendocrine characteristics. 18 NSCLC accounts for 88% of all lung cancer cases in the UK. NSCLC can further be divided into three major pathologic subtypes: (1) adenocarcinoma, (2) squamous cell carcinoma and (3) large cell carcinoma. Statistics from the National Lung Cancer Audit Annual Report 201618 report the distribution of NSCLC in the UK as follows: adenocarcinoma 36%, squamous 22%, other 11% and no pathology 31%. 18

Neuroendocrine tumours account for around one-quarter of primary lung neoplasms, including SCLC (20%), large cell neuroendocrine carcinoma (LCNEC) (a subtype of large cell carcinoma; 3%) and carcinoid tumours (typical and atypical carcinoid; 2%). 19 Lung neuroendocrine tumours arise from cells of the diffuse neuroendocrine system in the bronchial mucosa. 20

The World Health Organization (WHO) classification of lung tumours21 now recommends (2015 classification) greater restriction on the diagnosis of large cell carcinoma (former subtypes now reclassified), reclassification of adenocarcinoma subtypes, reclassification of squamous cell carcinoma subtypes and a number of smaller changes. 21

Grade

Grading of lung cancer for NSCLC distinguishes lung cancer cells into groups based on the cell’s appearance:22

• Grade 1 – cells look normal, will grow slowly and are less likely to spread (may also be called ‘low grade’ or ‘well differentiated’).

• Grade 2 – cells look abnormal and are more likely to spread (may also be called ‘moderate grade’ or ‘moderately well differentiated’).

• Grades 3 and 4 – cells look very abnormal, will grow quickly and are more likely to spread (may also be called ‘high grade’ or ‘poorly differentiated’).

Neuroendocrine tumours are also graded:19

• low grade (typical carcinoid)

• intermediate grade (atypical carcinoid)

• high grade (SCLC and LCNEC).

Stage

The stage of the lung cancer relates to the extent of the cancer and is currently classified by the NHS in the UK according to the international tumour, node, metastasis (TNM) system (Seventh Edition). 23,24 The Eighth Edition of the TNM system has been recently published25 and is due to be adopted in the UK. There are three components that are combined to make up the overall staging of lung cancer: (1) the size of the tumour (diameter in greatest dimension; TX, T0–T4), (2) whether or not it has spread to the lymph nodes (NX, N0–N3) and (3) whether or not it has distant metastasis (MX, M0, M1). 26 In the UK, the following percentages of cases recorded postoperatively by stage were reported by the National Lung Cancer Audit Annual Report 2016:18

• stage IA, 11%

• stage IB, 7%

• stage IIA, 4%

• stage IIB, 4%

• stage IIIA, 12%

• stage IIIB, 9%

• stage IV, 53%.

Note that, at present, there is no national lung cancer screening programme, so the vast majority of these cases are either clinically presenting or incidental findings.

Presentation

Diagnosis of lung cancer generally follows symptomatic presentation of (typically late-stage) lung cancer, but it can also be an incidental finding during other investigations.

Of the 73,063 lung cancer cases diagnosed in England (2012–13), 25,668 presented as an emergency, 20,420 were referred by a general practitioner (GP) through the 2-week wait pathway and 15,525 were otherwise referred by a GP. 27 The remaining cases presented through some other route or were identified only through a death certificate. Stages I–III were most often diagnosed following GP referral, whereas stage IV and unstaged cancers were most often diagnosed following emergency presentation. 27

The National Institute for Health and Care Excellence (NICE) guidelines on suspected cancer referral28 identify a number of key symptoms and findings that should prompt an urgent referral or chest X-ray (CXR), which include (not an exhaustive list):

• unexplained haemoptysis (coughing blood or blood-stained mucus)

• respiratory symptoms (cough, shortness of breath)

• chest symptoms and signs (chest pain, persistent or recurring chest infection)

• general symptoms and signs (fatigue, weight loss, appetite loss)

• finger clubbing.

In addition, there are decision aids for GPs that can help to identify the significance of combinations of symptoms. 29

Investigation

The 2011 NICE guidelines on lung cancer diagnosis and management30 recommend an investigative pathway for suspected lung cancer that includes contrast-enhanced chest computed tomography (CT) to further the diagnosis and contribute to staging the disease. Positron emission tomography–computerised tomography (PET-CT) is recommended for patients who are potentially suitable for treatment with curative intent. Other investigations are recommended according to the location and spread of the disease; many of these are invasive and involve endoscopy (including bronchoscopy) and biopsy.

Incidence

Lung cancer is the most common cancer in the world, with 1.8 million new cases diagnosed in 2012. 35

In the UK in 2014, 46,400 new cases of lung cancer were diagnosed (130 new cases per day on average). Lung cancer is the third most common cancer in the UK and, in 2014, it accounted for 13% of all new cases of cancer in the UK. The directly age-standardised rates of lung cancer incidence in England (2014) were 91.6 and 65.2 per 100,000 person-years for men and women, respectively, but the incidence rate climbs with age. 1

The incidence of lung cancer has declined for men compared with historical levels, but has climbed for women. 1 Men continue to be at a higher risk of lung cancer than women. Across the UK, the age-standardised rate of lung cancer decreased from 2012 to 2014, to approximately 2006 levels.

Mortality

Lung cancer is the most common cause of deaths from cancer in the EU (20.8% of all cancer-related deaths). 36 Lung cancer accounted for 5.4% of the total number of deaths in the EU (2013), equating to more than one-quarter of a million people (268,744 people) or 55.2 deaths per 100,000 inhabitants. 36

There were 35,895 lung cancer deaths in the UK in 2014,37–39 accounting for 22% of all cancer deaths; lung cancer is the most common cause of cancer death. Crude mortality rates indicate that there are 62 lung cancer deaths for every 100,000 males and 50 for every 100,000 females. 37–40 In England in 2014, 15,856 men and 12,993 women died from lung cancer (19,563 and 16,332, respectively, for the UK). 37–39

Prevalence

The prevalence of lung cancer (i.e. the proportion of people previously diagnosed with lung cancer who are still alive at a given time) is relatively low because survival is generally poor. Worldwide, the overall ratio of mortality to incidence is 0.87, because of the high fatality rate associated with lung cancer.

In 2006, it was estimated that 38,141 people were alive in the UK who had been diagnosed with lung cancer within the previous 10 years, 15,802 of whom had been diagnosed within the previous year. 41

Significance for lung cancer patients

The prognosis for those diagnosed with lung cancer is disappointing despite recent advances in oncology and surgery. This is because lung cancer is typically diagnosed when the cancer is in the later stages, and in older people, who often have concomitant diseases that subsequently limit therapeutic options.

For individuals with NSCLC, the primary symptoms (or most frequent) are appetite loss (98%), fatigue (98%), shortness of breath (94%), cough (93%), pain (90%) and blood in sputum (70%). From the literature, it appears that fatigue is the primary symptom that has an impact on daily living for those with lung cancer, followed by pain (chest pain/pain swallowing). 42 Quality-of-life (QoL) scores deteriorate as the number of chemotherapy cycles increases (i.e. the longer the treatment lasts, the lower the QoL). 42 Being treated for lung cancer also has a detrimental impact on an individual’s social life, family life and ability to work. 42

Significance for the NHS

Luengo-Fernandez et al. 43 estimated that lung cancer had the highest economic cost of any cancer in the UK, but health-care costs are greater for colorectal cancer and breast cancer (lung cancer leads to significantly more productivity losses).

Inclusion criteria

Exclusion criteria

Studies were excluded if they did not match the inclusion criteria. In addition, certain studies were not considered, particularly:

• animal models

• preclinical and biological studies

• non-systematic reviews, editorials, opinions

• non-English language papers

• reports published as meeting abstracts only, as there is unlikely to be sufficient methodological details to allow critical appraisal of study quality.

At least two reviewers independently screened the titles and abstracts (if available) of all reports identified by the search strategy. Full-text copies of all studies deemed to be potentially relevant were obtained and two reviewers independently assessed them for inclusion. Any disagreements were resolved by consensus or by a third reviewer.

Decision problem

The set of outcomes was expanded to include the following, after consultation with clinical experts:

• number of patients who were more amenable to surgical treatment

• surgical resection rate

• smoking cessation and patients’ smoking behaviour change

• adherence rate to screening

• overdiagnosis

• complications in those who underwent an invasive procedure

• radiation dose of screening

• radiation-related patient outcomes

• adverse psychological impact.

Search methods

The following resources were not searched:

• National Research Register

• Food and Drug Administration website

• European Medicines Agency website.

In addition, the WHO trial registry was searched.

Subgroups

The prospectively registered protocol indicated that heterogeneity would be explored through consideration of the study populations, methods and interventions. In the review, heterogeneity was specifically explored through consideration of study quality, which relates to study methods, but was not explicitly listed in the protocol.

Characteristics of included studies

Table 1 presents the summary information of characteristics of included trials for the systematic review of clinical effectiveness. All of the included studies were RCTs. Nine studies were conducted in European countries and three studies were conducted in the USA. Two trials (one of which was a pilot trial) were conducted in the UK. Only a minority of included trials contributed to important comparative outcomes.

Low-dose CT screening was a key component of the screening programmes. Most of the included trials used usual care as a comparator, whereas three trials used CXR as a comparator. The sample size of included trials ranged from 190 to 53,434. The included trials recruited participants with age ranging from 47 to 80 years. The number of screening rounds ranged from 1 to 10. Most trials adopted 1-year interval screening. However, one trial [Multicentric Italian Lung Detection (MILD)]69 used both annual and biennial screening and the NELSON trial56 performed screening at baseline, 1 year, 3 years and 5.5 years of follow-up. When reported, the duration of follow-up ranged from 2 to 9.53 years.

All included trials recruited high-risk populations. The characteristics of study populations are shown in Appendix 4. The percentage of male participants ranged from 32% to 100%. All the studies recruited current smokers and former smokers. Most trials recruited participants through targeted mailings of questionnaires via GPs and family doctors, media, internet and newspaper advertisements. The characteristics of recruitment methods are shown in Appendix 4.

As seen from this table, there were wide variations in definitions of high risk between trials. For example, the National Lung Screening Trial (NLST)71 (which was conducted in the USA) used two variables (age and smoking history) to define high risk:

• aged 55–74 years

• current smokers with at least a 30 pack-year smoking history

• former smokers (who had quit within the previous 15 years) with at least a 30 pack-year smoking history.

However, UKLS55 used the Liverpool Lung Project lung cancer risk prediction algorithm to predict high risk. This risk prediction rule has been validated in three independent studies from Europe and North America and demonstrated its predictive benefit. The following variables were included in this risk prediction model:

• age

• sex

• prior diagnosis of pneumonia

• family history of lung cancer

• smoking duration

• prior diagnosis of malignant tumour

• personal history of other cancer and non-malignant lung diseases

• early onset (< 60 years of age) family history of lung cancer.

In this trial,55 participants were selected based on the prediction result (i.e. ≥ 5% risk of developing lung cancer in the next 5 years). It should be noted that such variations in the definition of ‘high-risk’ participants can lead to different prevalences of lung cancer being detected at baseline between different trials.

Furthermore, the Detection and Screening of Early Lung Cancer with Novel Imaging Technology and Molecular Essays (DANTE) trial61 defined high-risk participants as those smokers or former smokers (aged 60–74 years) of at least 20 pack-years who had quit < 10 years before recruitment. Similar criteria were adopted by the Italian lung cancer screening (ITALUNG) trial. 72 The Despiscan trial62 defined high-risk patients as those aged 50–75 years who were current or former smokers (having quit < 15 years from enrolment) with cigarette consumption of ≥ 15 cigarettes per day for ≥ 20 years. The Danish Lung Cancer Screening Trial (DLCST)63 defined high-risk participants as those current or previous smokers (aged 50–70 years) with ≥ 20 pack-years of smoking. Previous smokers had to have quit after the age of 50 years and < 10 years prior to the start of the study. Patients had to be able to climb 36 steps without pause. In this trial, forced expiratory volume in 1 second (FEV₁) had to be at least 30% of predicted normal at baseline. It should be noted that the MILD trial69 recruited younger participants (≥ 49 years) who were current or former smokers (having quit smoking within 10 years of recruitment) with ≥ 20 pack-years of smoking. 69

The DANTE trial61 recruited only male participants, whereas most trials recruited both male and female participants. The NELSON trial57 recruited at first only men, and later also women (aged 50–75 years), who were current and former smokers with ≤ 10 years of cessation, who smoked > 15 cigarettes per day for > 25 years or > 10 cigarettes per day for > 30 years.

The characteristics of screening programmes are shown in Appendix 4. Most studies compared LDCT screening with usual care (no screening), whereas three studies compared LDCT screening with CXR screening. As seen in Appendix 4, Table 43, the definitions of a positive scan varied across studies in terms of nodule sizes. For example, the NELSON trial57 defined positive CT scans as those non-calcified nodules that had a solid component of > 500 mm³ (> 9.8-mm diameter) or volume-doubling time of < 400 days. If the volume of largest solid nodule or the solid component of a partially solid nodule was 50–500 mm³ (4.6–9.8 mm in diameter) or > 8 mm in diameter for non-solid nodules or volume-doubling time was 400–600 days, these results were treated as indeterminate test results. In NLST,71 any non-calcified nodule measuring ≥ 4 mm in any diameter and radiographic images were classified as positive, suspicious for lung cancer. Other abnormalities (e.g. adenopathy or effusion) could be positive or suspicious. As per the protocol, abnormal findings suspicious for lung cancer that were stable across the three screening rounds were classified as minor abnormalities rather than positive findings.

There were variations in imaging evaluation and interpretation strategy across included trials. When reported, two radiologists independently interpreted and reported the results in most trials. If there was disagreement, final interpretation was based on joint consensus.

The diagnostic follow-up strategies for suspicious abnormality findings varied between studies. When reported, most studies used further diagnostic imaging {e.g. high-resolution CT or chest fludeoxyglucose (¹⁸F) positron emission tomography ([¹⁸F] FDG-PET)} and/or invasive biopsy with rapid on-site examination.

Computed tomography parameters

The key CT technical specifications of the included studies are presented in Tables 2 and 3 (see the Glossary for definitions of CT parameters). Given the selective reporting of CT vendors, the available information suggests that the trials transition from single (single bank of detectors) to multislice (multiple banks of detectors) technology during a time of rapid CT development. The DANTE,61 Garg et al. ,64 ITALUNG72 and German Lung Cancer Screening Intervention (LUSI)68 trials have incorporated single-slice technology. The advantages of multislice over single-slice technology mainly include the same acquisition in shorter time, better z-axis resolution and the capacity to scan larger volumes in the same time. Specific to CT of the thorax, multislice technology allows for quicker scanning, resulting in reduced breathing artefact and better quantification of thoracic lesions where present.

In general, the slice thickness ranges between 1 and 3 mm. These are generally considered ‘thin’ slices and were considered superior to ‘thick’ slices. Reconstructed slice thickness (which includes the consideration of pitch and collimation) is complicated in helical multidetector compared with single-detector scanning. In helical scanning, reducing slice thickness increases z-axis resolution but this results in a trade-off with increased image noise and possibly dose. The DANTE trial61 adopted a slice thickness of 5 mm, which is at odds with the other trials.

In the last decade, the development of multislice technology expanded the applications of CT, leading to the increased number of examinations and radiation exposure. Given the concern about the rise of medical radiation, automatic tube current modulation was designed to achieve the same image quality for individuals with different biological make-up/patient attenuation characteristics, reducing radiation exposure. In the setting of LDCT screening, there are two broad strategies in performing CT thorax, either (1) tube current modulation (as described) or (2) a fixed-tube current approach. In this report, three main strategies of dose reduction have been identified:

1. fixed-tube current and voltage regardless of the body mass index (BMI)

2. fixed-tube current and voltage depending on the BMI

3. automatic tube current and voltage depending on the BMI.

Considering the selective reporting, each LDCT thorax strategy results in different radiation output but the reported doses are generally lower than the doses of standard CT thorax in the literature.

Some trials conducted volumetric analysis using software mostly developed by the CT vendors. DLCST63 used semiautomated software designed by Philips, whereas MILD,69 NELSON56,57 and UKLS55 used semiautomated software made by Siemens. LUSI68 used computer-aided detection software made by MEDIAN. The comparability of volumetric measurements using different software is not known.

Estimation of radiation dose is achieved by multiplying the dose length product by a conversion factor; none of the studies reported whether or not the commonly used conversion factor (0.014) was used in estimating the average radiation dose. However, all studies used radiation doses that would be considered low by CT standards and are less than annual background radiation exposure. Image quality has not been considered as an outcome in the trials, which is influenced by patient demographics.

Ongoing studies

A total of 125 ongoing trials were identified in the search and investigated further. Of these, only two were considered relevant to this review. One is the MILD trial69 (NCT02837809), which is already included, and the second is a RCT based in China (NCT02898441), which is due for completion by December 2018. This study compares LDCT screening for lung cancer with usual care. The anticipated recruitment is 6000 participants and the primary outcome is lung cancer incidence. The 5-year follow-up is planned for lung cancer incidence, lung cancer mortality and all-cause mortality.

Outcome measures in included trials

In order to demonstrate the variability of reporting in the included trials, Table 4 displays which outcome is measured and whether or not it includes ≥ 5 years of follow-up. Additional information, when available, is given on whether or not the outcome was predefined in a protocol and, if so, whether it was categorised as a primary or secondary outcome.

Of the 12 included studies, predefined outcomes could not be verified for five studies. 61,62,64,66,68 Only the ITALUNG trial65 had a full protocol published,65 whereas the remaining trials listed outcomes in relevant clinical trials registries. 55,57,63,67,69,70 Five studies62,64,66–68 have no data on the outcomes of interest. A sixth trial (ITALUNG65) has recently reported results for several outcomes; however, as they were published after the inclusion date for the current review, these results were not incorporated in the analyses.

Lung cancer mortality, all-cause mortality and cancer incidence with > 5 years’ follow-up are reported in four studies. 61,63,69,70 When lung cancer mortality is cited as a primary outcome, data for two studies57,65 were not identified. However, one of these is the ITALUNG trial,65 as previously mentioned, and the second is the more recent NELSON trial,57 for which results may be published imminently. As a secondary outcome in the LungSEARCH trial,67 data for lung cancer mortality remain unreported.

Cancer incidence is defined as a secondary outcome for three studies, with results available for two studies63,70 in this review and the third being the ITALUNG trial. 65

Data for stage distribution are provided for three studies61,63,70 with ≥ 5 years’ follow-up. Although the LungSEARCH trial67 has stage distribution defined as a primary outcome, no results were available.

Complete resection is the least reported outcome, with only two studies61,70 providing data for ≥ 5 years.

Smoking cessation was reported in three studies,56,61,63 with a fourth study defining this as a secondary outcome; however, the results have not been published. 69

With regard to HRQoL, this is reported for four studies;55,57,63,70 however, the follow-up is < 5 years and actually < 1 year for the NLST trial. Furthermore, the NELSON trial57 included only a subsample of participants for this outcome.

Risk of bias of included studies

All the studies were assessed for risk of bias using the Cochrane Risk of Bias tool. 76 However, as indicated in Table 4, only a minority of the included studies contributed data to the consideration of the main outcome measures assessed through a comparison of the intervention arm with the control. Thus, the reporting of quality focuses on the risk of bias for the studies contributing results of the main outcomes, particularly separating mortality, psychological consequences/HRQoL and smoking cessation, as we noted that not only were there different included studies for these outcomes, but also differing threats to validity depending on the outcome. This stemmed from variation in the objectivity of the outcome and different losses to follow-up between outcomes within a trial.

Comparative outcomes

Lung cancer mortality

Four RCTs (DANTE, DLCST, MILD and NLST) assessed the effects of LDCT screening compared with either usual care (no screening) or the best available care (CXR screening), and reported lung cancer mortality at long-term follow-up. 61,63,69,71 Over the long-term follow-up, it was likely that CXR could be an element that constituted usual care for the early detection of lung cancer in a high-risk population. Therefore, usual care in this context was not dissimilar to the best available care when CXR was used in early detection of lung cancer. For this reason, we performed statistical pooling of all the four RCTs on mortality outcomes. All trials were conducted in participants at high risk for lung cancer. We only performed statistical pooling for trials that reported lung cancer mortality data with ≥ 5 years of follow-up, using the random-effects model.

Figure 2 shows the overall pooled result of four RCTs comparing LDCT screening with controls. When compared with controls (usual care/best available care), LDCT screening was associated with a non-statistically significant reduction in lung cancer mortality (pooled RR 0.94, 95% CI 0.74 to 1.19) with up to 9.80 years of follow-up. There was moderate heterogeneity in the magnitude of effects (I² = 43.3%).

FIGURE 2.

Lung cancer mortality: overall results.

It is important to note that, given the moderate heterogeneity observed with this outcome (I² = 43.3%), the pooled non-statistically significant decrease in lung cancer mortality (pooled RR 0.94, 95% CI 0.74 to 1.19) should be treated with caution.

Exploration of heterogeneity

The quality of trials

Among these four RCTs, the MILD trial69 was judged to be of poor quality, whereas the remaining trials were judged to be of moderate to high quality (see Risk of bias for lung cancer and overall mortality). 55 We explored the impact of trial quality on the robustness of overall results.

Figure 3 presents the sensitivity analysis by excluding the poor-quality trial (MILD69). The results of the sensitivity analysis showed that, compared with controls (usual care/best available care), LDCT screening demonstrated a statistically significant reduction in lung cancer mortality (pooled RR 0.85, 95% CI 0.74 to 0.98). The level of heterogeneity was considerably reduced (I² = 6.9%). This suggests that variation in trial quality could be a potential source of heterogeneity.

FIGURE 3.

Lung cancer mortality: sensitivity analysis by excluding the low-quality trial (MILD69).

Annual low-dose computed tomography screening versus usual care

We further explored the heterogeneity on the basis of frequency of screening interventions (e.g. annual screening or biennial screening). We conducted a subgroup analysis based on the annual LDCT screening compared with usual care only. Figure 4 presents the pooled results of this subgroup analysis. When compared with usual care (no screening), LDCT screening programme demonstrated a non-statistically significant increase in lung cancer mortality outcome, with a pooled RR of 1.15 (95% CI 0.79 to 1.67) from three RCTs. The level of heterogeneity was reduced (I² = 38.9%).

FIGURE 4.

Lung cancer mortality subgroup analysis: annual LDCT screening vs. usual care (no screening).

All-cause mortality

Four RCTs (DANTE, DLCST, MILD and NLST) assessed the effects of LDCT screening compared with either usual care (no screening) or the best available care (CXR screening), and reported all-cause mortality outcome with ≥ 5 years of follow-up (see Appendix 4). 61,63,69,71 Likewise, considering that CXR could be an element that constituted usual care for the early detection of lung cancer in high-risk populations, we performed statistical pooling of all these four RCTs on all-cause mortality outcome in a similar fashion.

Figure 5 shows the overall pooled result of four RCTs comparing LDCT screening with controls. When compared with controls (usual care/best available care), LDCT screening demonstrated no statistically significant difference in all-cause mortality outcome (pooled RR 1.00, 95% CI 0.87 to 1.16) with up to 9.80 years of follow-up. There was substantial heterogeneity associated with this outcome (I² = 57.0%). Sensitivity analysis and subgroup analysis were also performed to explore the potential sources of heterogeneity.

FIGURE 5.

All-cause mortality: overall results.

Given the substantial heterogeneity detected between studies, the results from this pooled analysis should be treated with caution.

Exploration of heterogeneity

The quality of trials

We assessed the impact of trial quality on the robustness of overall results as a means to explore heterogeneity. The details of trial quality are presented in Quantity and quality of research available, and Figure 6 shows the results of removing the low-quality trial (MILD). 55 Based on the pooled data, the LDCT screening programme demonstrated a non-statistically significant decrease in all-cause mortality (pooled RR 0.95, 95% CI 0.89 to 1.00) compared with controls (usual care/best available care).

FIGURE 6.

All-cause mortality: sensitivity analysis by excluding the low-quality trial (MILD69).

It should be noted that the level of heterogeneity was considerably reduced (I² = 0%), suggesting that variation in trial quality could be a potential source of heterogeneity between studies.

Annual low-dose computed tomography screening versus usual care

A subgroup analysis are performed on the basis of annual LDCT screening compared with usual care. Three trials evaluated the effect of annual LDCT screening versus usual care (no screening).

The pooled results of this subgroup analysis are shown in Figure 7. Annual LDCT screening showed a non-statistically significant increase in all-cause mortality compared with usual care, with a pooled RR of 1.15 (95% CI 0.84 to 1.58). There was substantial heterogeneity (I² = 75.2%) observed in this outcome.

FIGURE 7.

All-cause mortality subgroup analysis: annual screening LDCT vs. usual care.

Cancer detection

We performed meta-analyses for three studies providing comparative cancer incidence data (with a control group) with ≥ 5-years’ follow-up, as this was judged sufficiently long to obtain mature data. 61,63,71 Compared with controls (usual care/best available care), LDCT screening was associated with a statistically significant increase in lung cancer detection rate (pooled RR 1.38, 95% CI 1.02 to 1.86; I² = 79.7%) with ≥ 5 years’ follow-up (Figure 8). When compared with usual care (no screening) only, the sensitivity analysis demonstrated a consistent result (pooled RR 1.58, 95% CI 1.15 to 2.19) over 5 years after last screening, with a reduction in the level of heterogeneity (I² = 54.6%).

FIGURE 8.

Cancer detection difference between LDCT and controls: overall results.

Stage distribution

We performed meta-analyses for trials that provided relevant data for ≥ 5 years’ follow-up as this was judged sufficiently long to obtain mature data. Figure 9 illustrates the change of lung cancer stage distribution between LDCT screening and control arms. When pooling data from three trials with ≤ 6 years’ follow-up, lung cancers detected in the LDCT screening arms (including interval cancers and cancers diagnosed after the final screening round) were more likely to be early stage (I and II) than those in the control arm, and this was statistically significant (RR 1.73, 95% CI 1.27 to 2.37, I² = 61%). Correspondingly, these cancers were less likely to be late stage (RR 0.67, 97% CI 0.60 to 0.75, I² = 22%).

FIGURE 9.

Early cancer stage (stage I and II) distribution between LDCT screening and control arms. M–H, Mantel–Haenszel.

We performed sensitivity analyses to explore the potential sources of heterogeneity. When pooling data of two trials comparing LDCT screening with usual care (no screening) only as part of investigation of heterogeneity, LDCT screening was associated with a statistically significant increase in early stage (I and II) cancer detection (RR 2.09, 95% CI 1.34 to 3.24) with ≤ 5 years’ follow-up. The level of heterogeneity was considerably reduced (from I² of 61% to 39.1%).

The change in stage distribution in these analyses could arise through two mechanisms: (1) cancers that would have presented clinically as late stage are instead detected at the earlier stages (stage shift) and (2) cancers that would never have presented clinically are instead detected at the earlier stages (overdiagnosis). Figure 10 shows the impact of screening on the risk of late-stage lung cancer diagnosis. This is minimally affected by overdiagnosis because late-stage lung cancers are associated with poor survival. The results from three studies indicate that, on average, there is a risk reduction of 15% associated with LDCT screening, which is just statistically significant (RR 0.85, 95% CI 0.73 to 1.00; I² = 17%).

FIGURE 10.

Late-stage lung cancer risk between LDCT screening and control arms. M–H, Mantel–Haenszel.

A sensitivity analysis comparing LDCT screening with usual care (no screening), that is, excluding NLST,71 indicated no statistically significant evidence of an impact on the late-stage lung cancer risk (RR 1.00, 95% CI 0.75 to 1.34; I² = 0%).

Health-related quality of life and psychological consequences

Four trials (NELSON, NLST, DLCST and UKLS) evaluated the psychological consequences on patients of LDCT screening and patients’ HRQoL (Table 9). 55,57,63,71 Two trials (NELSON and NLST) assessed the impact of LDCT screening on patients’ HRQoL measures. 57,71 Both NELSON and NLST trials assessed this HRQoL outcome at short- and long-term follow-up. Four trials (DLCST, NELSON, UKLS and NLST) assessed adverse psychological impact associated with LDCT screening. 55,57,63,71

All included studies suffered from additional challenges to their validity relative to mortality outcomes, resulting from subjectivity of outcomes in trials that were not blinded and loss to follow-up. Three studies (NLST, NELSON and DLSCT) evaluated these outcomes in subsamples of their whole-trial populations. Hence, the size of the whole trial is not a good guide to quantity of evidence provided by each study on these outcomes. The sample size of each study was UKLS (n = 4061), DLSCT (n = 3929), NLST (n = 2812) and NELSON (n = 1466).

As a general comment, when normal values for psychological consequences and HRQoL were provided, the levels encountered were generally well outside the scale values, indicating marked psychological distress or QoL outside that which would be expected in the normal population.

Change in smoking behaviour

Lung cancer screening may have a potential impact on participants’ smoking behaviour, for example, providing a new opportunity for attempts to quit smoking. One potential drawback of lung cancer screening is that it may reduce smokers’ motivation to quit smoking by inducing a sense of assurance, thereby delaying smoking cessation. Three trials (DLCST, NELSON and NLST) reported relevant data on patients’ behaviour change in smoking within the trial period (Table 10). 57,63,71

The NLST71 and DLCST63 provided randomised evidence on all participants randomised (n = 4104 and n = 53,456, respectively). Evidence from NLST was reported for each of the two contributing research networks, LSS and ACRIN. The evidence for NELSON was on a much smaller scale (n = 1284) based on a sample of the whole trial. Additional evidence from UKLS is reported to be in the process of publication.

The randomised evidence for smoking behaviour was more open to bias than evidence on mortality. The absence of blinding would have influenced the validity of self-reported smoking status relative to more objectives. In addition, DLCST63 and NELSON56,57 were open to further bias through loss to follow-up, although this was greater in the no-screening arm. The fact that NLST71 was compared with an active intervention, CXR, rather than no screening in DLCST63 and NELSON56,57 is also noteworthy.

Non-comparative outcomes

Nodule detection and lung cancer detection rate

Four trials57,61,63,71 reported the number of participants with non-calcified lung nodules identified on screening over the study period. Figure 11 presents the rate of participants with non-calcified lung nodules over the study threshold in the LDCT arm. As shown in this figure, there were wide variations in the percentages of participants with non-calcified lung nodules over the study threshold in the LDCT arm. It ranged from 11% to 69%, indicating that all these trials may have used different criteria to define a positive LDCT scan.

FIGURE 11.

Positive scans detected in the LDCT screening arm.

Diagnostic follow-up evaluations

There were variations in the diagnostic methods used to follow nodules that were identified as ‘positive’ scans. Most positive scans were resolved by comparison with prior scans or further diagnostic imaging. Figure 12 presents the proportions of participants with additional diagnostic CT scans in the CT screening arm. It ranged from 5% to 52%. It should be noted that both NELSON and NLST trials had similar proportions (33% to 34%) of participants with additional diagnostic CT scans in the CT screening arm. However, UKLS achieved the lowest proportion (5%) of participants who underwent further diagnostic CT scans. This reflects the effective nodule management protocols being used in the UK setting.

FIGURE 12.

Proportion of participants with additional diagnostic CT scan.

Figure 13 presents the proportions of participants who underwent surgical biopsy or procedures for diagnosis in the CT screening arm. It ranged from 2% to 4%. This figure appears to be consistent across these trials.

FIGURE 13.

Proportion of patients undergoing surgical biopsy or procedures for diagnosis.

Lung cancer detection rate

Figure 14 shows the lung cancer prevalence rate at baseline. The lung cancer prevalence rate ranged from 1% to 2% across different trials, indicating that included trials used different definitions of high-risk populations. For example, UKLS55 used a risk prediction model to define ‘high risk’, whereas both the NLST70 and the NELSON56 trial used only two variables (age and smoking status) to define ‘high risk’. We discuss this further in Discussion.

FIGURE 14.

Lung cancer prevalence rate at baseline.

Figure 15 presents the cumulative lung cancer detection rates for LDCT screening. Six trials reported the cumulative lung cancer detection rates of all the screening rounds. 57,61,63,65,69,71 The cumulative lung cancer detection rate ranged from 1.7% to 5.2% for the LDCT screening arms.

FIGURE 15.

Cumulative lung cancer detection rate in CT screening.

As seen in the figure, the NELSON trial reported that the cumulative lung cancer detection rate of the three screening rounds was 2.6%. There was a relatively stable detection rate of lung cancer across the three screening rounds. NLST reported a detection rate of 2.5% across all three screening rounds. DLCST63 reported a detection rate of 3.4%. The ITALUNG trial65 reported that the cumulative lung cancer detection rate was 2.7%. The MILD trial69 reported that the cumulative lung cancer detection rate was 2.4% for annual screening and 1.7% for biannual screening. UKLS reported that the cumulative lung cancer detection rate was 2.1% for the pilot trial period. The DANTE trial61 reported that cumulative lung cancer detection rate was 5.2%.

Interval cancer findings

Two trials reported interval cancer findings (i.e. those cancers were detected during the two rounds of screenings). One large US-based trial70 reported that, in the LDCT screening group, among 1060 total lung cancer cases detected, there were 367 (35%) participants who either missed the screening or received the diagnosis after their trial screening phase. In the CXR group, among 941 total lung cancer cases detected, there were 525 (56%) participants who either missed the screening or received the diagnosis after their trial screening phase.

Another large trial (NELSON) also reported that, among 200 total lung cancer cases detected, 52 (26%) cancers were detected between two different screening rounds (i.e. received the diagnosis of lung cancer after their trial screening phase). 56 The NELSON trial56 defined interval cancers as (1) lung cancers diagnosed after negative screening results, (2) lung cancers diagnosed after indeterminate screening results but without any follow-up LDCT investigations or diagnostic examinations in the screening arm or (3) lung cancers diagnosed after a positive screening test if the diagnostic investigation on the positive screening result did not lead to a diagnosis of lung cancer at that stage but a confirmed diagnosis was made later as patients’ symptoms triggered further diagnostic work-up that resulted in a diagnosis of lung cancer.

False positives

False positives are one of the key outcomes for the evaluation of important adverse effects associated with LDCT screening for early detection of lung cancers. For those patients with positive scans, further evaluations often involve more non-invasive evaluations (e.g. imaging scans) and more invasive procedures (e.g. bronchoscopy, fine-needle biopsy and/or surgery). It is important to note that for patients with false positives, these further evaluations may be even harmful as a result of serious complications (e.g. serious infections and postoperative deaths associated with surgery).

Five trials55,57,61,63,69 reported the percentage of scans in the screening arm performed over the trial period that had a false positive result, ranging from 1.2% to 23% (Figure 16). It should be noted that such variations could be because of differences in the definition of a positive scan among included trials (e.g. different thresholds of lung nodule sizes being categorised as a positive scan).

FIGURE 16.

Proportion of scans in the screening arm with a false-positive result.

One large trial57 (NELSON with a total of 24,354 CT scans) reported that 59.4% (293 out of 493; 95% CI 54.8% to 63.9%) of positive screen results were false positive for a follow-up period of 5.5 years. The NELSON trial reported that the positive screenings had a predictive value of 40.6%. This trial reported that 55% to 65% of positive screen results were false positive across four screening arounds. A total of 493 positive results across three rounds of the NELSON trial led to a diagnosis of lung cancer in 200 patients.

In the NELSON trial,57 around 24.5% (n = 67) of those participants with false-positive screen results underwent an invasive procedure in the diagnostic work-up. A total of 91% (n = 61) of these invasive procedures were surgeries (including thoracotomies, mediastinoscopies, sternotomy and video-assisted thoracoscopies) and the remaining six procedures were transthoracic biopsies.

In NLST, across the three rounds, 96% of the positive scans in the LDCT group and 95% of those in the radiography group were false positive. These percentages differed little by screening round. The sensitivity of LDCT screening varied from 93% to 94%, whereas the sensitivity of CXR screening varied from 64% to 74% across three rounds. The specificity of LDCT screening varied from 73% to 84%, whereas the specificity of CXR screening varied from 91% to 95% across three rounds. The NLST reported complications from diagnostic procedures used to evaluate a positive LDCT scan. In this trial, around 1.4% of participants had at least one complication in the CT screening group and 1.6% in the CXR screening group.

However, the high positive rate in NLST reflects the fact that there was no distinction being made for those indeterminate findings or interval imaging findings from false positives.

It is important to note that recent trials (such as NELSON and UKLS) have made a distinction of the definitions of indeterminate findings or interval imaging rate from false positives. For example, in the NELSON trial, the LDCT screening result was indeterminate in 10.8% (2629 out of 24,354) of the all scans across three rounds of CT screening. A CT scan in the NELSON trial was considered indeterminate if the volume of the largest solid nodule or the solid component of a partially solid nodule was 50–500 mm³ or > 8 mm in diameter for non-solid nodule. Indeterminate results led to invitations for a repeat scan (after 6–8 weeks or after 12 months, depending on the nodule size and screening round) in order to determine the final result as positive or negative.

In UKLS,55 a false positive was counted if a participant did not have lung cancer but was referred to the multidisciplinary team (MDT) and/or was subjected to repeated imaging scans before 12 months had elapsed. The UKLS reported that the interval imaging rate for the category 3 (larger, potentially malignant) nodules was 23.2%.

False negatives

Another concern of using CT screening for lung cancer is the potential adverse effect associated with false-negative results. Four trials (DANTE,78 MILD,92 NLST93 and NELSON94) reported that the sensitivity of CT screening for the detection of lung cancer ranged from 69% to 94%. The data indicated that the false-negative examination rates ranged from 6% to 21%. Three trials57,61,69 reported the percentage of scans in the low-dose screening arm that had a false-negative result, ranging from 0.1% to 1.3% (Figure 17).

FIGURE 17.

Proportion of scans in the screening arm with a false-negative result.

Most trials reported a false-negative result as a LDCT scan that was negative within 1 year. However, there were no studies evaluating the potential harm associated with false-negative examinations. One potential adverse effect associated with false reassurance of patients is that it can cause delayed examinations of future suspicious symptoms in patients. Given this consideration, participants should be aware of the importance of reporting symptoms (such as cough, haemoptysis or weight loss) even if they have a negative screening scan.

Key outcomes

Other outcomes

Overall analysis

The overall network meta-analysis included six trials: three trials comparing LDCT with usual care, one trial comparing LDCT with CXR and two trials comparing CXR with usual care. Figure 18 presents the network of available intervention comparisons for lung cancer mortality.

FIGURE 18.

Network map. The size of nodes is proportional to the number of individuals randomised to each intervention and the thickness of each line is proportional to the number of direct comparisons in trials.

Figure 19 presents the cumulative probability of three screening strategies (LDCT, usual care and CXR) for the lung cancer mortality outcome. The estimated RR of LDCT compared with usual care was 0.95 (95% CI 0.82 to 1.11). LDCT was ranked first according to the estimated surface under the cumulative ranking curve values, with a 74.8% probability of being the best intervention in terms of lung cancer mortality reduction. Usual care had a 74.7% probability of being the second best strategy among the three interventions. However, CXR screening had a 99.7% probability of being the worst intervention in terms of lung cancer mortality reduction.

FIGURE 19.

Rankogram for the outcome of lung cancer mortality. Ranking indicates the probability of being the best treatment, the second best, and so on, among the different interventions under evaluation. Sensitivity analysis includes NLST-eligible subpopulation from PLCO. (a) Usual care; (b) LDCT; and (c) CXR.

Overall, the results of network meta-analysis showed that LDCT screening was ranked as the most effective intervention for the outcome of lung cancer mortality compared with other screening strategies (usual care and CXR).

Both consistency and inconsistency models were fit for lung cancer mortality data. By applying the design-by-treatment model, we did not find any evidence of inconsistency. The global test for inconsistency gives a p-value of 0.29, indicating no evidence of inconsistency.

Sensitivity analysis

Sensitivity analysis was performed by including data of the NLST eligible subgroup from another large trial (PLCO) comparing CXR with usual care. 107 The sensitivity analysis showed similar results to the overall analysis: LDCT was ranked as the most effective strategy and CXR screening was ranked as the worst strategy in terms of lung cancer mortality outcome.

Figure 19 presents the results of sensitivity analysis of cumulative probability of three screening strategies for the lung cancer mortality outcome. The estimated RR of comparing LDCT with usual care from the sensitivity analysis of network meta-analysis was 0.93 (95% CI 0.76 to 1.14). Based on the estimated surface under the cumulative ranking curve values, LDCT screening was ranked first: it had a 75.3% probability of being the best intervention in terms of lung cancer mortality reduction. Usual care had a 68.3% probability of being the second best strategy among the three interventions. Similarly, CXR screening had a 87.7% probability of being the worst intervention in terms of the lung cancer mortality outcome.

For sensitivity analysis, both consistency and inconsistency models were also fit for lung cancer mortality data. By applying the design-by-treatment model, we did not find any evidence of inconsistency. The global test for inconsistency gives a p-value of 0.18, suggesting that there was no evidence of inconsistency.

Trial- and model-based evaluations

Simple data abstraction templates were developed in advance of study selection.

Data were abstracted by one reviewer (TS) based on main publications (i.e. without referring to any supplementary materials).

The quality of studies was assessed using the Consensus on Health Economic Criteria (CHEC)-list for economic evaluations,109 with certain modifications/guidelines for assessment (see Appendix 7).

Only the main publication was examined for quality assessment (i.e. supplementary materials were not checked).

Systematic reviews of economic evaluations

Simple data abstraction templates were developed in advance of study selection.

Data were abstracted by one reviewer (TS) based on main publications (i.e. without referring to any supplementary materials).

No quality assessment was conducted of systematic reviews of economic evaluations.

Trial- and model-based analysis

A common theme in the study results is that LDCT screening is more costly and more effective than no screening. Studies had sharply diverging conclusions about the cost-effectiveness of screening. There is some evidence that studies based on the Early Lung Cancer Action Project (ELCAP) cohort study135 predict improved cost-effectiveness for screening117–119,122,125,127,128 versus studies based on NLST110–113,115,124,127 or lung cancer natural history models.

Three different natural history models have been used to predict the cost-effectiveness of LDCT screening. The Lung Cancer Policy Model was used by McMahon et al. 114 and suggested that LDCT would not be cost-effective. The Cancer Risk Management Model (renamed OncoSim) was used by Goffin et al. 112,113 and suggested that biennial LDCT would be cost-effective. The Microsimulation Screening Analysis (MISCAN) – Lung model was used by ten Haaf et al. 115 and suggested that annual LDCT would be cost-effective.

Many studies identified that the cost of LDCT scans, the effectiveness of screening (identifying early-stage lung cancers) and the prevalence and incidence of lung cancer are key factors affecting the cost-effectiveness of screening. A number of studies considered age and smoking history and found these to be influential also.

Some studies incorporated smoking cessation as an adjunct intervention, but these studies did not generally consider the cost-effectiveness of screening versus smoking cessation, or screening with smoking cessation versus smoking cessation and, therefore, were not appropriate evaluations of lung cancer screening.

Systematic reviews of economic evaluations

Existing systematic reviews found significant variation in methodology and results of economic evaluations. Earlier economic evaluations did not have access to good-quality estimates of clinical effectiveness, and the studies that incorporated the results of NLST produced ICERs of < US$100,000 per QALY. More recent economic evaluations were generally more methodologically robust (see Table 51, Appendix 7).

The reviews concentrated on the cost-effectiveness of screening versus no screening, without any significant attention being paid to the impact of frequency of screening.

Target population and subgroups

The target population was people who are at higher risk of lung cancer relative to the general population. Specifically, the model considered people aged between 55 and 80 years with a history of smoking (i.e. current or former smokers).

Subgroups in this population are identified by further restricting the age range and by imposing a minimum threshold on the predicted risk of lung cancer for an individual to be eligible for screening.

Setting and location

The setting was the NHS in the UK.

Initial invitations to screening may be sent from primary care (as initial identification of people who may be at high risk could be from primary care records). Screening CT scans are performed in secondary and tertiary settings. Cancer care is performed in secondary and tertiary settings. Palliative care may be delivered in secondary settings, in the community or in hospices.

Study perspective

The perspective on costs was NHS and Personal Social Services (PSS). The economic evaluation therefore did not include any effect on tax revenue, pensions, productivity or out-of-pocket expenses for affected individuals (e.g. transport).

Direct health effects on individuals who were contacted through a screening programme were included.

No direct health effects on family or carers were included as no good-quality evidence was found to support their inclusion.

Although some screening studies have shown an impact on the smoking behaviour of participants (e.g. Clark et al. 90), and despite it being very likely that increased quitting reduces the risk of lung cancer mortality, no attempt has been made to model smoking behaviour in the model.

Comparators

Four screening programme designs were compared with no screening in 12 population alternatives, thereby creating 48 intervention strategies and one control (no screening programme) strategy that represents current practice.

Chest X-ray was not considered as a comparator as it is not considered a relevant policy option (the briefing note for this project lists only no screening as a comparator) and because RCT evidence has shown that CXR is not expected to be clinically effective. 107

Discount rate

Costs and QALYs were discounted at 3.5% per year. These are the conventional discount rates for technology appraisal in England,137 Scotland138 and Wales139 (Northern Ireland typically endorses NICE technology appraisals) and are derived from the UK Treasury discount rate. 140

Time horizon

Individuals are modelled until death. Most simulated individuals are dead by 100 years of age (e.g. 98.9% of women simulated from the age of 80 years die before the age of 100 years).

Choice of health outcomes

The primary health outcomes were HRQoL and life-years attained in each strategy, expressed in QALYs, as is the preference in UK cost-effectiveness decision analyses. 137 In the incremental analysis, the outcome was QALYs gained versus no screening.

In addition, secondary health outcomes were:

• screening programme sensitivity

• number of lung cancers diagnosed per 100,000 entrants

• interval cancers diagnosed per 100,000 entrants

• mortality per 100,000 entrants

• 5-year survival from diagnosis of lung cancer

• substage distribution

• average age at diagnosis, death from lung cancer, death from other

• lead time.

Analysis methods

The economic evaluation employed a cost-effectiveness (cost–utility) analysis, in which the costs and QALYs for each alternative (combination of population and intervention strategies) are estimated and then a cost-effectiveness frontier is constructed by eliminating strategies that are dominated or extendedly dominated. The cost-effectiveness of each alternative is then assessed using the incremental cost-effectiveness ratio (ICER), the ratio of incremental costs to incremental QALYs.

• Main analysis: cost-effectiveness analysis of all alternatives, in which the ICER is calculated both versus the most effective alternative on the cost-effectiveness frontier that is less effective than the current option, and versus current practice (no screening).

• Secondary analyses:

  • Cost-effectiveness analyses of intervention strategies for each choice of population strategy (e.g. a cost-effectiveness of four intervention strategies and no screening assuming a minimum age at entry of 60 years, a maximum age at entry of 75 years and a minimum risk of 4%).

  • Net monetary benefit (NMB) maximisation (at willingness to pay of £20,000 per QALY) conducted for each intervention strategy to identify the optimal choice of age limits and predicted risk thresholds.

These analyses are conducted in the base-case analysis (in which all parameters are fixed at their base-case values and a large cohort of patients is simulated).

Sensitivity analyses are also conducted:

• Deterministic sensitivity analysis: main analysis conducted with deterministic changes to parameter values (the results remain stochastic as individual patients continue to be simulated).

• Probabilistic sensitivity analysis (PSA): multiple cohorts simulated, each with a single set of parameter values sampled probabilistically from suitable distributions reflecting parameter uncertainty. A main analysis was conducted and cost-effectiveness acceptability curves were produced.

Software

The model was implemented in Microsoft Excel. Supporting analyses were conducted in Stata 14.2, R 3.3.3 (The R Foundation for Statistical Computing, Vienna, Austria), and JAGS 4.2 (Martyn Plummer, Lyon, France). 141

Overview

A cohort of individuals is simulated with a range of baseline characteristics (including age and predicted risk of lung cancer). The age range of simulated individuals is 55–80 years, as no screening programmes are being evaluated that would include individuals outside this age range.

Each individual is concurrently simulated with four screening intervention arms and a control (no screening) arm. By simulating the individuals concurrently through all arms there is a reduction in stochastic variation.

The costs, QALYs and other outcomes for each full programme (combination of population strategy and intervention) are estimated using a decision tree. Costs of administering the screening programme (sending letters with questionnaires, analysing the questionnaires to estimate the lung cancer risk, sending invitations to those at high risk) are accumulated through the decision tree, and long-term costs and QALYs are estimated at the leaves of the decision tree by identifying appropriate individuals simulated in the cohort and assigning them appropriately either to the screening intervention (if they meet all criteria and join the screening programme) or to no screening.

It was assumed that the two uptake probabilities (relating to a person completing a risk questionnaire and agreeing to join the programme after being informed they are at high risk) are independent of age and sex, although gender has been identified as a potential factor influencing the decision to participate in a lung cancer screening programme (see Chapter 7).

Figure 21 shows how age thresholds and the risk threshold are used to identify individuals at high risk of lung cancer and map them to screening interventions.

FIGURE 21.

Population selection diagram.

Simulating individuals

The DES modelling framework was employed, but without interactions between individuals (i.e. no queues or shared resources). This method of simulation involves sampling times to future events according to the current state of the individual (and any relevant history). The earliest of these events is modelled as occurring and the model ‘clock’ advances to that event. Times to events are then either reduced by the amount the clock has advanced or are resampled (as appropriate). Figure 22 provides a diagram of the model.

FIGURE 22.

Model diagram for simulating individuals. LC, lung cancer.

Individuals began the simulation without clinically diagnosed lung cancer, although they could have occult lung cancer.

A natural history model of lung cancer was utilised to generate outcomes for individuals in the absence of screening (see Natural history). The action of lung cancer screening is to identify preclinical (occult, asymptomatic) lung cancer earlier than it would be clinically diagnosed. The sensitivity of the screening test affects the likelihood that a cancer will be detected at the time of screening.

If a lung cancer is detected by screening in an earlier stage than it would have presented clinically, then the time to lung cancer mortality (i.e. survival) is extended, as survival is related to stage. If a lung cancer is detected by screening in the same stage as it would have presented clinically, the time to lung cancer mortality is extended by the lead time. In any case, the age at lung cancer mortality is modelled as never being earlier when a cancer is screen detected than if the cancer had presented clinically.

In the base case it is assumed that there is no heterogeneity between patients in the rate at which their cancers progress or present, but in a scenario analysis a random effect is included for each simulated individual across their rates of progression, as shown in Equation 1:

In this equation, λ^{k→k + 1}_i is the rate of progression from state k to state k + 1 for individual i. ϵ_i is the error term for between-individual variability in the log-rate of progression, with variance σ²_p. Note that the variance is assumed to be the same for all progression rates.

Note that in the base case there is no lung cancer mortality from the preclinical lung cancer states. This is justified in Natural history. Note also that there is no explicit modelling of cancer progression after diagnosis, as the costs and outcomes are intended to be averaged across lung cancers diagnosed in each stage.

The model does not include incidental findings resulting from screening. It is likely that such findings would lead to increased health-care resource use in the short term (although these could in some cases be offset by savings in the long term), and could lead to improved QoL if the disease has been causing reduced QoL and treatment is effective.

Five screening programme designs were developed following consultation with the expert advisory group. These varied according to the target number of screens, the interval time between screens and the duration of the programme (Table 15). UKLS55 adopted a single-screen design; the NLST and NELSON trials adopted a triple screen design, with increasing screening intervals in NELSON.

Twelve population subgroups combining age at entry (minimum ages 55 and 60 years, maximum ages 75 and 80 years) and predicted risk (minimum risk 3%, 4% and 5%) were specified. Only patients in the starting cohort whose age and risk profiles met the criteria of these populations were included in the respective strategy analyses.

Starting characteristics

Each individual was simulated with the following random starting characteristics:

• age

• sex

• baseline disease state

• risk score.

In the base case, the age distribution of individuals was estimated from UKLS for the participants returning a questionnaire. 55 Truncated normal regression was performed on the age of participants with cut-off points at 50 years and 76 years using the Stata command truncreg. The estimated age [mean and standard deviation (SD)] was 61.94 years (SD 9.00 years). The ages were sampled using a truncated normal distribution but with cut-off points at 55 years and 80 years to reflect the widest range of eligibility criteria.

A scenario analysis was conducted using an age distribution fitted to the approximate age distribution of smokers in the UK143,144 (aged 55–80 years) and using least squares regression to estimate an underlying normal distribution of 61.62 (SD 15.19) years. This scenario predicts a similar mean age but a greater spread.

The sex distribution of individuals was similarly estimated from UKLS participants returning a questionnaire, producing an estimate of 48.2% being men.

The baseline disease state was estimated by sampling the age at preclinical lung cancer incidence and the age at entry to the programme. If the age at preclinical lung cancer incidence was less than the age at entry to the programme, then the stage of preclinical lung cancer was estimated using probabilities inversely proportional to the expected time spent in each stage in the absence of screening. If the age at preclinical lung cancer incidence was greater than the age at entry to the programme, then the individual started with no lung cancer.

The risk score was estimated as detailed in Effectiveness estimates.

Model parameters

Details of model parameters are given throughout this chapter and Appendix 9 provides a full listing of model parameters, their base-case values and their PSA distributions.

Lung cancer survival

Survival from lung cancer was estimated according to the stage of lung cancer at diagnosis. The key data source was the International Association for the Study of Lung Cancer (IASLC) study used to develop the TNM Seventh Edition,24 which provides mature survival estimates based on a significant number of patients. The drawback of this data set is that the patients are primarily not from the UK, although they were drawn from 46 sources across 19 countries.

Kaplan–Meier curves for survival according to clinical stage were extracted and annual values up to 10 years were isolated. A Weibull plot was constructed by plotting cumulative hazard (logarithmic scale) versus time (logarithmic scale). The lines were close to straight and close to parallel and, therefore, a proportional hazards Weibull model was judged to be appropriate. A weighted linear regression was performed on log(cumulative hazard) with log(time) and stage as independent variables. Each point was weighted by the number of patients diagnosed in the stage multiplied by Kaplan–Meier survival (as an approximation of the number of patients contributing to the data point).

The resulting survival curves closely match the extracted survival data, as shown in Figure 24.

FIGURE 24.

Comparison of survival data and Weibull fit. (a) Extracted Kaplan-Meier data; and (b) Weibull fit.

To reduce unnecessary variability between intervention arms and to achieve a consistent improvement in survival when a stage shift is achieved, a single quantile was sampled randomly per lung cancer and used to sample all survival times for the different stages of the lung cancer.

Mortality from undiagnosed lung cancer

The rate of death from preclinical lung cancer is a highly uncertain quantity, as it is believed that most individuals dying from lung cancer will do so with a diagnosis of lung cancer (i.e. they will have clinically presented). This suggests that the rate of death from preclinical lung cancer should be much lower than the rate of clinical presentation, even if this is significantly lower than the mortality rate for clinical lung cancer, or even close to zero. This appears paradoxical, as diagnosing lung cancer should not accelerate mortality, as the aim of treatment is (in many cases) to prolong life expectancy. The paradox is resolved by not assuming that the time to clinical presentation and lung cancer mortality are independent, but in fact that individuals would present shortly before dying from a previously undiagnosed lung cancer.

As our estimate of lung cancer survival should include all patients, even those with very limited survival post diagnosis, it is reasonable to estimate the rate of clinical presentation assuming a very low probability of dying from lung cancer prior to diagnosis.

In the base case it is assumed that there is no hazard of dying from preclinical lung cancer, but as soon as an individual is diagnosed with lung cancer, they then face the hazard of death from lung cancer.

Other cause mortality

Individuals at high risk of lung cancer are likely to be at a higher risk of mortality from other causes as well (e.g. cardiovascular disease, other respiratory disease, other cancers).

We estimated the risk of death from causes other than lung cancer for smokers in the following ways:

• We adjusted the overall risk of death for smoking.

• We estimated and removed the proportion of the mortality rate attributable to lung cancer.

• We fitted a parametric (Gompertz) model to the resulting mortality profile from age 30 years.

The baseline risk of death was taken from the interim life tables for England and Wales based on data for the years 2010–12. 147

The adjustment for the risk of death caused by smoking was taken from the Institute and Faculty of Actuaries ‘00’ tables,148 which allow comparison of the mortality rates in male and female smokers versus males and females generally. The population here is individuals with permanent life assurance policies, which may mean that the data are not wholly representative, but we used the estimates after the ‘select period’ (during which individuals are at a lower risk of mortality because of selection bias), and these estimates were only used to adjust the life tables for England and Wales, which are based on national data. The adjustment was performed by estimating the mortality rate ratio between smokers and the general population for each year of age from 30 years, and then applying this to the England and Wales life tables.

The proportion of mortality caused by lung cancer was estimated using cause of death data from the Office for National Statistics,149 estimating the number of lung cancers attributable to smoking for each age group using population attributable fractions of 86.3% and 72.2% for men and women, respectively,143 using estimates of smoking prevalence (current smokers and former smokers) in each age group143 to estimate the population size in each age group, and finally dividing the number of lung cancers attributable to smoking by the population size of smokers in each age group to obtain a mortality rate from lung cancer in smokers.

The mortality rate from lung cancer in smokers was subtracted from the overall mortality rate in smokers to produce a final estimate of the risk of death from causes other than lung cancer (Figure 25).

FIGURE 25.

Survival rates for smokers, excluding lung cancer as a cause of death.

Graphical inspection of the mortality profiles (log-hazard versus time) demonstrated exponential growth in the mortality rate from age 30 years, suggesting that a Gompertz model would be appropriate. Gompertz models were fitted separately for men and women using least squares regression on log-hazard for ages 30–100 years.

In the model, the age at death from other causes is sampled conditionally on the individual being alive at the start of the simulation as shown in Equation 2, where λ and γ are the Gompertz parameters, U is a uniform random variable between 0 and 1 and t₀ is the age of the individual at the start of the simulation:

Uptake of screening

The probability of an individual responding to an initial letter regarding lung cancer screening with a lung cancer risk questionnaire was estimated from UKLS,55 as this included contacting a significant number of people in two qualitatively different areas and included a risk questionnaire. In this trial, 247,354 individuals were sent invitation letters, of whom 75,958 responded positively with a completed questionnaire. Therefore, it was assumed that 30.7% of individuals would respond to an initial letter and complete a risk questionnaire.

The probability of an individual meeting all criteria for entry to a screening programme (having initially responded and completed a risk questionnaire) was also estimated from UKLS. 55 In this trial, 8729 individuals were classified as high risk, of whom 4061 individuals subsequently gave informed consent to participate in the trial. Therefore, it was assumed that 46.5% of individuals would join a programme if invited.

The model also assumes that once an individual takes up screening, that they are 100% compliant (i.e. there are no missed screens).

Risk prediction

To achieve a favourable balance of benefits, harms and costs, it is necessary to target screening towards individuals at high risk of lung cancer. The ability to accurately discriminate between individuals at low and high risk is a key component of the effectiveness of a screening programme.

In the model it is assumed that the Liverpool Lung Project (version 2) (LLPv2) risk prediction tool150 will be used, as this is the only risk prediction tool that has been used to select a population for a RCT of lung cancer screening (in UKLS55). Other risk prediction tools are available and may have different data requirements and different performance in terms of discriminatory ability. The choice of the LLPv2 tool does not represent an endorsement.

Rather than estimating the performance of LLPv2 from the case–control study in which it was developed150 or the validation studies,151 we instead considered its performance among responders to an invitation to screening, that is, in the UKLS population. 55

In UKLS, 41 lung cancers were detected in the baseline screen out of 1994 individuals scanned. 55 A further nine lung cancers were detected within 3 years of follow-up in these individuals (Professor John K Field and Dr Michael Marcus, University of Liverpool, 2017, personal communication).

However, the lung cancer outcomes are unknown in individuals:

• meeting the risk threshold (4.5%) and being randomised but not receiving a CT scan (n = 2061)

• not being randomised

• not meeting the risk threshold

• not replying to the questionnaire.

It is therefore not possible to directly estimate the sensitivity and specificity of the LLPv2 at any particular threshold, or to construct a receiver operating characteristic curve.

We were provided with data from UKLS on certain baseline characteristics, the risk prediction score and their lung cancer outcome (if known).

We then sought to estimate a statistical model for the risk score according to those baseline characteristics and the lung cancer outcome. As our model simulates individuals for their lifetime through different screening programmes, we are easily able to identify whether or not a simulated individual develops clinical lung cancer within a certain duration of follow-up, and then to ‘back-estimate’ their risk prediction score. In this way, simulated individuals who have a baseline lung cancer or develop clinical lung cancer early in the model will have a higher predicted risk score if the risk prediction tool is effective.

Equation 3 describes the statistical model we sought to fit, which is a linear regression on the logit of the risk score:

where r₁ is the predicted risk for individual i, β^(r)₀ is the intercept term, β^(r)_x are the coefficients for baseline characteristics x_i (age, sex, smoking status), β^(r)_y is the coefficient for the outcome y_i (lung cancer or not) and ϵ_i is an error term with variance σ²_r.

All analyses in this section were conducted using Stata version 14.2.

Fitting this model to only individuals for whom all data are available (i.e. individuals receiving a baseline screen) is expected to produce inefficient and biased estimates of the coefficients, because these individuals are not representative (they have been selected because of their high risk). The results of this regression are given in Table 16 and suggest that increasing age, being male, being a current or former smoker (vs. having never smoked) and developing lung cancer all predict a higher risk score, although the coefficient for smoking status is not statistically significant (as very few individuals received a CT scan who were not smokers, i.e. there is a lack of power).

To account for the missing outcome (lung cancer) for the vast majority of individuals for whom data were available, multiple imputation was employed. The lung cancer outcome was imputed using logistic regression on the age, sex, logit risk score and smoking variables and 50 imputations were performed. The linear regression on logit risk score was then conducted again using the imputed data sets.

Table 16 gives the results of the multiple imputation regression. This shows stronger effects for all predictors, and they are all now statistically significant. The error term also now has increased variance. The central estimate for the lung cancer outcome coefficient is approximately half the root-mean-squared error, so there will be significant overlap in the predicted risks for individuals developing and not developing lung cancer.

Accuracy of low-dose computed tomography

In the model it was assumed that LDCT tests would be imperfect, and that sensitivity and specificity would both be < 100%.

Sensitivity was included in the model as the probability that an individual with preclinical lung cancer would be diagnosed with lung cancer following CT screening. If an individual is not diagnosed with lung cancer following CT screening, their preclinical disease may continue to progress until they present clinically or are diagnosed in a subsequent screening round.

Specificity was included in the model as the probability that an individual without preclinical lung cancer would receive a result that required further testing.

It was assumed that follow-up tests would be perfect (e.g. nobody receiving a false-positive result by LDCT screening would go on to receive treatment).

The sensitivity of LDCT screening was estimated in the calibration exercise described in Natural history as 70.9% in the base case, but 97.3% in the scenario analysis in which heterogeneity was included.

The specificity of LDCT screening was estimated from UKLS to be 62.4%. 55 This is lower than the 78% estimated in the calibration exercise based on NLST and, therefore, could be a conservative estimate, but it was judged that a UK-based estimate would be more relevant, as specificity can be relatively accurately estimated (compared with sensitivity) with reference standards that can be employed in trials.

Impact on survival

Numerous studies have shown that survival is higher for screen-detected cancers than non-screen-detected cancers, including those of the same stage (we have confirmed this for NLST using the data on which the natural history model is calibrated), but there are three significant reasons why these survival estimates would be biased:

• Lead time bias – the screen-detected cancers are detected earlier and, therefore, even if the age at death is unchanged, the duration from date of diagnosis to date of death is extended.

• Length bias – if some cancers are more aggressive than others, these are less likely to be detected by screening, as they spend less time in the preclinical stage before reaching advanced stages and being diagnosed.

• Overdiagnosis (an extreme form of length bias) – some slow-growing cancers may never be clinically relevant for a patient in the absence of screening because the patient dies from another cause.

On this basis, it was decided that the model should assume the same survival for lung cancer in each stage, whether screen detected or clinically presenting, with the following caveat: the age of lung cancer mortality should not be brought forward by screening and, therefore, there is a lower bound on survival of A + B, where A is the expected survival in the later stage (in which the cancer would have presented absent screening) and B is the lead time. This also applies if the lung cancer is screen detected in the same state as it would have clinically presented.

Figure 26 illustrates the approach through three examples: (a) lung cancer is detected significantly earlier and in a substantially earlier stage in the screening arm (shown in green), resulting in a predicted age of lung cancer mortality beyond that in the control arm; (b) lung cancer is detected somewhat earlier and in a somewhat earlier stage in the screening arm – the predicted survival in the earlier stage would lead to an earlier age of mortality and, therefore, survival is extended in the screening arm to match the age of lung cancer mortality in the control arm; and (c) lung cancer is detected slightly earlier but in the same stage in the screening arm – the predicted survival is extended exactly by the lead time to match the age of lung cancer mortality in the control arm.

FIGURE 26.

Illustration of approach to modelling lung cancer survival.

This modelling approach directly translates a stage shift into a reduction in lung cancer mortality, but NLST is the only RCT thus far to report a positive result for this outcome. 70 To explore the possibility that lung cancer mortality may not be reduced (or may only be reduced a small amount), we also performed two scenario analyses. In the first, there was assumed to be no effect on lung cancer mortality from screening (i.e. the predicted age of lung cancer mortality is insensitive to whether or not patients receive screening). In this instance, it is still possible for screening to lead to benefits for patients, as HRQoL is better for earlier-stage detected cancers (see Measurement and valuation of preference-based outcomes). In the second scenario, the potential effect on lung cancer mortality is halved (i.e. if, in the base case, a lung cancer patient would be predicted to live x years longer in the screening arm than in the control arm, in the scenario they are modelled as living x/2 years longer).

Screening

We identified two studies84,153 that reported EQ-5D measures relating to screening for lung cancer.

Mazzone et al. 153 reported on the impact of lung cancer screening by CXR with computer-aided detection versus placebo screening in a RCT. They used US population values elicited using the TTO method. 154 The authors’ key findings were that HRQoL was not affected by study arm (likely because patients were blinded) but that EQ-5D utility dropped significantly following notification that a lung nodule was detected (from 0.940 to 0.877).

Reporting results from the NELSON study, van den Bergh et al. 84 described patient-reported outcomes for individuals in the LDCT screening arm who did not receive a positive result (i.e. received negative or indeterminate results). The authors did not report EQ-5D utility values, but instead VAS scores and a number of anxiety and distress measures. These measures suggested that across all participants there was a worsening in patient-reported measures between giving consent and 1 week prior to CT scan, although this was reversed shortly after receiving their CT scan (prior to receiving results). They also suggest that participants receiving an indeterminate result had worse patient-reported outcomes after receiving their result and before receiving a follow-up scan. The generic QoL SF-12 showed no statistically significant difference over time (between those receiving an indeterminate and negative result) in the physical or mental component scales.

We judged that it was important to include some estimate of the impact of screening on HRQoL as there was some evidence to support it.

We therefore assumed that lung cancer screening itself would be associated with a small, temporary disutility of 0.01 (based on VAS drop from 79.3 to 78.8 in NELSON84) lasting for 2 weeks (i.e. a loss of 0.01 × 2/52 = 0.00038 QALYs). This is probably a negligible loss for a single participant, but considering that in screening very few patients benefit, the average benefit from screening may also be considered ‘negligible’.

We also assumed that receiving a false-positive result (the nearest representation of an indeterminate result in the model) would result in a temporary disutility of 0.063 (based on EQ-5D drop in Mazzone et al. 153) that would last for 3 months (it is anticipated that within 3 months an individual would have had some follow-up to give them reassurance). This corresponds to a loss of 0.063 × 3/12 = 0.0158 QALYs. As this represents a significant loss of utility (greater than the base-case disutility owing to stage IV lung cancer) and because the study is of CXR rather than LDCT, a scenario analysis was conducted in which this disutility is not included at all.

General (smokers)

All individuals included in the economic evaluation are current or former smokers. It is therefore important to note that such individuals are unlikely to be at perfect health, or even at the same average health of the population.

We estimated the impact of smoking (current or previously regular) on EQ-5D utility from the Health Survey for England 2014: Health, Social Care and Lifestyles. Summary of Key Findings,155 controlling for sex and age. Linear regression was conducted with appropriate weighting and stratification based on the survey design.

The effect of smoking was estimated to be –0.048. Men had slightly higher utility values (+0.029) and an age profile was observed in which utilities generally declined with age.

For the baseline age category (75–84 years), the estimated utility for female smokers was 0.753 and for male smokers was 0.782. These utility values were used for women and men in the model regardless of the current smoking status (which was not modelled) and age. It would have been possible to model utility as a function of age, but it was judged that this additional complexity would not greatly affect results and, therefore, it was decided to focus on other aspects of the economic evaluation.

Lung cancer

The ideal study to inform utility values relating to lung cancer would include lung cancer patients as well as matched patients without lung cancer (matched on at least age, sex and smoking history) and would estimate the effect of the stage of lung cancer as well as the effect of different treatments and time-dependent effects (e.g. time before death). In addition, EQ-5D utilities measured in patients and valued by a representative sample of the UK population using the TTO method would be preferred to be in line with the NICE reference case,137 which is used for the vast majority of economic evaluations of health technologies in the UK.

Unfortunately, no such study was found in our review of the literature.

It was decided, therefore, to focus on studies that gave evidence of the effect of lung cancer stage on utility values.

Only one such study, by Chouaid et al. ,156 explicitly measured HRQoL in UK patients (among patients from other countries) using EQ-5D and valued using a UK population TTO tariff. This study included 255 NSCLC patients with stage IIIB or stage IV lung cancer and produced utility estimates of 0.77 and 0.70 for these stages, respectively.

Another study, by Grutters et al. ,157 used the UK population TTO tariff but measured HRQoL in 245 NSCLC Dutch patients. Only two patients with stage IV lung cancer were included. The study estimated utility values of 0.77, 0.74, 0.70 and 0.86 for stages I, II, III and IV, respectively.

Three studies158–160 utilised a TTO tariff elicited from a US population sample. 154 Jang et al. 158 measured HRQoL in 172 NSCLC patients in Canada and produced utility estimates of 0.80, 0.78, 0.73 and 0.75 for stages I to IV, respectively. Yang et al. 160 measured HRQoL in 518 NSCLC patients in Taiwan and produced utility estimates of 0.85, 0.83 and 0.83 for operable stage I, II and III, respectively, and 0.72 and 0.75 for inoperable stage III and IV NSCLC. Tramontano et al. 159 measured HRQoL in 2396 lung cancer patients in the US and produced utility estimates of 0.81, 0.77, 0.77 and 0.76 for stages I to IV, respectively. This study also used the Short Form questionnaire-6 Dimensions (SF-6D) and UK population value set161 (derived using standard gamble) and estimated utility values of 0.71, 0.68, 0.67 and 0.66 for stages I to IV, respectively. In addition to differences in the population valuing the health states (USA vs. UK) and the valuation method (TTO vs. standard gamble), the EQ-5D asked patients to rate their health today, whereas the SF-6D asked patients to recall their health for the previous 4 weeks.

Of all these studies, the study by Tramontano et al. 159 is by far the largest and is the only study to not restrict to NSCLC. As such, we believe this is the best study with which to estimate lung cancer utility values according to stage, despite it using a US value set rather than a UK value set.

All of the studies identified will have been at some risk of bias for a number of reasons. First, as lung cancer can be associated with particularly poor HRQoL, it is possible that patients with worse HRQoL are underrepresented as they will be less likely to participate in studies or to be able to complete health questionnaires. Second, most of these estimates do not account for other differences between patients besides their cancer stage. It may be that late-stage cancer is associated with other factors that affect HRQoL, such as age, sex, income and other respiratory conditions.

It is also possible that generic health measures (such as EQ-5D and SF-6D) are not sensitive to aspects of lung cancer that negatively affect HRQoL, such as shortness of breath and fatigue.

It is notable that the identified studies did not produce a large difference in the utility of stage IV lung cancer versus earlier stages, the largest difference (0.07) being measured by Chouaid et al. 156 for stage IV versus stage IIIB NSCLC. This contrasts with a systematic review by Sturza that aimed to estimate utility values associated with lung cancer. 162 Using meta-regression across a large number of studies using a variety of HRQoL measures and valuation methods, Sturza estimated a difference in utility value of 0.25 between metastatic and non-metastatic lung cancer. There are a number of ways in which this estimate may be inflated, for example because it relies on meta-regression rather than investigation within a study population, because it includes estimates not derived using EQ-5D and because it includes multiple methods of valuation.

In the base-case analysis, it was decided that the utility values from Tramontano et al. 159 would be used to estimate the disutility for later stages of lung cancer versus stage I, whereas in a scenario analysis, a disutility of 0.252 for stage IV would be applied based on the estimate by Sturza. 162

It was further assumed that individuals with stage I cancers (mostly asymptomatic) would have the same utility as smokers without lung cancer. This was judged to be a pragmatic approach as the average utility for smokers without lung cancer estimated above is lower than the utility estimated for stage I lung cancer, and it would lack face validity to increase the utility for individuals with lung cancer versus individuals without lung cancer.

The utility values are applied for the remainder of the lung cancer patient’s life according to their stage at diagnosis. This is a simplifying assumption as lung cancer often progresses despite treatment, and HRQoL is likely to decline as patients approach death.

It was also assumed that individuals with preclinical lung cancer would experience some disutility as a result of lung cancer symptoms. The same utility values for clinical lung cancer of a particular stage were applied to the preclinical lung cancer stage also (i.e. it is assumed that a diagnosis of lung cancer does not intrinsically affect HRQoL).

Screening programme costs

Screening programme costs included programme administration around participation as well as LDCT examinations for subsequent joiners (Table 18). Administration comprised the postal invitation to self-assess to all potentially high-risk candidates, the scoring of responder questionnaires and a subsequent follow-up letter of invitation or decline. These costs were allocated appropriately using the decision tree described in Model structure and probability of uptake described in Effectiveness estimates. Since the probability of uptake (a product of the probability of responding and then the probability of joining the programme) in the base case was based on a trial population and may not accurately predict real-world uptake, scenario analyses tested lower and high rates of uptake, as described in Deterministic sensitivity analyses.

The LDCT screening examinations were the same unit cost whether they were the first or subsequent examination, and were assumed to be of the kind directly accessed in NHS secondary and tertiary care settings (currency code RD20A). 163 In scenario analyses, the base-case unit cost of LDCT was varied up and down according to the upper and lower quartile values of the HRG currency code distribution. As it could be presumed direct access would be preceded by clinical consultation, as is not the case via invitation, this was added as an associated cost equating to 15 minutes with a band 5 hospital nurse. 169

Hospital costs

Following a referral, secondary care is the setting for most of the NHS expenditure on people with lung cancer (83% in 2012/13). 164 In order to incorporate the burden of referral and symptom management on primary care, an average of two GP consultations was assumed per clinical presentation (unit cost £36169) and included in this category. The cost of symptomatic referrals suspicious for lung cancer that are subsequently found to be negative for lung cancer was not included.

The rate of resource consumption for the hospital-based aspects of the care pathway, namely the diagnosis, treatment and clinical follow-up, was based on a retrospective 1-year cohort study of all emergency, inpatient and outpatient costs (not palliative) from the records of 3274 lung cancer patients between January 2008 and October 2013. 164 This study was limited to the experience of a single English teaching hospital, so may not reflect nationwide variation in disease management, but was not limited to NSCLC as was the study by McGuire et al. 165 Summary costs at 90 days were chart extracted by substage at diagnosis and attributed to all true diagnoses, whether clinical presentations or true-positive screen detections, irrespective of survival time with lung cancer as many of these costs are quickly accrued (more than half of first-year costs come in the first 90 days). People surviving beyond 90 days were attributed further costs in the first year proportionate to their survival up to a limit of 2 years. Second-year costs were adjusted downwards from index year costs using the rate of change observed between year 1 and year 2 in a separate retrospective cohort study of NHS lung cancer resource consumption. 165 Two years was judged a reasonable cut-off point for disease costs in the base case, given the front-loaded nature of resourcing following a lung cancer diagnosis. McGuire et al. 165 estimated that second-year costs were just 13% of first-year costs. However, in a scenario analysis the subsequent year costs were maintained at a flat rate for survivors up to 5 years.

A summary of hospital costs by substage of cancer and period post presentation is given in Table 19.

People who received a false-negative screen were zero cost until a true diagnosis of lung cancer, whereupon costs accrued as described above. Those who received a false-positive screen were resourced as observed in UKLS. 55 Of 951 false-positive cases, there were 72 resultant cancer MDTs, 1466 outpatient CT scans, seven needle biopsies, 13 PET scans, one endobronchial ultrasound bronchoscopy and 21 follow-up outpatient consultations. The cost of investigative resourcing following a false-negative screen was £184.63. A breakdown of the unit costs and their consumption is given in Table 20.

End-of-life costs

In the literature search, we identified a recent and directly applicable study of the cost of caring for people with cancer in England and Wales. 166 In this study, the average per-patient cost of palliative resourcing over an average period of 180 days from the start of strong opioid treatment was £4589, based on 68,340 hospital episodes relating to lung cancer. The personalised social services component of this figure (£1380) was included in the base case but excluded in a scenario analysis in which the payer perspective was limited to health care. In a further scenario analysis, the end-of-life cost attributed to lung cancer mortality was excluded altogether in order to reflect the possibility that other deaths in this population could be equally costly owing to the preponderance of comorbidity in people with a history of smoking.

Main analysis

Cost-effectiveness

Four of the modelled screening strategies were on the cost-effectiveness frontier (i.e. strategies that can give the maximum NMB for at least one choice of the cost-effectiveness threshold) and ‘no screening’ was also on the frontier (the least costly and least effective option).

Table 23 gives key results for strategies on the cost-effectiveness frontier. In this analysis, none of the screening strategies would be considered cost-effective versus no screening at a threshold of £20,000 per QALY. At a threshold of £30,000 per QALY, S–60–75–3% would be cost-effective versus no screening (ICER £28,169 per QALY), as would S–55–75–3% (ICER £28,784 per QALY), but in a fully incremental analysis only S–60–75–3% would be cost-effective with a threshold of £30,000 per QALY.

TABLE 23 - Base-case cost-effectiveness results

Only strategies on the cost-effectiveness frontier are shown. All strategies were predicted to lead to health benefits (vs. no screening), ranging from 0.0003 to 0.0012 QALYs per person. Although such gains would not generally be considered significant, these gains are concentrated in people who join the screening programme (ranging from 1.2% to 4.0% of the population), and are diagnosed with lung cancer at an earlier stage and, therefore, receive more substantial health benefits. For example, individuals participating in the S–60–75–3% screening programme are predicted to gain an average 0.054 life-years (≈3 weeks)/0.027 discounted QALYs compared with no screening, and to die from lung cancer 0.16 years (≈8 weeks) later.

Strategy	Costs (£)	QALYs	ICER (vs. no screening) (£)	Incremental costs (vs. previous) (£)	Incremental QALYs (vs. previous)	ICER (vs. previous) (£)
No screening	1103	8.502
S–60–75–3%	1126	8.503	28,169	23	0.0008	28,169
S–55–75–3%	1129	8.503	28,784	3	0.0001	35,453
S–55–80–3%	1135	8.503	30,821	6	0.0001	44,087
T–55–80–3%	1151	8.503	40,034	17	0.0002	95,292

Figure 27 presents the cost-effectiveness plane with all strategies, and Figure 28 shows the strategies on the cost-effectiveness frontier.

FIGURE 27.

Cost-effectiveness plane for base-case results.

FIGURE 28.

Cost-effectiveness frontier for base-case results.

A summary of selected clinical outcomes is presented for the screening strategies on the cost-effectiveness frontier in Table 24.

TABLE 24 - Clinical outcomes for participants of strategies on the cost-effectiveness frontier

Screening programme strategy	Strategy
	S–60–75–3%	S–55–75–3%	S–55–80–3%	T–55–80–3%
Per participant
Number of screens	1.00	1.00	1.00	2.70
Number of false positives	0.33	0.33	0.33	0.95
Lead time (years)	0.299	0.299	0.295	0.395
Life-years gained	0.0537	0.0568	0.0524	0.0762
Additional lung cancer survival (%)	0.80	0.45	0.43	0.64
Additional 5-year lung cancer survival (%)	16.1	16.4	16.1	21.0
Additional survival time with lung cancer (years)	1.87	1.89	1.85	2.44
Change in age at lung cancer diagnosis	–1.70	–1.69	–1.62	–2.03
Change in age at death from any cause	0.05	0.06	0.05	0.08
Change in age at death from lung cancer	0.16	0.20	0.23	0.41
Per 100,000 participants
Proportion of diagnoses arising from screening (%)	44.4	44.3	47.1	62.5
Number of screen-detected cases	1710	1785	2335	3185
Number of interval cancers	0	0	0	215
Additional lung cancer diagnoses	295	300	450	590
Lung cancer deaths averted	170	100	120	180
Life-years gained	0.2683	0.2839	0.2621	0.3809

Lung cancer mortality reduction

The average number of lung cancer deaths (per 100,000 participants) was 15,200, 15,100, 14,600 and 15,000 for the single, triple, annual and biennial strategies, respectively, with a comparable 15,800 for no screening.

Across the different screening programmes a reduction in lung cancer mortality of 2.9% to 8.7% was predicted (RR) among the participating individuals versus no screening. The results for the strategies on the cost-effectiveness frontier are shown in Table 25.

The average lung cancer mortality reduction for single-screen strategies was 4.2%, while for triple screen strategies it was 4.4%, for annual strategies it was 7.7% and for biennial strategies it was 5.2%.

Lung cancer stage and survival

Screening strategies were associated with an increased probability of lung cancer being diagnosed in the early stages (I and II) versus later stages (III and IV). The average ORs of early diagnosis (geometric mean) were predicted to be 2.44, 3.29, 5.62 and 3.83 for single, triple, annual and biennial screening programmes, respectively.

Table 26 presents the average stage distributions for screening programmes by frequency of screening. As can be seen, the most significant impact is seen in the increase in lung cancers detected at stage IA and the decrease in lung cancers detected at stage IV.

As would be expected, lung cancer survival was predicted to be higher in the screening arms. Lung cancer survival at 5 years was predicted to be 20.3%, 26.2%, 32.3% and 29.1% for single, triple, annual and biennial screening programmes (on average) versus a comparable average of 4.7% for no screening.

Lung cancer diagnoses

Lung cancer screening programmes led to increased lung cancer diagnoses across the lifetime of participants (i.e. what would be considered overdiagnosis) versus no screening. The average RRs of a lung cancer diagnosis were 1.11, 1.15, 1.20 and 1.18 for single, triple, annual and biennial screening programmes (geometric mean), respectively.

Per 100,000 participants, there were on average 19,200 lung cancers diagnosed in the single-screening arms, 19,700 in the triple screening arms, 20,600 in the annual screening arms and 20,300 in the biennial screening arms. The comparable figure for no screening was 17,200.

On average, 47.7% of lung cancer diagnoses in the single-screen programmes were screen detected, compared with 64.2%, 80.6% and 72.1% for triple, annual and biennial strategies, respectively (these could be considered the screening programme sensitivities). Interval cancers accounted for 3.9%, 5.6% and 11.8% of diagnoses in the triple, annual and biennial strategies, respectively.

Number of screening tests and false positives

Screening programmes were associated with an average of 1.00, 2.68, 8.03 and 4.55 LDCT screens per participant for single, triple, annual and biennial screening programmes, respectively, and with 0.32, 0.93, 2.96 and 1.60 false-positive or indeterminate results.

Average ages at events

The average age at diagnosis of lung cancer was lower in the screening arms (which would be expected unless there was significant overdiagnosis in older participants). The average ages at diagnosis were 74.6, 74.1, 73.6 and 73.9 years for single, triple, annual and biennial programmes, respectively, versus a comparable 76.2 years in the absence of screening.

The average age at death from lung cancer was higher in the screening arms. The average ages at death from lung cancer were 77.6, 77.8, 78.0 and 77.9 years for single, triple, annual and biennial programmes, respectively, versus a comparable average of 77.5 years for no screening.

The average age at death from other causes was not significantly affected (around 82 years), but was slightly higher in the screening arms. The only explanation for this in the model is that some lung cancer patients were dying from other causes in the screening arms, whereas they died from lung cancer in the no-screening arm, and that these patients were on average older at time of death than the people already dying from other causes.

Lead time is calculated in the model as the difference between the age at which an individual is diagnosed with lung cancer in the no-screening arm (or dies from other causes, whichever is earlier) and the age at which the individual is diagnosed with lung cancer in the screening arm. Lead time is therefore time spent by the individual with a known diagnosis of lung cancer that they would not have had in the absence of screening. The average lead time in the single-screening arms was 0.32 years, whereas it was 0.44 years in the triple screening arms. The average lead time in annual screening arms was 0.58 years, whereas it was 0.20 years in the biennial screening arms.

Costs

The costs per participant relating to LDCT screening ranged from £104 (single-screen programmes) to £603–794 (annual screening programmes).

Lung cancer costs (excluding end of life) also generally rose in line with the frequency of screening. If there are savings in the cost of treating some screen-detected cancers because they were detected at an earlier stage, these are outweighed by the increased number of lung cancers diagnosed (i.e. overdiagnosis).

The costs of end-of-life care are decreased as the frequency of screening increases, because there is a reduction in the number of people dying of lung cancer (the model assumes end-of-life costs only for individuals dying of lung cancer, not for those dying of other causes with lung cancer).

The costs for the screening programmes on the cost-effectiveness frontier are shown in Table 27. The programmes are predicted to lead to population lifetime cost increases of £299M to £634M for a relevant population of 13 million smokers aged 55–80 years. The costs of running the screening programme (invitations, risk scoring, LDCT scans) make up less than half of the increased cost, with the rest being attributable to increased costs associated with lung cancer (excluding end-of-life care).

Secondary analyses

Univariate sensitivity analyses

Univariate sensitivity analyses were conducted by rerunning the base case with a single parameter increased or decreased by 10% of its base-case value or by log(1.1) (≈0.095) depending on whether or not it could change sign and/or was estimated on a logarithmic scale.

Each run of the model simulated 6000 individuals and, therefore, it is possible that stability was not reached for strategy mean costs and QALYs; however, the results should still be indicative of the likely direction and magnitude of the impact on cost-effectiveness from changing each parameter.

The impact on cost-effectiveness was assessed by evaluating the INMB of the strategy S–60–75–3% (the optimal screening strategy in the base case, although not cost-effective) versus no screening (at £20,000 per QALY) and comparing it to the base-case value (–£7).

When the INMB is > £0, it indicates that S–60–75–3% is cost-effective with no screening at £20,000 per QALY.

The results of the univariate sensitivity analysis are presented in a tornado diagram in Figure 29.

FIGURE 29.

Tornado diagram for univariate sensitivity analyses. Parameters sorted by range of INMB produced (including the base-case value). See Appendix 9 for key to parameter names.

Four out of the five most influential parameters relate to the natural history for smokers (lung cancer survival, other cause mortality, preclinical lung cancer incidence). The cost of LDCT screens (c_LDCT; see Appendix 9 for parameter labels) is fairly influential and, as would be expected, screening is expected to be more cost-effective when the cost is lower.

There appears to be some asymmetry in the tornado diagram, which may suggest non-linearity in a number of the parameters. For some parameters, the resulting range of INMB does not include the base case. This may be due to Monte Carlo variation, or may be a result of non-linearities.

The specificity of screening (sens_LDCT) appears to be more influential than the sensitivity (spec_LDCT), but in both cases improved diagnostic performance leads to better cost-effectiveness. Likewise, the performance of the risk prediction (risk_lungcancer) positively affects cost-effectiveness, as would be expected.

It should be noted that parameters were all varied by approximately the same degree, regardless of how precisely they were estimated, and that no correlation between parameters has been incorporated. These issues are both addressed within the PSA (see Probabilistic sensitivity analysis).

Scenario analyses

A number of scenario analyses were conducted, in which changes to the structure or sets of parameter values were made. For each scenario analysis 10,000 individuals were simulated and the impact of the scenario analysis is assessed by presenting the INMB of S–60–75–3% versus no screening, as well as for up to two alternative screening strategies: (1) the strategy giving the highest INMB versus no screening of all screening strategies, and (2) the strategy on the cost-effectiveness frontier giving the highest INMB versus no screening of all screening strategies.

The results of these scenario analyses are presented in Table 30 and discussed in detail below.

Very few scenario analyses led to any screening strategy being predicted to be cost-effective versus no screening at a threshold of £20,000 per QALY (this happened only if false-positive and indeterminate results were predicted to have no effect on HRQoL, or if there was no discounting).

Age distribution

In this scenario the age distribution of responders was presumed to match the age distribution of smokers in the UK population, as described in Model structure.

In this scenario, S–60–75–3% was extendedly dominated, and there were two screening strategies on the cost-effectiveness frontier: S–60–80–3% (ICER £36,526 per QALY) and T–60–80–3% (£69,956 per QALY).

Risk-prediction accuracy

The accuracy of risk prediction was increased by changing the risk_lungcancer parameter in the risk model.

In this scenario, S–60–75–3% was dominated and there were three screening strategies on the cost-effectiveness frontier: T–60–75–5% (£25,056 per QALY), T–60–75–3% (£51,077 per QALY) and T–55–75–3% (£13M per QALY).

Programme uptake

In one scenario analysis the uptake of the screening was halved. In this scenario, S–60–75–3% was dominated and there were three screening strategies on the cost-effectiveness frontier: S–60–80–3% (£41,040 per QALY), T–60–80–3% (£97,030 per QALY) and B–60–80–3% (£263,700 per QALY).

In another scenario analysis the uptake of the screening was increased. In this scenario, S–60–75–3% was dominated and there were three screening strategies on the cost-effectiveness frontier: T–60–75–3% (£49,409 per QALY), T–60–80–3% (£84,823 per QALY) and T–55–80–3% (£382,200 per QALY).

Incorporating heterogeneity in lung cancer progression

In this scenario analysis, the natural history model was recalibrated assuming heterogeneity between patients in the rate of lung cancer progression. This also affected the estimated sensitivity of LDCT screening (increasing it substantially).

In this scenario, S–60–75–3% was dominated and there was one screening strategy on the cost-effectiveness frontier: S–60–80–5% (£167,136 per QALY).

Impact on mortality

In one scenario analysis, the survival benefit from early detection was eliminated (i.e. the survival is extended only by the lead time because of screening). In this scenario all screening strategies were dominated by no screening.

In another scenario analysis, the survival benefit from early detection was halved. In this scenario, S–60–75–3% was dominated and there were two strategies on the cost-effectiveness frontier: S–60–75–5% (£74,157 per QALY) and T–60–75–5% (£121,200 per QALY).

Impact on health-related quality of life

In the first of these scenario analyses, it was assumed that there would be a short period of disutility following a lung cancer diagnosis. In this scenario, S–60–75–3% was dominated and there were four screening strategies on the cost-effectiveness frontier: S–60–75–5% (£22,190 per QALY), S–60–75–4% (£50,776 per QALY), S–55–75–3% (£58,040 per QALY) and S–55–80–3% (£113,800 per QALY).

In the second of these scenario analyses, a substantial disutility was assumed for stage IV lung cancer. In this scenario, five screening strategies were on the cost-effectiveness frontier: S–60–75–3% (£22,496 per QALY), S–60–80–3% (£34,254 per QALY), A–60–75–3% (£66,781 per QALY), A–60–80–3% (£93,480 per QALY) and A–55–80–3% (£120,200 per QALY).

In the third of these scenario analyses, it was assumed that there would be no impact on HRQoL in the run-up to subsequent screens (as participants would have ‘acclimatised’ to screening). In this scenario analysis, there were three screening strategies on the cost-effectiveness frontier: S–60–75–3% (£36,680 per QALY), S–60–80–3% (£44,784 per QALY) and S–55–80–3% (£1M per QALY).

In the fourth of these scenario analyses, it was assumed that there would be no impact on HRQoL as a result of false-positive or indeterminate results. In this scenario analysis there were four screening strategies on the cost-effectiveness frontier: S–60–75–3% (£16,759 per QALY), B–55–75–3% (£33,577 per QALY), A–55–75–3% (£63,262 per QALY) and A-55–80–3% (£105,800 per QALY).

Cost of follow-up care

In this scenario analysis, follow-up costs for lung cancer were included for up to 5 years (compared with 2 years in the base case). S–60–75–3% was extendedly dominated and there were three screening strategies on the cost-effectiveness frontier: S–60–75–5% (£31,086 per QALY), S–60–80–3% (£39,648 per QALY) and S–55–80–3% (£67,566 per QALY).

End-of-life costs

In the first of these scenario analyses, PSS costs were excluded from the cost of the end of life (reducing the cost from £4589 to £3209). S–60–75–3% was extendedly dominated and there were two screening strategies on the cost-effectiveness frontier: S–60–80–3% (£35,003 per QALY) and T–60–80–3% (£76,903 per QALY).

In the second of these scenario analyses, end-of-life costs were eliminated completely (which gives an approximation to the case in which end-of-life costs are included for other cause mortality at the same cost as for lung cancer mortality). In this scenario analysis S–60–75–3% was extendedly dominated and there were five screening strategies on the cost-effectiveness frontier: S–60–80–4% (£31,014 per QALY), S–55–80–4% (£31,485 per QALY), S–55–80–3% (£37,778 per QALY), T–55–80–4% (£244,500 per QALY) and T–55–80–3% (£751,000 per QALY).

Computed tomography screening costs

In the first of these scenario analyses, the cost of a LDCT scan was taken from the lower quartile cost across trusts in the NHS reference costs (£70 vs. £99 in the base case). 163 In this scenario, there were two screening strategies on the cost-effectiveness frontier: S–60–75–3% (£39,303 per QALY) and S–60–80–3% (£74,212 per QALY). These results appear anomalous, as it would be expected that cost-effectiveness would be improved by reducing the cost of LDCT (and this was confirmed in the univariate sensitivity analysis). The scenario analysis has not been reconducted as this would potentially introduce bias into the results.

In the second of these scenario analyses, the cost of a LDCT scan was taken from the upper quartile cost across trusts (£121 vs. £99 in the base case). In this scenario, S–60–75–3% was extendedly dominated and there were four screening strategies on the cost-effectiveness frontier: S–60–75–4% (£31,607 per QALY), S–60–80–4% (£45,210 per QALY), T–60–80–4% (£61,927 per QALY) and T–60–80–3% (£230,700 per QALY).

Time horizon

When a time horizon of 10 years was used (compared with a lifetime time horizon in the base case), there were five screening strategies on the cost-effectiveness frontier: S–60–75–5% (£31,290 per QALY), S–60–75–3% (£33,307 per QALY), S–60–80–3% (£45,075 per QALY), B–60–80–3% (£143,900 per QALY) and B–55–80–3% (£383,100 per QALY).

Discount rate

When costs and QALYs are not discounted, S–60–75–3% is dominated and there are five strategies on the cost-effectiveness frontier: T–60–75–5% (£15,001 per QALY), T–60–75–4% (£15,160 per QALY), T–60–75–3% (£63,533 per QALY), B–60–75–3% (£254,200 per QALY) and B–55–75–3% (£477,400 per QALY).

Systematic review

Lung cancer screening programmes are predicted to lead to health benefits for participants compared with no screening, but also increased costs. Study estimates of cost-effectiveness differed substantially, and there are a number of key parameters in existing studies that may not be generalisable to the UK setting.

Independent economic evaluation

Forty-eight lung cancer screening strategies were considered, each representing a unique combination of frequency (single, triple, annual or biennial), minimum age for entry (55 or 60 years), maximum age for entry (75 or 80 years) and threshold for predicted risk (3%, 4% or 5%).

Lung cancer screening programmes are predicted to lead to health benefits for participants compared with no screening (including potential anxiety impacts as a result of screening), but they are also predicted to lead to increased costs (including increased costs associated with lung cancer).

Participants in lung cancer screening are estimated to have a reduction in lung cancer mortality of 4.2–7.7% (depending on the frequency of screening), but also to receive more lung cancer diagnoses than in the absence of screening (i.e. lung cancer is overdiagnosed).

It is expected that 1.2–4.0% of the approximately 13 million smokers aged 55–80 years would actually undergo screening (depending on the population criteria), as a significant number would not respond to an invitation to screening, and another significant proportion would not meet the threshold for lung cancer risk.

Lung cancer screening with a single LDCT scan for smokers aged 60–75 years with a ≥ 3% risk of lung cancer is predicted to be cost-effective at a threshold of £30,000 per QALY (the upper end of thresholds commonly considered in the UK for non-ultra-orphan, non-end-of-life treatments), but not at a threshold of £20,000 per QALY (the lower end of thresholds commonly considered in the UK). This screening programme was estimated to cost £28,000 per QALY gained. However, when the probable range of inputs is considered in a PSA, there were no screening strategies that were cost-effective at thresholds of £20,000 or £30,000 per QALY.

Annual and biennial screening programmes are not estimated to be cost-effective at any threshold, and this appears to be mostly as a consequence of the presumed detrimental impact of false-positive and indeterminate results on HRQoL.