Technical report. Modelling Cost-Effectiveness of Different Strategies for Preoperative Staging of Lung Cancer. December 2010

Transcription

1 CAST - Centre for Applied Health Services Research and Technology Assessment University of Southern Denmark J. B. Winsløws Vej 9B, 1st. floor DK-5000 Odense C Phone: Fax: Technical report Modelling Cost-Effectiveness of Different Strategies for Preoperative Staging of Lung Cancer December 2010 Rikke Søgaard, MSc, MPH, PhD CAST, University of Southern Denmark Barbara Malene Bjerregaard Fisher, MD, PhD Hvidovre Hospital, Denmark

2 Colophon Title Modelling Cost-Effectiveness of Different Strategies for Preoperative Staging of Lung Cancer Authors Rikke Søgaard and Barbara Malene Bjerregaard Fischer Institute CAST Centre for Applied Health Services Research and Technology Assessment Publisher University of Southern Denmark Date of Release December 2010 Printed at University of Southern Denmark, Print and Sign ISBN no

3 Table of Contents Introduction... 1 The disease... 1 The technology... 1 Health economic perspective... 1 Objective... 2 Model structure... 3 Complexity of staging... 3 The choice of a modelling approach... 3 Relevant comparators... 4 Structural assumptions General methodology for cost effectiveness evaluation Populating the model Probability revision Study population True staging distribution Probability for postoperative mortality Diagnostic accuracy of tests Mediastinoscopy EUS-FNA EBUS-FNA CT PET PET-CT Outcome parameters Survival Quality-adjusted life years Cost parameters Specification of probability distributions Model validation Internal validation External validation Prospective validation Acknowledgements References... 32

4 Introduction The disease Lung cancer is the leading cause of cancer related deaths in the western world with non-small cell lung cancer (NSCLC) accounting for around 75-80% of cases and small cell lung cancer accounting for the rest. Once lung cancer is detected, it is important to ascertain the spread of the disease to guide the optimal mix of treatment modalities. Assuming there are no signs of distant metastases diagnostics focus on whether or not there is involvement of the mediastinum (mediastinal lymph nodes). In cases with involvement the current clinical guidelines suggests non-surgical treatment of, for example, chemotherapy, radiotherapy and/or palliative care. In cases with no involvement of the mediastinum the recommendations are to attempt surgical resection of the tumour. In case of inaccurate diagnostics some patients will be referred to surgical resection, which proves to be futile due to the disease having progressed to an extent where resection is not possible. The technology When a patient is suspected to have NSCLC a series of clinical, histological and/or cytological examinations are usually carried out along with a Computed Tomography (CT) scan to establish the diagnosis. To assess the spread of the disease these are often supplemented with a positron emission tomography (PET) scan. PET is a nuclear medicine imaging technique, which produces a three-dimensional image or picture of functional processes in the body. When combined with a radiopharmaceutical it can form images of the distribution of that pharmaceutical around the body to identify for example cancer activity. However as a stand-alone technology it does not provide anatomical details for which reason the rationale of the CT remains. The latest generation of scanners combines the nuclear imaging benefits of PET with the anatomical accuracy of CT in a single procedure with the potential benefits of improved diagnostic performance and a faster track for staging. Two recent randomised controlled trials have demonstrated the superiority of the combined technology over earlier generations of scanners in terms of fewer futile surgeries. 1, 2 Health economic perspective The cost effectiveness of the PET technology for staging NSCLC has been examined in several studies during the past 15 years There is however some variation across studies in terms of the characteristics of the study population, the choice of comparator and the outcome measure adapted. A model for the cost effectiveness of different staging strategies in the setting of England and Wales has been established for patients with normal-sized lymph nodes on CT. 12 The study demonstrated that using PET to select candidates for mediastinoscopy resulted in cost savings and higher gains of quality adjusted life years (QALY) as compared with not using PET. Whether this result holds for a population with nodal involvement on CT has been examined in several models. In a German model the cost per life year (LY) gained was estimated at 143 in CT-negative patients and at in 1

5 CT-positive patients, which corresponds to a factor 78 increase. 4 More moderate ratios between subgroups with and without nodal involvement on CT have been found in a Scottish model ( versus 10475) and in an Australian model (A$ versus A$14581). 11, 13 But still, it should be noted that these values may approach or even exceed decision-makers threshold for willingness to pay. The choice of comparator appears to vary across studies, which may be justified if clinical recommendations vary across settings. Most of the models include multiple comparators from simple regimens of sending all patients directly to surgery through regimens that distinguish between whether there is concordance between CT- and PET-findings. It seems that the question of whether PET should be confirmed by mediastinoscopy impacts the cost effectiveness significantly. For example in the German model, the cost per life year of adding PET to conventional work up was estimated at 143 whereas the cost of adding a confirmatory mediastinoscopy to the PET regimen was estimated at In the Scottish model, on the other hand, it was concluded that selective confirmatory mediastinoscopy (if PET demonstrated nodal involvement) was both cheaper and more effective than no confirmatory mediastinoscopy. 11 In terms of outcome parameter economic evaluations conducted alongside clinical trials usually have no other options than adapting the clinical parameter. At least two such studies, using the number of futile thoracotomies as the primary outcome parameter, have been reported. A Dutch study compared PET with conventional work up and demonstrated a 50% reduction in the number of futile thoracotomies and at the same time a cost saving at on average 1289 per patient. 5 An analogous Australian study in contrast concluded that PET lead to increased costs and no reduction in the number of futile thoracotomies. 10 Overall, it seems beyond doubt that PET is a cost effective technology in CT-negative patients whereas it is more uncertain for CT-positive patients. It is also beyond doubt that cost effectiveness depends on the chosen comparator and the setting for which the analysis is conducted due to cultural and structural differences between countries. To conclude on the attractiveness of PET in the Danish health care system an analysis of all relevant comparators, which is based on parameter values specific to Denmark, is required. Furthermore, the review of the modelling literature in NSCLC has made two methodological issues for improvement apparent: that none of the models are specified as probabilistic and that none of the proposed models have been subject to validation. The combined PET-CT has only recently become available in all Danish centres. Whether current practice is defined as mediastinoscopy, endoscopic ultrasound with fine needle aspiration (EUS- FNA), endobronchial ultrasound with fine needle aspiration (EBUS-FNA) or PET, an a priori hypothesis would be that PET-CT is cost effective over current practice due to providing a faster track for staging and possibly also due to more accurate staging, ultimately leading to fewer futile thoracotomies and saved health care resources. Objective The objective of this study was to analyse the cost effectiveness of different strategies, including the combined PET-CT, for preoperative staging of patients with a confirmed diagnosis of NSCLC from a Danish health care sector perspective. 2

6 Model structure Complexity of staging Staging has traditionally been undertaken using the TNM system, classifying patients in five stages (from 0 to IV) according to tumour size and location (T-status), nodal involvement (N-status) and the presence of distant metastases (M-status). 15 According to clinical guidelines only patients in stages I and II should be referred to thoracotomy. This implies that patients in stages III and IV are considered inoperable and should be referred to non-surgical management. The specific arguments for being inoperable are in short: involvement of central and contralateral nodes (N2/3 disease rather than N0/1 disease), presence of distant metastases (M1 rather than M0), or a tumour involving mediastinal structures making resection impossible. Staging is rarely determined upon a single test because different tests have different rationales in terms of accuracy in assessing one (or more) of the TNM-parameters: Mediastinoscopy is used to assess nodal involvement and considered to have high specificity whereas its sensitivity is poor due to the fact that not all nodes can be assessed. EBUS-FNA and EUS-FNA have recently gained terrain as potential substitutes for mediastinoscopy that are less invasive and less expensive. Occasionally, these tests lead to the detection of distant metastases but the extent is uncertain. Chest-CT is usually used to assess all dimensions of the TNM-system. The sensitivity to detect nodal involvement has been questioned as it relies on morphological rather than histological information. Full-body PET is used to assess all dimensions of the TNM-system and is generally considered more accurate on N- and M-status than CT. Full-body PET-CT is the latest advance combining the PET with the CT in an integrated procedure. It is used to assess all dimensions of the TNM-system and is believed to be more accurate in all dimensions, as compared to other imaging techniques. The choice of a modelling approach Decision analysis has been defined as a systematic approach to inform decision-making under uncertainty. 16 It has become an important framework for cost effectiveness evaluation in health care as it provides some advantages over a trial-based approach of e.g. allowing for the comparison of all relevant alternative options, including all available evidence, using an appropriate time horizon, and explicitly addressing the decision uncertainty. 17 A decision model for cost effectiveness evaluation is thus a mathematical model that defines a series of possible consequences that would flow from alternative options and, based on model 3

7 inputs, the probability-weighted cost and outcome of each consequence are added up for the alternative options to return expected costs and outcomes. 18 There are three types of uncertainty relating to the guidance output of a decision model. 18 This has not been addressed in previous models of the cost effectiveness of different strategies for staging NSCLC. The first type arises from unexplained variability between individual patients i.e. the random chance that patients with the same underlying parameters will experience different outcomes (first-order uncertainty). The second type is parameter uncertainty referring to the conventional statistical variation around a point estimate (second-order uncertainty) and the third type is model uncertainty, which refers to the assumptions imposed by the modelling framework and therefore often referred to as structural uncertainty. To policy makers, the first type of uncertainty is irrelevant whereas the two other types are of crucial importance. Parameter uncertainty was incorporated in the present analysis by making the model probabilistic i.e. by assigning distributions to model parameters rather than just expected values. The only way of addressing structural uncertainty is to be explicit about model assumptions and discuss their impact to the guidance output. Relevant comparators The decision scenario depends on the point of departure in terms of current practice. Until recently, essentially all patients were first assessed based on a set of conventional modalities for diagnosis (clinical anamnesis, chest x-ray, bronchoscopy and/or transthoracic biopsy and a CT-scan) and after this initial work-up the patient were referred for more thorough staging procedures of for example mediastinoscopy. During the last 5-10 years this have been gradually changing, partly, due to an increased awareness on the importance of a fast diagnosis, but also due to the emergence of new technologies such as PET, EUS and EBUS. Many patients with a solitary pulmonary nodule (SPN) or similar indices of lung cancer are today referred directly to centres specialized on fast diagnosis and eventual staging of lung cancer. Although all specialized centres now have a PET or a PET-CT scanner to their availability clinical practice did vary in terms of whether it was used routinely until revision of the Danish guidelines in 2010 ( Moreover, it is currently recommended that patients with N1-findings on PET or PET-CT undergo confirmatory invasive staging due to increased risk of N2/N3 disease. 19 Relevant comparators after conventional work up thus comprise mediastinoscopy, EUS-FNA, EBUS-FNA, conventional PET (with confirmatory invasive test of N-status) and combined PET- CT (with confirmatory invasive test of N-status). Two other questions rise: 1) whether the combined PET-CT makes the initial CT-scan redundant as anatomical details are now provided in the PET test and 2) whether the confirmatory test after PET should be for those with suggested nodal involvement (in strict consensus with the clinical guidelines) 19 or for all patients. Altogether, this results in eight relevant comparators for which the feasibility of including them in a health economic model depends on the availability of parameter estimates. The majority of the literature on staging stems from studies of populations with a very high prevalence of lung cancer (typically %) whereas the results regarding diagnosis of SPN are typically based on studies with a prevalence of approximately 50%. Mediastinoscopy, EUS-FNA, CT and PET are all relatively old modalities for which test properties have been established in several clinical studies. 4

8 Linear EBUS making real-time fine needle aspiration possible, was however not commercially available until The diagnostic performance of PET without prior CT has been examined in at least one trial 21 and recently two randomised clinical trials investigating the performance of PET- CT as compared with mediastinoscopy have been reported. 1, 2 The proposed model structure includes eight strategies (see Figure 1): A. Upon conventional workup, including CT where M1-patients are referred directly to nonsurgical treatment, all M0-patients are referred to mediastinoscopy. Patients without nodal involvement are referred to thoracotomy while patients with nodal involvement are referred to non-surgical management. B. Upon conventional workup, including CT where M1-patients are referred directly to nonsurgical treatment, all M0-patients are referred to EUS-FNA. Hereafter the strategy is identical to strategy A. C. Upon conventional workup, including CT where M1-patients are referred directly to nonsurgical treatment, all M0-patients are referred to EBUS-FNA. Hereafter the strategy is identical to strategies A and B. D. Upon conventional workup, including CT where M1-patients are referred directly to nonsurgical treatment, all M0-patients are referred to PET. In case of upstaging to M1 the patient is referred to non-surgical management and in case of M0 nodal involvement determines the pathway; if nodal involvement is suggested the patient is referred to a confirmatory EBUS-FNA and otherwise directly to thoracotomy. E. Upon conventional workup, including CT where M1-patients are referred directly to nonsurgical treatment, all M0-patients are referred to PET-CT. Hereafter the strategy is identical to strategy D. F. Upon conventional workup without CT, all patients are referred to PET-CT. Hereafter the strategy is identical to strategies D and E. G. A strategy identical to strategy D except that all patients (and not just those with suggested nodal involvement) proceed to confirmatory EBUS-FNA. H. A strategy identical to strategy E except that all patients (and not just those with suggested nodal involvement) proceed to confirmatory EBUS-FNA. A mortality risk was built in for all occurrences of thoracotomy. Figures 1 to 9 illustrate each of the strategies in detail. 5

9 Figure 1 Simplified diagram of the proposed model structure Note: CT=computed tomography, Test+/Test- =test result presence/absence of disease, EUS- FNA=endoscopic ultrasound with fine needle biopsy, EBUS- FNA=endobronchial ultrasound with fine needle biopsy, PET=positron emission tomography, PET- CT=combined CT and PET, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N0=no nodal involvement, N1/2/3 and N2/3=nodal involvement. The pathway to the left of a bolded bar represents a clone definition in order to reuse the pathway elsewhere in the decision tree. 6

10 Figure 2 Strategy A of the proposed model structure: all patients to mediastinoscopy after M1- negative CT Note: CT=computed tomography, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. The pathway to the left of a bolded bar represents a clone definition in order to reuse the pathway elsewhere in the decision tree. 7

11 Figure 3 Strategy B of the proposed model structure: all patients to EUS- FNA after M1- negative CT Note: CT=computed tomography, EUS- FNA=endoscopic ultrasound with fine needle biopsy, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. Clone 1 is defined in strategy A, Figure 2. 8

12 Figure 4 Strategy C of the proposed model structure: all patients to EBUS- FNA after M1- negative CT Note: CT=computed tomography, EBUS- FNA=endobronchial ultrasound with fine needle biopsy, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. Clone 1 is defined in strategy A, Figure 2. 9

13 Figure 5 Strategy D of the proposed model structure: all patients to PET after M1- negative CT + confirmatory EBUS- FNA before thoracotomy if PET indicates N1/2/3 Note: CT=computed tomography, PET=positron emission tomography, EBUS- FNA=endobronchial ultrasound with fine needle aspiration, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. Clone 1 is defined in strategy A, Figure 2. 10

14 Figure 6 Strategy E of the proposed model structure: all patients to PET- CT after M1- negative CT + confirmatory EBUS- FNA before thoracotomy if PET indicates N1/2/3 Note: CT=computed tomography, PET- CT=combined positron emission tomography and CT, EBUS- FNA=endobronchial ultrasound with fine needle aspiration, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. Clone 1 is defined in strategy A, Figure 2. 11

15 Figure 7 Strategy F of the proposed model structure: all patients directly to PET- CT + confirmatory mediastinoscopy before thoracotomy if PET indicates N1/2/3 Note: CT=computed tomography, PET- CT=combined positron emission tomography and CT, EBUS- FNA=endobronchial ultrasound with fine needle aspiration, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. Clone 1 is defined in strategy A, Figure 2. 12

16 Figure 8 Strategy G of the proposed model structure: all patients to PET and EBUS- FNA after M1- negative CT Note: CT=computed tomography, PET=positron emission tomography, EBUS- FNA=endobronchial ultrasound with fine needle aspiration, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. Clone 1 is defined in strategy A, Figure 2. 13

17 Figure 9 Strategy H of the proposed model structure: all patients to PET- CT and EBUS- FNA after M1- negative CT Note: CT=computed tomography, PET- CT=combined positron emission tomography and CT, EBUS- FNA=endobronchial ultrasound with fine needle aspiration, Test+/Test- =test result presence/absence of disease, TN=true negative, FN=false negative, TP=true positive, FP=false positive, M1=distant metastases, N2/3=nodal involvement. Clone 1 is defined in strategy A, Figure 2. 14

18 Structural assumptions A decision-analytic model will always be a simplified representation of the real world scenario and, accordingly, a number of assumptions have to be made in order to arrive at a model, which is not overly simplified but still transparent in its structure. This paragraph comments on the assumptions made in order to arrive at the proposed structure. Assumptions related to the population of the model are commented on in the section concerning parameter estimates. It was assumed that the initial CT performed in all strategies except strategy F was used only to exclude patients with obvious M1 disease and not to inform nodal involvement. There were two arguments for that. First, the procedure that follows (mediastinoscopy, EUS-FNA, EBUS-FNA, PET or PET-CT) is in practice considered to be conclusive per se. Second, to use the information from the CT, estimation of joint probabilities for test accuracies of CT and the respective strategyspecific test would be required. Calculation of joint probabilities implies an assumption about at least conditional independence between tests (tests are independent conditional upon the presence or absence of disease). This assumption has been found not to hold in a recent meta-analysis, which also concluded that the value of initial CT-findings diminished once PET-findings were known. 22 Similar conclusions have been made for the mediastinoscopy and the ultrasonography-based technologies: the performances of the invasive methods are not independent of findings on CT or PET and the result of an invasive test will (almost) always overrule imaging The overall consequence of this assumption may be an underestimation of the diagnostic accuracy of all strategies relative to strategy F. It was overall assumed that there is no morbidity or mortality risk associated with the diagnostic tests. In earlier years of mediastinoscopy there has been a suspected mortality risk but this was not confirmed from a literature review of more recent studies. It is also known that there is an increased risk of future cancers related to undergoing the radiologic examinations. However, as these risks were apparent for all strategies they more or less cancel out (initial CT or PET-CT were performed in all strategies). However, the PET-based strategies do include, all in all, relatively more tests and the potential impact of not considering morbidity and mortality risks of diagnostic tests thus could favour these strategies. The perhaps most important assumption was related to the choice of a decision tree to model longterm cost effectiveness. While such structure is the most common choice for modelling diagnostic strategies it has one drawback: that the elapse of time is not explicit. This usually means that all relevant events occurring at future points in time have to be built in the model as explicit branches in order to be considered. In other words, what was included in the analysis is what can be seen from the decision tree and, it is implied that all nodes represent one-off events. Should nonmodelled future events be unevenly distributed across strategies this may cause bias. In particular, it was not build in the model that relapses might occur after a successful thoracotomy and incur extra costs for staging and treatment (the effect on life expectancy was integrated as estimates of survival includes patients with and without recurrence). The choice was not voluntary but given by a lack of data on recurrence rates. The overall consequence on the incremental cost effectiveness could be a bias against the most accurate strategies as the number of thoracotomies with relapse within 5 years postoperative would presumably be higher for strategies sending more false negatives to surgery. 15

19 Similarly, the model structure did not include costs of confirmatory tests for patients detected with distant metastases on the initial CT (PET in strategy F). Clinical practice often confirm such findings by Magnetic Resonance (MR) scans, ultrasonography or others tests but, since the practice is truly varied and not consensus-based it was decided not to attempt defining a standard. The overall consequence on the incremental cost effectiveness could be a bias against strategy A, B, C where fewer tests are included in the strategies and more supplemental tests therefore are likely to be appended. General methodology for cost effectiveness evaluation The economic evaluation adapted a Danish health care sector perspective for a time horizon of five years. Outcomes were defined as LY and QALY. Discounting of 3% per year was applied. All monetary estimates were inflated to (1 = 7.5 DKK) using the consumer price index where relevant. 16

20 Populating the model Probability revision The probability of a positive test result P(Test+) was calculated from the true disease distribution (prevalence) and the test s sensitivity and specificity: P(Test+) = sensitivity * prevalence + (1 specificity) * (1 prevalence) The probability of a negative test result P(Test-) was calculated as the complement to a positive test result: which is equivalent to: P(Test-) = 1 P(Test+) P(Test-) = specificity * (1 prevalence) + (1 sensitivity) * prevalence The distribution between true positive (TP) and false positive (FP) given a positive test result was calculated using Bayes theorem. Its specific form can be defined as: TP = sensitivity * prevalence / P(Test+) FP = 1 - TP Similarly, the distribution between true negative (TN) and false negative (FN) given a negative test result was calculated as: TN = specificity * (1 prevalence) / P(Test-) FN = 1 TN Study population There are two relevant choices of study population, depending on whether diagnostics and staging are seen as integrated regimens: patients with indices of lung cancer, who need staging if diagnosed, or patients with a diagnosis of lung cancer, who needs staging. In some health care systems patients are referred directly to centres specialized in diagnosis and eventual staging of lung cancer at an early stage in the diagnostic process (upon indices of lung cancer, for example, a SPN). In other health care systems patients are not routinely referred until they have a histologically or cytologically verified diagnosis. The latter definition is traditionally used in the modelling literature and is associated with prevalence estimates approaching 100%. The former definition, on the other hand, is increasingly seen in clinical studies, perhaps representing a more modern approach with fast track and highly specialised diagnostics. The prevalence in this 17

21 setting would be in the region of 20-30% as many of the referred patients will be cancer-free or will have another cancer than lung cancer. 1 A decision was made to use the traditional definition of the study population to be able to compare findings with the existing literature and for pragmatic purposes of moderating model complexity, as patients with other types of cancer or no cancer would require separate pathways in the model. The study population was accordingly defined as 65-year-old patients with a histologically or cytologically verified diagnosis of NSCLC who are fit for surgery. True staging distribution The handling of two dimensions of disease stage M-status and N-status required some considerations. Usually in decision trees involving a test providing imperfect information, a prior (pre-test probability of disease) is used along with information on test accuracy to calculate the posterior (the post-test probability for disease). The probability revision from prior to posterior is based on Bayes theorem, which was introduced in the section above. In a scenario including one test for the presence/absence of distant metastases and another test for the presence/absence of nodal involvement, the posterior of test 1 cannot be adapted as the prior for test 2. Nor can individual priors for the two tests be used unless the two dimensions are truly independent. The two dimensions are highly dependent as patients with presence/absence of distant metastases are significantly more likely to have/not have nodal involvement (and vice versa). Also, the assessment of the two dimensions often occurs simultaneously due to the radiologic tests covering both dimensions. As a decision tree defines activities sequentially, the proposed structure was specified arbitrarily with assessment of M-status as a step 1 and assessment of N-status as a step 2. This required different priors in M1-positive and M1-negative populations. It was decided that including a third dimension would make the model to complex, thus it was decided not to include the T-status. The assessment of nodal involvement had different cut-off values depending on the strategy and staging modality; for the invasive modalities (mediastinoscopy, EUS-FNA and EBUS-FNA) a N0/1- versus N2/3-dichotomisation was defined and for PET/CT in strategies D, E and F a N0- versus N1/2/3-dichotomisation was defined. The two systems of dichotomies required individual priors. Priors were estimated empirically from The Danish Lung Cancer Registry (DLCR), which is a national clinical database established in The database includes patients with a confirmed diagnosis of lung cancer (and patients who have had a confirmed diagnosis that was found to be incorrect after thoracotomy). For the present purpose, the analysis was restricted to N=17900 patients with a TNM-assessment, who had been diagnosed between 2003 and Of these 1026 patients were excluded due to unknown M-status (Mx) and in the remainder sample 1431 patients had an unknown N-status (Nx). The estimation of conditional probabilities is based on the 1 The prevalence was informed from an informal survey at two of the largest centres in Denmark where patients are being referred if they have a SPN or similar indices of lung cancer. In Aarhus it was reported that 19% of the referred patients (179/965) were diagnosed with lung cancer in At Bispebjerg the corresponding number was 31% (433/1398). 18

22 assumption that the TNM-status in the database represents a true disease status. Table 1 shows the empirical estimates. Table 1 Disease statuses in a population with a histologically or cytologically confirmed diagnosis of lung cancer according to the Danish Lung Cancer Registry Prior Population Positive status Prevalence Distant metastases (M1) Nodal involvement (N1/2/3) Conditional on no distant metastases (M0) Conditional on distant metastases (M1) Nodal involvement (N2/3) Conditional on no distant metastases (M0) Conditional on distant metastases (M1) Conditional on no distant metastases (M0) and nodal involvement (N1/2/3) Conditional on distant metastases (M1) and nodal involvement (N1/2/3) Note: Estimations were based on the variable utnm in the database. Cases with unknown M- status (Mx) were excluded (n=1026) and of the remainder sample n=1431 had unknown N- status (Nx). Probability for postoperative mortality The DLCR was furthermore used to inform the probabilities for mortality within 30 days after thoracotomy. Of 6406 thoracotomies performed between 2003 and 2009, 250 patients have died within 30 after the surgery, which results in a probability at The probability of postoperative mortality was assumed to be constant across all thoracotomies, i.e., be independent on the disease status and what tests were performed prior to arriving at the thoracotomy. Diagnostic accuracy of tests Estimates of test accuracies were as far as possible adapted from published systematic literature reviews and meta-analyses. Where several systematic reviews were identified the most recent was chosen as a main rule. Literature estimates were compared with Danish observations from the PERALUST trial 1 to judge context validity where possible. Tables 2 and 3 show the accuracies specified for the model and details about their source are given in the following. 19

23 Table 2 Test accuracies for detection of nodal involvement N0/1 versus N2/3 N0 versus N1/2/3 Sensitivity Specificity Sensitivity Specificity Source Mediastinoscopy NR NR Detterbeck et al EUS- FNA NR NR Puli et al EBUS- FNA NR NR Gu et al PET NR NR Unpublished pooled average PET- CT NR NR Unpublished pooled average Note: EUS- FNA = endoscopic ultrasonography with fine needle aspiration, EBUS- FNA= endobronchial ultrasound with fine needle aspiration, PET = positron emission tomography, PET- CT=combined computed tomography and PET, NR = not relevant for the present application. Table 3 Test accuracies for detection of distant metastases M0 versus M1 Sensitivity Specificity Source CT Herder et al PET Herder et al PET- CT Cerfolio et al. 2007, 2 PERALUST Note: CT = computed tomography, PET = positron emission tomography, PET- CT=combined computed tomography and PET. Mediastinoscopy A systematic literature review from 2007 included 19 studies (n=6505) and found a pooled sensitivity of 0.78 and a pooled specificity of 1.00 for the ability of mediastinoscopy to correctly detect nodal involvement. 28 These values are significantly higher than the values observed in the Danish PERALUST trial; sensitivity 0.30 (95% CI 0.16; 0.51) and specificity 1.00 (95% CI 0.93; 1.00) but it was assumed that this was due to the trial not being intended to inform measures of accuracy (sample size issues, research protocol rather than daily practice etc.). EUS-FNA Two systematic reviews have recently been reported for the ability of EUS-FNA to detect nodal involvement. In 2007, Micames et al. reported a review of 18 studies (n=1201) and estimated the pooled accuracies at 0.83 (95% CI 0.78; 0.87) for sensitivity and 0.97 (95% CI 0.96; 0.98) for specificity. 25 In 2008, Puli et al. reported a review of 32 studies (n=2680) and estimated the pooled sensitivity at 0.88 (95% CI 0.86; 0.90) and the pooled specificity at 0.96 (95% CI 0.95; 0.97). 29 In the Danish PERALUST trial a sensitivity of only 0.67 (95% CI 0.42; 0.84) was observed while the specificity was found to be 1.00 (95% CI 0.77; 1.00). Among patients who also had a PET/CT a sensitivity of 0.93 (95% CI 0.70; 0.99) was found. The discrepancy with the sensitivity of the Danish trial was, again, considered to be due to the trial not being intended to inform measures of EUS-FNA accuracy. 20

24 EBUS-FNA Three systematic reviews on the diagnostic accuracy of EBUS-FNA to detect nodal involvement have been recently published. Adams et al. reported a pooled sensitivity of 0.88 (95% CI 0.79; 0.94) and a specificity of 1.00 (95% CI 0.92; 1.00), based on 10 original studies. 30 Varela-Lema et al. identified and described 14 studies with sensitivities ranging from 0.85 to 1.00 but did not report a pooled estimate. 31 The most recent work has been done by Gu et al. who reported pooled estimates for sensitivity at 0.93 (95% CI 0.91; 0.94) and for specificity at 1.00 (95% CI 0.99; 1.00). 24 This meta-analysis was based on 11 studies and 1299 patients. CT In the present context CT is used only to exclude patients with obvious M1 disease. The relevant literature on this dimension appeared to be not only limited but also characterised by protocol variations in terms of the anatomical focus of the scan (thorax, upper abdomen, brain etc.) and eligibility criteria (whether some patients were already excluded from the study population due to other test results). Herder et al. examined test performance in a population relevant to the present model (n=465) and reported a sensitivity of 0.80 (95% CI ) and a specificity of 1.00 (95% CI ). 21 PET No systematic reviews were found to assess the accuracy of PET in relation to distinguishing N0 versus N1/2/3. Four original studies were instead identified and retrieved; two retrospective cohort studies 32, 33 34, 35 including 50 and 36 patients, respectively, and two prospective cohort studies including 256 and 110 patients respectively. The pooled average of these studies (see Figure 10) was used for the model: sensitivity 0.63 (95% CI 0.55; 0.70) and specificity 0.68 (95% CI 0.63; 0.74). The corresponding findings of the PERALUST study were 0.60 (95% CI 0.44; 0.74) for sensitivity and 0.80 (95% CI 0.65; 0.90) for specificity. The ability of PET to detect distant metastases was assessed in only one of the four abovementioned studies. 34 However, as a larger and more recent study from a European setting (also adapted for CT test accuracy) has been reported this was used for the model; Herder et al. reported a sensitivity of 0.73 (95% CI 0.60; 0.84) and a specificity of 0.98 (95% CI 0.92; 0.99). 21 It should be noted, though, that these estimates are different from older estimates. For example, Reed et al. reported in 2003 that the sensitivity across US was observed at 0.83 (95% CI 0.61; 0.94) and the specificity at 0.93 (95% CI 0.89; 0.95). 35 Figure 10 Sensitivity and specificity of PET to identify N0 versus N1/2/3 21

25 PET-CT No systematic reviews were found assess the accuracy of PET-CT in relation to distinguishing N0 versus N1/2/3. Five original studies were instead identified and retrieved These were the four studies also examining accuracy of stand-alone PET plus an additional prospective study in 122 patients by Yang et al. 36 The pooled average of these studies was estimated at 0.75 (95% CI 0.69; 0.80) for sensitivity and 0.78 (95% CI 0.73; 0.82) for specificity (see Figure 11). In the PERALUST trial sensitivity and specificity to detect nodal involvement were observed at 0.58 (95% CI 0.42; 0.72) and 0.78 (95% CI 0.63; 0.88). The ability of PET-CT to detect distant metastases appeared to be informed by almost no empirical studies. In a cohort study, Cerfolio et al. estimated the sensitivity at 0.92 (n=129) but reported no estimate for the specificity. 34 This was therefore estimated from the Danish PERALUST trial at 0.98 (n=87). Figure 11 Sensitivity and specificity of PET- CT to identify N0 versus N1/2/3 Outcome parameters Survival Ad hoc extractions of mean 5-year survival were obtained from The Danish Lung Cancer Registry as only the proportions alive is reported in official reports of the registry. Survivals were stratified according to the true disease distribution and, for individuals without distant metastases and N2/3, according to the treatment modality received. Table 4 present the obtained estimates for a 5-year follow up. Patients with M1 or M0 and N2/3 should be referred to non-surgical treatment but due to misclassification some patients will be referred to a futile thoracotomy before getting the nonsurgical treatment. The expected survival for these patients was assumed to be the same as for patients directly referred to non-surgical treatment, except for the mortality risk associated with the thoracotomy. 22

26 Table 4 Observed 5- year survivals in years after being diagnosed with lung cancer for different strata of patients registered in the Danish Lung Cancer Registry Survival stratum Mean SD Sample size M0, N0/1 After thoracotomy After non- surgical management M0, N2/ M day mortality after thoracotomy Note: Estimates are based on original analysis of the Danish Lung Cancer Registry with vital status updated February 9, M0 and M1=distant metastases no/yes, N0/1 and N2/3=nodal involvement no/yes, SD = Standard deviation. Quality-adjusted life years A systematic literature search was conducted to identify studies on health-related quality of life in NSCLC. PubMed was searched on July 24, 2010 using the following search terms: "non small cell lung cancer" AND "quality of life" AND (utility OR utilities OR preference*), which returned 33 hits. Upon assessment of abstracts it was clear that very few of the studies reported genuinely preference-weighted outcomes and that the few who did primarily examined outcomes of nonsurgical treatment. Only 5 studies were retrieved for possible adaptation; these are presented in the following. Two studies appeared to have attempted to derive the value that society places on the avoidance of severe symptoms associated with NSCLC in stage III and IV. 37, 38 Both studies estimated a baseline value for stable disease with no toxicity (0.653 and 0.626, respectively) and established models for the additional decrement in utility experienced upon various symptoms associated with non-surgical treatment. The lowest value was derived for progressive disease with no treatment response at Van den Hout et al. followed 299 patients randomised to different radiotherapy regimens and observed baseline utilities of 0.62 and 0.52 in the two groups. 39 These value appeared to be relatively stable towards endpoint observations of about 0.60 for both groups. Belani et al. reported a similar study for patients receiving chemotherapy but reported only changes over time and not descriptive utilities. 40 Trippoli et al. reported a cross sectional study in 95 patients undergoing chemotherapy, radiotherapy or surgical resection. Utility values were found to range from 0.50 in the subgroup of patients with more than 12 months since diagnosis to 0.68 in those with no distant metastases. 41 Altogether, the literature search revealed that very few studies had observed quality of life in patients undergoing staging or surgical treatment for NSCLC. Populating the model would therefore require assumptions about not only the utility decrement associated with a given test or treatment, but also the duration of such decrement. It was therefore decided that using hypothetical values for practically all other health states than non-surgical treatment would be unreasonable. Estimation and analysis of QALYs was thus refrained from. 23

27 Cost parameters The cost of individual procedures for staging was estimated from the Diagnosis-Related Grouping (DRG) case-mix system. The system is based on a cost database, which is regularly updated and validated by the National Board of Health and respective clinical societies. As no precision estimates for the resulting tariffs are given this work assumed a standard deviation corresponding to 10% of the average cost for diagnostic tests and 25% of the average cost for treatments. Table 5 list the tariffs. It was assumed that patients who had thoracotomy despite (unknown) M1 disease would receive non-surgical management during the follow up time. No overall tariff exists for non-surgical treatment as it depends on what modalities are indicated. For patients who were inoperable due to N2/3 (but have a true status of M0) it was assumed that chemotherapy and radiotherapy with at least 30 fractions were administered. For patients who were inoperable due to M1 disease it was assumed that only chemotherapy was administered although in practice some patients might receive radiotherapy. The practice of such would however be limited and, when occurring, only 1 (inexpensive) fraction is usually offered. The tariff for chemotherapy is the same whether it is used for in- or outpatient treatment. A standard chemotherapy-regimen consisting of cisplatin + vinorelbine on day 1 and vinorelbine alone on day 8, repeated every three weeks until 4 cycles have past, was adapted. Table 5 Costs of procedures adapted from the 2010 Diagnosis- Related Grouping case- mix system Procedure Items Tariff DKK Total DKK Total EUR Mediastinoscopy KTGE EUS- FNA UXUC EBUS- FNA GCA CT PG14H PET PG17G PET- CT PG17H Thoracotomy Grey zone Non- surgical regimen M0 patients Chemotherapy PG12E Chemotherapy complex PG12C Radiotherapy BWGC Total NA NA NA Non- surgical regimen M1 patients Chemotherapy PG12E Chemotherapy complex PG12C Total NA NA NA Note: Tariffs were extracted for diagnosis DC34. EUS- FNA = endoscopic ultrasonography with fine needle aspiration, EBUS- FNA= endobronchial ultrasound with fine needle aspiration, CT = computed tomography, PET = positron emission tomography, PET- CT=combined PET and CT, M0 = no distant metastases, M1 = distant metastases, NA = not applicable. The appropriate number of cost categories to include depend on whether a difference is hypothesised between strategies. Previous international models of the cost effectiveness of different strategies for staging include costs related to the diagnostic procedures only It was therefore investigated whether other health care utilisation would be associated with the strategy. The sample of the PERALUST trial was used to examine whether costs would differ between the two arms of conventional staging (n=91) and PET-CT-based staging (n=98). 1 24

28 Resource use was extracted from national registries: the National Health Insurance Registry (number of primary health care services including visits to general practitioners, physiotherapists, practicing specialist doctors, dentists etc.), the National Patient Registry (number of secondary health care of in- and outpatient services), and The Registry of Prescribed Medicine (consumption of prescription medicine). The observed resource use was valued using prices of collaborative agreements for primary care, Diagnosis-Related-Grouping tariffs for hospital services and market prices of medication. There were no significant differences between groups in any of the resource or cost categories and it furthermore appeared that extra-hospital service utilisation constituted less than 2% of total health care utilisation. For that reason it was considered appropriate to include costs related to diagnostic procedures and lung cancer related therapy only (and assume that all other costs were even across strategies). A secondary argument of a more pragmatic character was that original Danish observations on resource use were available only for the two strategies of the PERALUST trial and that assumptions thus had to be made for the other strategies. Specification of probability distributions It is sometimes claimed that the specification of probability distributions for individual parameters in a health economic model is arbitrary. However given the properties of individual types of parameters the choice is usually limited within a few alternatives. The following describes the available alternatives, though limited to standard distributions included in the modelling software Tree Age Health Care Pro (version 2009), and the choices made for the model. Table 6 lists all parameters used in the model and their assigned distributions. Where relevant the conventional equation for conversion between standard errors and standard deviation was used: 42 SE=SD / n Probabilities Probabilities follow a continuous distribution bounded by 0 and 1. The only relevant specification is the Beta distribution with a mean of r/n, where r is occurrences and n is sample size. A convenient property of TreeAge is that approximation is possible from conventional means and standard errors if r and n are unknown. 43 Beta(n,r) [integer form] or Beta(α,β) [real number form] where α=mean^2*(1-mean)/(se^2) and β=mean*(1-mean)/(se^2)- α 25

29 Table 6 Model parameters and assigned distributions Description Type Parameter 1 Parameter 2 Prevalence M1 Beta Prevalence N1/2/3 in M0 Beta Prevalence N1/2/3 in M1 Beta Prevalence N2/3 in M0 Beta Prevalence N2/3 in M0 and N1/2/3 Beta Prevalence N2/3 in M1 Beta Prevalence N2/3 in M1 and N1/2/3 Beta Probability of mortality within 30 days after thoracotomy Beta Sensitivity of CT to detect M1 Beta Sensitivity of EBUS- FNA to detect N2/3 Beta Sensitivity of EUS- FNA to detect N2/3 Beta Sensitivity of mediastinoscopy to detect N2/3 Beta Sensitivity of PET to detect M1 Beta Sensitivity of PET to detect N1/2/3 Beta Sensitivity of PET- CT to detect M1 Beta Sensitivity of PET- CT to detect N1/2/3 Beta Specificity of CT to detect M0 Beta Specificity of EBUS- FNA to detect N0/1 Beta Specificity of EUS- FNA to detect N0/1 Beta Specificity of mediastinoscopy to detect N0/1 Beta Specificity of PET to detect M0 Beta Specificity of PET to detect N0 Beta Specificity of PET- CT to detect M0 Beta Specificity of PET- CT to detect N0 Beta Survival 30- day postoperative mortality Gamma Survival M0N0/1 patients after non- surgical treatment Gamma Survival M0N0/1 patients after thoracotomy Gamma Survival M0N2/3 patients Gamma Survival M1 patients Gamma Cost of CT Normal Cost of EBUS- FNA Normal Cost of EUS- FNA Normal Cost of mediastinoscopy Normal Cost of non- surgical management M0 Normal Cost of non- surgical management M1 Normal Cost of PET Normal Cost of PET- CT Normal Cost of thoracotomy Normal The real number form of the Beta distribution is only defined for a number of occurrences r < the number of observations n; hence distributions of 1.00 were approximated by (r- 1)/n. M1=distant metastases, N2/3=nodal involvement, CT=computed tomography, EUS- FNA=endoscopic ultrasound with fine needle biopsy, EBUS- FNA=endobronchial ultrasound with fine needle biopsy PET=positron emission tomography, PET- CT=integrated PET and CT, M0=no distant metastases and N0/1=no nodal involvement. 26