Dasatinib, nilotinib and standard-dose imatinib for the first-line treatment of chronic myeloid leukaemia: systematic reviews and economic analyses

Molecular mechanism

The molecular characteristic of CML is the presence of an acquired BCR–ABL fusion gene [oncogene fusion protein consisting of breakpoint cluster region (BCR) and Abelson oncogene (ABL)] in multipotent stem cells. More than 90% of people who are diagnosed with CML have an acquired (non-inherited) chromosomal abnormality caused by a reciprocal translocation between chromosomes 9 and 22 in an individual stem cell. The result is a shortened 22q, which is called the Philadelphia chromosome. 2,3 More specifically, the Abelson oncogene (ABL1), which is located on chromosome 9, translocates to the BCR gene on chromosome 22. The result is a fusion gene, BCR–ABL, and its corresponding protein, a constitutively active BCR–ABL tyrosine kinase. BCR–ABL tyrosine kinase is not controlled by normal cellular mechanisms and its presence leads to enhanced cell proliferation, resistance to apoptosis (programmed cell death) and genomic instability. These are key features in the pathophysiology of CML. 4,5 Within the CML population, approximately 10% of people do not have a demonstrable Philadelphia chromosome but have a complex of different translocations that still results in the formation of the BCR–ABL gene and its product. 6

Diagnosis

Chronic myeloid leukaemia is diagnosed by the presence of a characteristic pattern of cells in the blood and bone marrow in conjunction with specific cytogenetic and molecular abnormalities.

At presentation, patients typically have an enlarged spleen and a raised white cell count, with higher than normal numbers of immature WBCs. Bone marrow biopsy typically shows very little fat present and the bone marrow space occupied entirely by large numbers of leukaemia cells. 7

The presence of the Philadelphia chromosome is important both in terms of diagnosis and for monitoring responses to treatment. It is usually demonstrated by cytogenetic techniques that involve examining bone marrow cells in mitosis under a microscope to allow visualisation of metaphase chromosomes. 8 This test can also identify additional clonal chromosomal abnormalities in Philadelphia chromosome-positive (Ph+) cells (clonal cytogenetic evolution), which may be important indicators of prognosis. The technique requires at least 20–30 bone marrow cells in mitosis, which can be difficult to achieve. 9 There are considerable sampling errors because of the relatively small numbers of cells examined and the infrequency of measurement because bone marrow examination is a relatively invasive, although minor, procedure. The sensitivity is approximately 5% if 20 metaphase chromosomes are examined. 6

Fluorescence in situ hybridisation (FISH) is a sensitive and quantitative method that is used to detect specific chromosomal aberrations not only in cells undergoing metaphase, but also in interphase nuclei. 6,10 It uses specific fluorescent probes to map the chromosomal location of genes and identify other genetic abnormalities. In the case of CML, the probe looks for the BCR–ABL fusion gene in bone marrow or peripheral blood cells. 10 This test is usually performed in addition to the conventional cytogenetic test and uses approximately 200 bone marrow or blood cells for interphase FISH. 6 The limit of detection is between 1% and 5% abnormal cells. 6

Reverse transcriptase-polymerase chain reaction (RT-PCR) to detect BCR–ABL transcripts is also sometimes used to provide confirmation of diagnosis in CML. Through this technique, the level of BCR–ABL transcripts in peripheral blood or bone marrow is measured and one CML cell in 100,000 normal cells can be detected. 6 This qualitative technique is a simplified version of real-time quantitative polymerase chain reaction (qPCR), which is used to detect and quantify the level of BCR–ABL transcripts in a sample, and can be used to monitor disease progression and molecular response to treatment more closely. All of the above diagnostic techniques are currently recommended in the UK for the confirmation of CML diagnosis. 8

Natural history and clinical presentation

With the advent of a new class of drugs called tyrosine kinase inhibitors (TKIs) for the treatment of CML, imatinib being the first (see Imatinib, below), the natural history of the disease has been markedly changed. Current evidence suggests that patients whose disease responds favourably to treatment with imatinib may remain essentially symptom free for at least 10 years. 7 The following paragraphs describe the natural history of the disease in the absence of imatinib treatment.

Traditionally, CML has been regarded as a progressive disease that evolves through three phases. The initial chronic phase (CP) during which the disease is stable and slow to progress is followed after a variable interval by progression through an accelerated phase (AP) to a rapidly fatal blast crisis (BC). In approximately, one-third of patients there is no demonstrable AP, with the disease progressing directly from CP to BC. Transition between the phases may be gradual or rapid.

Incidence

The Haematological Malignancy Research Network (HMRN), based in Yorkshire, estimates that 530 cases of CML are newly diagnosed in the UK each year, an annual age-standardised rate of 1.1 per 100,000 for men and 0.7 per 100,000 for women. 22

Figure 1 shows the annual estimated incidence of CML in the UK with age and sex distributions. The data are extrapolated from those collected within the HMRN region, whose population of 3.7 million is broadly representative of the UK as a whole. Approximately 60% of those diagnosed with CML are male. CML occurs in all age groups, although it is uncommon in those aged < 30 years; the median age at diagnosis is 58 years (this includes all phases). 22

FIGURE 1.

Annual estimated incidence in the UK, by age and sex. Source: Haematological Malignancy Research Network (www.hmrn.org).

Prognosis

There are two prognostic staging scores for CML in common use: the Sokal score23 and the Euro or Hasford score. 24 Details of how the scores are calculated are shown in Table 2. Both scores are used to determine if a patient is at low, intermediate or high risk of death and may also predict response to treatment. Both must be applied at diagnosis, prior to any treatment. The Sokal score is based on age, spleen size, and platelet and peripheral blood blast count. The Hasford score also includes data on eosinophil and basophil counts. The level and timing of haematological, cytogenetic and molecular responses provides important prognostic information and it is a widely accepted goal for patients to achieve a complete cytogenetic response (CCyR) within 18 months of CML therapy. 1,6 Both scores were developed prior to the introduction of tyrosine kinase inhibitors (TKIs) [the Hasford score in response to improvements in survival seen with interferon (IFN) treatment] but they appear to have some value in predicting response to treatment with TKIs.

At the 18-month follow-up of the IRIS trial26 (International Randomised Study of Interferon versus STI571) of imatinib and IFN, 49%, 67% and 76% of people with high-, intermediate- and low-risk Sokal scores, respectively, had achieved a CCyR. This relationship was maintained at the 48-month update with patients with a high Sokal score having a 69% probability of achieving a complete cytological response compared with 84% and 91% for patients with intermediate- and low-risk scores, respectively. 27 A similar relationship was seen with molecular response at 12 months; 38% of those in the high-risk group had a reduction from baseline of at least 3-log in BCR–ABL transcripts compared with 45% in the intermediate-risk group and 66% of those in the low-risk group (p = 0.007). 28

The trials in this assessment, ENESTnd (Evaluating Nilotinib Efficacy and Safety in clinical Trials – Newly Diagnosed patients) and DASISION (Dasatinib vs Imatinib in Patients With Newly Diagnosed Chronic Phase CML), use the Sokal and Hasford staging scores, respectively. 20,29 The ENESTnd study reported at 12 months that rates of CCyR for study arms nilotinib (300 mg), nilotinib (400 mg) and imatinib (400 mg) were 74%, 63% and 49%, respectively, for patients at high risk (Sokal). Rates of major molecular response (MMR) for study arms nilotinib (300 mg), nilotinib (400 mg) and imatinib (400 mg) were 41%, 32% and 17%, respectively, for patients at high risk (Sokal).

The DASISION study29 reported at 12 months that rates of CCyR for study arms dasatinib (100 mg) and imatinib (400 mg) were 78% and 64%, respectively, for patients at high risk (Hasford). Rates of MMR for study arms dasatinib (100 mg) and imatinib (400 mg) were 31% and 16%, respectively, for patients at high risk (Hasford) (see Chapter 3, Complete cytogenetic response and Major molecular response for full results). Comparability between ENESTnd20 and DASISION29 risk score responses should be treated with caution, with between-trial differences potentially resulting from the different risk group scoring systems adopted.

Survival

The most recently available survival statistics for all leukaemia in the UK are based on data collected from 2001–7. 30 The 5-year relative survival (survival of patients taking into account other causes of death) rate was 39.7% for men and 41.0% for women up to 2006, with a predicted rate of 42.7% for 2007. 30 The predicted 10-year survival rate for 2007 was 33.8% for men and 35.3% for women. 30 With fewer survival statistics available for CML, the IRIS trial26 of imatinib (see Imatinib, below) reports overall survival (OS) at 8 years of 85% for patients receiving imatinib.

Recent analysis of survival among patients with CML in the USA, derived from the 1973–2006 limited-use database of the Surveillance, Epidemiology and End Results Program of the United States National Cancer Institute suggests a dramatic recent increase in long-term survival for people with CML since the introduction of imatinib into routine clinical practice. 31 For all age groups combined, 5-year relative survival increased from 32.5% in 1990–2 to 54.6% in 1999–2006 (p < 0.05). For the period 1999–2006, 5-year relative survival was approximately 78.0% for age groups 15–44 year and 45–54 years, 63% for 55–64 years, 39.5% for 65–74 years and 24.7% for the ≥ 75 years age group. 31 There were indications from the data of improvements in long-term survival in the older age groups but long-term prognosis remained poor and essentially unchanged for the oldest patients (≥ 75 years age group). 31

Surrogate outcomes

In the absence of long-term follow-up, the above measures of treatment response may be regarded as ‘surrogate outcomes’ for patient-relevant outcomes [OS, disease progression, quality of life (QoL)], with CyR and molecular response used as the primary outcomes in current trials. 18–21 The use of surrogate outcomes rather than more patient-relevant outcomes may be easier, more economical and provide earlier results. 34 This can lead to faster licensing time and dissemination of new treatments. 35 The use of surrogate outcomes is essential in Phase II and Phase III trials aiming to establish a drug’s potential benefit. 36 However, the use of surrogate outcomes can also be harmful when there is a lack of an independent causal association between a change in the surrogate outcomes and a change in the patient-relevant outcomes, thus the evaluation and validation of using a surrogate outcome is warranted. 34–36 The value of surrogate outcomes can be judged against a hierarchy of evidence, which ranges from biologically plausible relationships (weak evidence) to changes in the surrogate corresponding to equal changes in the patient-relevant outcome, assessed by clinical trials (strong evidence). 34,35

Schrover et al. 37 reported on the development of a predictive survival model for patients with chronic myeloid leukaemia in chronic phase (CP-CML), according to CyR rates in seven IFN-based RCTs. 37 They estimated a weighted odds ratio – for the survival of patients who achieved a major cytogenetic response (MCyR) when compared with those who did not – of 7 (95% CI 5 to 11) at 2 years and 5 (95% CI 3 to 8) at 4 years. Median survival was increased by 1.8 years for every 25 percentage point increase in MCyR rate. The predictive model reported by Schrover et al. 37 provides support for CyR predicting long-term survival within IFN class treatments for CP-CML. The evidence for the use of surrogate outcomes within the TKI (imatinib, dasatinib and nilotinib) class of CP-CML treatment is unclear, and may not be available particularly for the newer second-generation TKIs. Therefore, only imatinib may provide evidence for the use of surrogate outcomes within the TKI drug class (see Chapter 4, Assessment of evidence to support the use of complete cytogenetic response and major molecular response as surrogate outcomes).

Disease progression

Typically, disease progression describes the process in which the disease develops into the AP or to BC. Differences in the definition of AP have resulted in the use of more specific definitions of disease progression. The definition of progression used in a trial of this assessment relies on participants meeting any one of the four criteria:20

1. development of AP or BC CML

2. loss of CHR

3. increase in Ph+ bone marrow metaphases to more than 35%

4. increasing WBC count (a doubling of white cell count to > 20 × 10⁹/l) in the absence of complete haematological response.

Allogeneic stem cell transplant

Currently, the only known curative treatment for CML is allogeneic haematopoietic stem cell transplantation (alloHSCT), either from a matched related or unrelated donor. 38,39 Patient age, disease phase and duration, the degree of mismatch between patient and donor and therapy before transplantation all influence outcome. Younger patients in CP receiving a transplant from a matched sibling donor soon after diagnosis have the best prognosis. 40 Two studies have shown similar outcomes for transplantation in patients with CP-CML using either a fully matched related or unrelated donor, with 5-year survival rates of > 70% for people aged ≤ 50 years who undergo transplantation within a year of diagnosis. 41,42 Results are less promising for those in AP and BC. 39

The morbidity and mortality of alloHSCT is considerable; transplant-related mortality ranges from 15% to 40%. 43

Allogeneic haematopoietic stem cell transplantation is not a treatment option for many people, either for reasons related to age at diagnosis (the median age for diagnosis of CML is 59 years, and many patients are considered to be unsuitable for a transplant at diagnosis) or because of lack of a suitable donor. 6 UK recommendations propose the use of alloHSCT with failure of imatinib and or second-generation TKIs, or where younger patients have progressed to the AP. 8

Imatinib

Imatinib [originally STI571; trade name Gleevec® (USA) or Glivec® (Europe/Australia/Latin America), Novartis] is an orally administered TKI.

Dasatinib

Dasatinib (BMS-354825; trade name Sprycel®, Bristol-Myers Squibb) is a second-generation TKI.

Nilotinib

Nilotinib (AMN107; trade name Tasigna®, Novartis) is a second-generation TKI.

Interventions and comparators

Table 4 shows the three treatment pathways that will be evaluated using the decision model, assuming that second-line use of TKIs is or is not available within the NHS. For a description of how these admittedly simplified treatment sequences were arrived at, please see the cost-effectiveness analysis methods (see Chapter 7, Approaches to modelling treatments for chronic myeloid leukaemia).

Apart from those relating to cytogenetic or molecular response at 12 months, no important and statistically significant subgroup differences emerged in the clinical effectiveness evidence.

Population

Adults with newly diagnosed CP, Ph+ CML. If possible, newly diagnosed CP-CML without genetic mutation (non-Philadelphia chromosome) will also be considered. In reality, for consistency, the patient population modelled will have to closely mirror the populations in the main trials from which the effectiveness estimates are derived.

Outcomes

The main outcomes that will determine the development of the decision model are:

• lifetime quality-adjusted life-years (QALYs)

• lifetime care costs [NHS and Personal Social Services (PSS)].

However, the modelling may also usefully estimate the following outcomes in the short or long term:

• PCR

• time to progression

• OS

• response rates: cytogenetic, molecular and haematological

• time to treatment failure

• adverse effects of treatment.

Identification of studies

The search strategy comprised of the following main elements:

• searching of electronic databases

• contact with experts in the field

• scrutiny of bibliographies of retrieved papers and manufacturer submissions

• follow-up on mentions of potentially relevant ongoing trials noted in previous National Institute for Health and Clinical Excellence (NICE) guidance on imatinib for CML.

The main electronic databases of interest were MEDLINE (Ovid); EMBASE; The Cochrane Library including the Cochrane Database of Systematic Reviews (CDSR), Cochrane Central Register of Controlled Trials (CCRCT), Database of Abstracts of Reviews of Effects (DARE), NHS Economic Evaluation Database (NHS EED) and Health Technology Assessment (HTA) databases; National Research Register (NRR); Web of Science [including Conference Proceedings Citation Index (CPCI)]; Current Controlled Trials (CCT); ClinicalTrials.gov; FDA website; and the European Medicines Agency (EMA) website. These were searched from search end date of the last technology appraisal report on this topic, October 2002. 65

The searches were developed and implemented by a trained information specialist (CC) using the search strategy detailed in the technology appraisal by Thompson-Coon et al. 66 as the starting point (see Appendix 1 for full search strategy). This strategy was reviewed by PenTAG, including a clinical expert (CR).

Relevant studies were identified in two stages using predefined inclusion and exclusion criteria (see Appendix 2 for full research protocol). One reviewer (TP) examined all titles and abstracts, with two reviewers (TJ-H and LC) each examining approximately 50% each of all titles and abstracts (therefore all titles and abstracts were examined by at least two reviewers). Full texts of any potentially relevant studies were obtained. The relevance of each paper was assessed independently by two reviewers (TP and TJ-H) and any discrepancies resolved by discussion.

Inclusion and exclusion criteria

Data abstraction strategy

Data were extracted by one reviewer (TP) using a standardised data extraction form and checked independently by a second (TJ-H). Disagreements were resolved by discussion, with involvement of a third reviewer if necessary. Data extraction forms for each included study are included in Appendix 3.

Quality assessment strategy

The methodological quality of randomised controlled studies was assessed according to criteria specified by the CRD. 64 Quality was assessed by one reviewer (TP) and judgements were checked by a second (TJH or LC). Any disagreement was resolved by discussion, with involvement of a third reviewer if necessary.

Methods of data synthesis

Data were tabulated and discussed in a narrative review. Given the paucity of data, a meta-analysis was not conducted.

Mixed-treatment indirect comparisons were used as far as data allowed to facilitate comparison between the drugs for which there are no head-to-head data for dasatinib and nilotinib. From the data provided from the included trials, indirect comparisons are based on raw unadjusted results in the form of unadjusted odds ratios. The indirect log-odds ratio and corresponding variance were calculated using standard formulae presented in the appendix of Bucher et al. 67 Assuming the sampling distribution of the log-odds ratio to be normally distributed, the Wald method was used to construct 95% CIs for the odds ratio and calculate the p-value. A fixed-effect approach was used, which assumes that the relative effect of the interventions is the same across the two study populations. 67 To check this assumption we compared the baseline characteristics between trials. The participants were similar with respect to median age, the percentage of males, median time between diagnosis and randomisation, median white cell count and median platelet count. It was not possible to use more sophisticated methods (e.g. sensitivity analyses and subgroup analyses) to validate the assumption of similar relative effects, as we did not have access to the original data.

Handling company submissions to the National Institute for Health and Clinical Excellence

All clinical effectiveness data included in the pharmaceutical company submissions to NICE were assessed to see if they met the inclusion criteria and had not already been identified from published sources.

Identification of evidence

The electronic searches retrieved a total of 3227 titles and abstracts. Two additional papers were found by hand-searching of reference lists, with two papers retrieved from updated searches. No additional papers were found by searching the bibliographies of included studies. A total of 2510 papers were excluded on title and abstract. Full text of the remaining 35 papers was requested for more in-depth screening. The process is illustrated in detail in Figure 2.

FIGURE 2.

Flow diagram study of inclusion process for clinical effectiveness. EconLit, American Economic Association’s electronic bibliography.

Two clinical randomised controlled trials were included, one each studying dasatinib and nilotinib compared with imatinib (Table 5), with any additional abstracts or presentations related to the trials also included. 20,29 A further trial was identified, but was published only as a conference abstract. As sufficient detail was not available to make assessments of methodological quality, this was not formally included in the systematic review, with a summary of results available in Appendix 6. 68 Kantarjian et al. (dasatinib) provided an additional seven conference abstracts/presentations. 29,69–75 Saglio et al. (nilotinib) provided an additional 24-month follow-up paper and five conference abstracts/presentations. 20,76–81 One conference abstract of a systematic review assessing CML as a first-line treatment was identified and provided indirect comparison analysis of dasatinib and nilotinib. 82 Another paper also provided indirect comparison analysis of dasatinib and nilotinib. 83 Additional data were also retrieved from the industry submissions of Bristol-Myers Squibb (BMS, 2011, unpublished; dasatinib) and Novartis (2011, unpublished; nilotinib). 84,85

The details of studies retrieved as full papers and subsequently excluded, along with the reasons for their exclusion, are detailed in Appendix 5.

Assessment of effectiveness

Assessment of clinical effectiveness

Complete cytogenetic response

Cytogenetic responses are shown in Table 12. DASISION29 and ENESTnd20 report CCyR by 12, 18 and 24 months’ follow-up. DASISION29 reports confirmed CCyR (i.e. two assessments 28 days apart) for 12, 18 and 24 months’ follow-up, which ENESTnd20 does not. Both trials report CCyR by risk group categorisation by 12 months. CCyR is the primary outcome in the DASISION trial. 29

TABLE 12 - Complete cytogenetic response

PenTAG calculated.

Relative risk compared with imatinib.

ITT analysis.

Confirmed CCyR (i.e. two assessments 28 days apart).

Hasford risk-DASISION, Sokal risk-ENESTnd.

Shaded cells = not reported.

Study		DASISION29,73–75				ENESTnd20,79,81,85
Intervention		Dasatinib (100 mg)	Imatinib (400 mg)	p-value	RR (95% CI)a	Nilotinib (300 mg)	p-value	RR (95% CI)a,b	Nilotinib (400 mg)	p-value	RR (95% CI)a,b	Imatinib (400 mg)
CCyR rates 12 monthsc (%)		216/259 (83)	186/260 (72)	0.001	1.17 (1.06 to 1.28)	226/282 (80)	0.001	1.23 (1.11 to 1.36)	220/281 (78)	0.001	1.20 (1.08 to 1.34)	184/283 (65)
CCyR rates 18 monthsc (%)		218/259 (84)	203/260 (78)	0.093	1.08 (0.98 to 1.17)	240/282 (85)	< 0.001	1.15 (1.09 to 1.25)	230/281 (82)	0.017	1.11 (1.01 to 1.21)	209/283 (74)
CCyR rates 24 monthsc (%)		223/259 (86)	213/260 (82)	0.23	1.05 (0.97 to 1.13)	245/282 (87)	0.0018	1.13 1.04 to 1.22)	238/281 (85)	0.016	1.10 (1.01 to 1.19)	218/283 (77)
CCyR rates 12 months confirmedd		199/259 (77)	172/260 (66)	0.007	1.16 (1.04 to 1.30)
CCyR rates 18 months confirmedd (%)		202/259 (78)	182/260 (70)	0.037	1.11 (1.00 to 1.24)
CCyR rates 24 months confirmedd (%)		207/259 (80)	192/260 (74)	0.12	1.08 (0.98 to 1.19)
Risk group CCyR rates 12 monthse (%)	Low	81/86 (94)	66/87 (76)			(CiC information has been removed)			(CiC information has been removed.			(CiC information has been removed)
	Intermediate	97/124 (78)	88/123 (72)
	High	38/49 (78)	32/50 (64)			58/78 (74)			49/78 (63)			38/78 (49)
Risk group CCyR rates 18 months confirmedd,e (%)	Low	79/86 (92)	63/87 (72)
	Intermediate	88/124 (71)	87/123 (71)
	High	36/49 (73)	32/50 (64)
Risk group CCyR rates 24 months confirmede (%)	Low					94/103 (91)			97/103 (94)			94/104 (90)
	Intermediate					88/101 (87)			85/100 (85)			78/101 (77)
	High					63/78 (81)			56/78 (72)			46/78 (59)

Figures 3 and 4 summarise the CCyR data. We present these on two axes: available follow-up data (see Figure 3) and potential long-term survival (see Figure 4).

FIGURE 3.

Complete cytogenetic response (24 months) all patients.

FIGURE 4.

Complete cytogenetic response (10 years) all patients.

Dasatinib compared with imatinib

The DASISION trial29 reports that significantly more patients taking dasatinib (83%) achieved a CCyR than patients taking imatinib (72%) by 12 months’ follow-up [p = 0.001; relative risk (RR) 1.17, 95% CI 1.06 to 1.28]. 29 This difference was not significant by 18 months (84% vs 78%, p = 0.093; RR 1.08, 95% CI 0.98 to 1.17) or 24 months (86% vs 82%, p = 0.23; RR 1.05, 95% CI 0.97 to 1.13). 74,75 There was a significant difference for patients with a confirmed CCyR (i.e. two assessments, 28 days apart) by 12 months’ (77% vs 66%, p = 0.007; RR 1.16, 95% CI 1.04 to 1.30) and 18 months’ (78% vs 70%, p = 0.037; RR 1.11, 95% CI 1.00 to 1.24) follow-up. 29,73

By 24 months’ follow-up there was no significant difference for patients with a confirmed CCyR (80% vs 74%, p = 0.12; RR 1.08, 95% CI 0.98 to 1.19). 75 Differences between confirmed and non-confirmed CCyR suggest that more transitory responses may be seen with imatinib.

By 12 months’ follow-up, CCyR rates were higher for patients receiving dasatinib across all Hasford risk categories than with imatinib, with rates among those categorised as high risk of 78% and 64% for dasatinib and imatinib, respectively. 29 By 18 months’ follow-up, confirmed CCyR rates remained higher for patients receiving dasatinib across all Hasford risk categories compared with imatinib. 73

Nilotinib compared with imatinib (nilotinib 300 mg licensed for first-line treatment of chronic myeloid leukaemia)

The ENESTnd trial20 reports that significantly more patients taking nilotinib 300 mg (80%) achieved a CCyR than patients taking imatinib (65%) by 12 months’ follow-up (p = 0.001; RR 1.23, 95% CI 1.11 to 1.36). 20 By 18 months’ follow-up, rates of CCyR for nilotinib 300 mg and imatinib were 85% and 74%, respectively (p < 0.001; RR 1.15, 95% CI 1.09 to 1.25). 79 By 24 months, nilotinib 300 mg (87%) continued to be significantly superior to imatinib (77%) (p = 0.0018; RR 1.13, 95% CI 1.04 to 1.22). 81 For patients receiving nilotinib 300 mg, CCyR rates were higher across all Sokal risk categories than with imatinib by 12 and 24 months’ follow-up, with high-risk CCyR rates of 74% compared with 49% (12 months) and 81% compared with 59% (24 months) for nilotinib 300 mg and imatinib, respectively. 20,81,85

Major molecular response

Table 13 shows MMR in the two key trials. DASISION29 and ENESTnd20 report MMR at 12, 18 and 24 months’ follow-up. ENESTnd20 reports MMR at any time (12- and 24-month cumulative, MMR may be lost at specific time point). DASISION29 reports MMR at any time (12- and 18-month cumulative). Both trials report MMR by risk group categorisation at 12, 18 and 24 months. MMR is the primary outcome in the ENESTnd trial. 20

TABLE 13 - Major molecular response

PenTAG calculated.

Relative risk compared with imatinib.

ITT analysis.

Cumulative (MMR may be lost by time point).

Hasford risk, Kantarjian;29 Sokal risk, Saglio. 20

Shaded cells = not reported.

Study		DASISION29,73–75				ENESTnd20,79–81
Intervention		Dasatinib (100 mg)	Imatinib (400 mg)	p-value	RR (95% CI)a	Nilotinib (300 mg)	p-value	RR (95% CI)a,b	Nilotinib (400 mg)	p-value	RR (95% CI)a,b	Imatinib (400 mg)
MMR 12 monthsc (%)		119/259 (46)	73/260 (28)	< 0.001	1.63 (1.29 to 2.09)	125/282 (44)	0.001	2.02 (1.56 to 2.65)	121/281 (43)	0.001	1.97 (1.51 to 2.58)	62/283 (22)
MMR 18 monthsc (%)		145/259 (56)	96/260 (37)	< 0.001	1.52 (1.25 to 1.85)
MMR 24 monthsa (%)						175/282 (62)	< 0.001	1.67 (1.40 to 2.00)	165/281 (59)	< 0.001	1.58 (1.32 to 1.90)	105/283 (37)
MMR at any time (12 months)c,d (%)		135/259 (52)	88/260 (34)	< 0.001	1.54 (1.25 to 1.91)	(CiC information has been removed) (57)	(CiC information has been removed)	(CiC information has been removed)	(CiC information has been removed) (54)	(CiC information has been removed)	(CiC information has been removed)	(CiC information has been removed) (30)
MMR at any time (18-month)c,d(%)		148/259 (57)	107/260 (41)	< 0.001	1.39 (1.15 to 1.67)	186/282 (66)	< 0.001	1.65 (1.40 to 1.95)	174/281 (62)	< 0.001	1.55 (1.31 to 1.84)	113/283 (40)
MMR at any time (24-month)c,d (%)		166/259 (64)	120/260 (46)	< 0.001	1.39 (1.18 to 1.64)	201/282 (71)	< 0.001	1.63 (1.37 to 1.84)	187/281 (67)	< 0.001	1.52 (1.30 to 1.78)	124/283 (44)
Risk group MMR rates 12 monthse (%)	Low	48/86 (56)	31/87 (36)			(CiC information has been removed) (41)			(CiC information has been removed) (32)			(CiC information has been removed) (17)
	Intermediate	56/124 (45)	34/123 (28)
	High	15/49 (31	8/50 (16)
Risk group MMR rates 18 monthse (%)	Low	54/86 (63)	42/87 (48)			71/103 (69)			71/103 (69)			53/104 (51)
	Intermediate	69/124 (56)	49/123 (40)			69/101 (68)			63/100 (63)			39/101 (39)
	High	25/49 (51)	15/50 (30)			46/78 (59)			40/78 (51)			22/78 (28)
Risk group MMR rates 24 monthse (%)	Low	63/86 (73)	49/87 (56)			75/103 (73)			76/103 (74)			68/104 (65)
	Intermediate	76/124 (61)	62/123 (50)			75/101 (74)			67/100 (67)			44/101 (44)
	High	28/49 (57)	19/50 (38)			51/78 (65)			44/78 (56)			25/78 (32)

Figures 5 and 6 summarise the MMR data. We present these on two axes: available follow-up data (see Figure 5) and potential long-term survival (see Figure 6).

FIGURE 5.

Major molecular response (24 months), all patients.

FIGURE 6.

Major molecular response (10 years), all patients.

Dasatinib compared with imatinib

The DASISION trial29 reports that significantly more patients taking dasatinib (46%) achieved a MMR than those taking imatinib (28%) at 12 months’ follow-up (p < 0.0001; RR 1.63, 95% CI 1.29 to 2.09) and 18 months’ follow-up (56% vs 37%, p = 0.001; RR 1.52, 95% CI 1.25 to 1.85), and at 24 months’ follow-up. 29,74 A significant difference also seen for a MMR at any time at 12 months’ (52% vs 34%, p < 0.001; RR 1.54, 95% CI 1.25 to 1.91), 18 months’ (57% vs 41%, p = 0.001; RR 1.39, 95% CI 1.15 to 1.67) and 24 months’ follow-up (64% vs 46%, p = 0.001; RR 1.39, 95% CI 1.18 to 1.64)29,73,75

At 12 months’ follow-up, MMR rates were higher for patients receiving dasatinib across all Hasford risk categories than for patients receiving imatinib. 29 At 18 months’ follow-up, MMR rates remained higher for patients receiving dasatinib across all Hasford risk categories than for patients receiving imatinib, with MMR rates of 51% and imatinib 30% for dasatinib and imatinib, respectively, among those categorised as high risk. 73 At 24 months’ follow-up, MMR rates remained higher for patients receiving dasatinib across all Hasford risk categories than for patients receiving imatinib, with MMR rates of 57% and imatinib 38% for dasatinib and imatinib, respectively, among those categorised as high risk. 75

Nilotinib compared with imatinib (nilotinib 300 mg licensed for first-line treatment of chronic myeloid leukaemia)

The ENESTnd trial20 reports that significantly more patients receiving nilotinib 300 mg (44%) achieved a MMR than patients taking imatinib (22%) at 12 months’ follow-up (p = 0.001; RR 2.02, 95% CI 1.56 to 2.65). 20 At 24 months’ follow-up, MMR rates continued to be significantly higher for patients receiving nilotinib 300 mg (62%) than patients receiving imatinib (37%) (p = 0.001; RR 1.67, 95% CI 1.40 to 2.00). 80 A significant difference was also seen for a MMR at any time between nilotinib 300 mg and imatinib at 12 months’ [57% vs 30%, (CiC information has been removed)], 18 months’ (66% vs 40%, p < 0.001; RR 1.65, 95% CI 1.40 to 1.95) and 24 months’ follow-up (71% vs 44%, p = 0.001; RR 1.63, 95% CI 1.37 to 1.84). 79,81,85

(CiC information has been removed.) At 18 months’ follow-up, MMR rates were higher for patients receiving nilotinib 300 mg across all Sokal risk categories than for those receiving imatinib, with MMR rates of 59% and 28% for nilotinib 300 mg and imatinib, respectively, among those categorised as high risk. 79 At 24 months’ follow-up, MMR rates remained higher for patients receiving nilotinib 300 mg across all Sokal risk categories than for patients receiving imatinib, with MMR rates of 65% and 32% for nilotinib and imatinib, respectively, among those categorised as high risk. 81

Survival

This section reports on OS, PFS and EFS. PFS is usually defined as all cause death or progression to AP/BC, but definition may be subjective to the trial. EFS is defined by the researchers of the trials and usually includes all cause death, progression to AP/BC and loss of response. Results and details of the trial survival definitions are shown in Table 15. Figures 7 and 8 summarise the OS data.

TABLE 15 - Survival (progression free, event free, overall)

Defined as death from any cause, progression to AP/BC, loss of CCyR, loss of partial CyR or loss of CHR. 77

Defined as no progression, failure or intolerance.

Progression defined as a doubling of white cell count to more than 20 × 10⁹, absence of CHR, increase in Ph+ metaphases to more than 35%, progression to AP/BC, death from any cause. 29

PFS defined as progression to AP/BC or death by any cause. 20

Shaded cells = not reported.

Study	DASISION29,73,75			ENESTnd77,79,81
Intervention	Dasatinib (100 mg)	Imatinib (400 mg)	p-value	Nilotinib (300 mg)	p-value	Nilotinib (400 mg)	p-value	Imatinib (400 mg)
EFSa 12 months				97.6%	0.09	99.6	0.001	95.7%
EFSa 24 months				96.4%	0.12	97.8	0.01	93.6%
EFSb 24 months	84.8%	83.8%
PFSc 12 months	96%	97%
PFSc 18 months	94.9%	93.7%
PFSc,d 24 months	93.7%	92.1%		98%	0.07	97.7%	0.04	95.2%
OS 12 months	97%	99%
OS 18 months	96%	97.9%		98.5%	0.28	99.3%	0.03	96.9%
OS 24 months	95.3%	95.2%		97.4%	0.64	97.8%	0.21	96.3%

FIGURE 7.

Overall survival (24-month axis).

FIGURE 8.

Overall survival (10-year axis).

We present these on two axes, available follow-up data (see Figure 7) and potential long-term survival (see Figure 8), as an indication of the immaturity of these data in relation to expected long-term survival.

Dasatinib compared with imatinib

The DASISION trial29 reports (see Table 15) that PFS and OS were not statistically different between dasatinib and imatinib at 12 months (PFS 96% vs 97%; OS 97% vs 99%) 18 months (PFS 94.9% vs 93.7; OS 96 vs 97.9%) and 24 months’ follow-up (PFS 93.7% vs 92.1%; OS 95.3% vs 95.2%), as calculated by PenTAG. 29,73,75

Nilotinib compared with imatinib (nilotinib 300 mg licensed for first-line treatment of chronic myeloid leukaemia, nilotinib 400 mg licensed for second-line treatment of chronic myeloid leukaemia)

At 12 months’ follow-up, the ENESTnd trial20 reports (see Table 15) no significant difference in EFS compared with imatinib (95.7%) for nilotinib 300 mg (97.6%; p = 0.09) but significantly higher EFS for nilotinib 400 mg (99.6%,; p = 0.001), with differences maintained at 24 months’ follow-up. 77,81 PFS at 24 months was also not significantly different for nilotinib 300 mg (98%; p = 0.07) but significantly higher for nilotinib 400 mg (97.7%; p = 0.04) compared with imatinib (95.2%). 81

At 18 months, OS was not significantly different for nilotinib 300 mg (98.5%; p = 0.28) and significantly higher for nilotinib 400 mg (99.3%; p = 0.03) than for imatinib (96.9%). 79 At 24 months OS was not significantly different for either dose of nilotinib compared with imatinib (97.4%, p = 0.64; 97.8%, p = 0.21; 96.3%, respectively). 81

Adverse events

Results for AEs are shown in Tables 16 and 17 for the DASISION29 and ENESTnd20 trials, respectively. Both trials report the measurement of similar haematological and non-haematological events, with ENESTnd20 also reporting biochemical abnormalities.

Health-related quality of life

Health-related quality of life was not reported in the DASISION29 or ENESTnd20 trials.

Indirect comparison of dasatinib and nilotinib

No trials compared dasatinib and nilotinib head to head. However, an indirect comparison of nilotinib to dasatinib was carried out using results from the DASISION29 and ENESTnd trials. 20

The primary outcomes reported are MMR at 12 months and CCyR by 12 months. Because the DASISION trial29 reported CCyR as well as confirmed CCyR, two sets of results are reported for the CCyR outcome. 29 As shown in Table 18, there was no difference between dasatinib and nilotinib for CCyR, MMR or CMR rates for 12 months’ or 24 months’ follow-up.

Oxford Outcomes conducted an indirect comparison of dasatinib and nilotinib based on the data from the DASISION and ENESTnd trials. 20,29,82 The indirect results for 12 months’ follow-up showed no statistical difference between dasatinib and nilotinib for CCyR or MMR data.

(CiC information has been removed.)

Signorovitch et al. 85 report on the indirect comparison of the DASISION29 and ENESTnd20 trials, with individual patient data for patients receiving nilotinib (ENESTnd20). 85 Individual patient data for patients receiving 300 mg nilotinib were weighted to match the baseline characteristics reported for patients receiving dasatinib including age, gender, ECOG performance and haematology laboratory values. After matching patients receiving 300 mg, nilotinib, compared with dasatinib, had significantly higher rates of MMR (56.8% vs 45.9%; p = 0.001) and OS (99.5 vs 97.3; p = 0.046). CCyR was not assessed due to different measurement procedures of the trials. We have analysed CCyR as we believe they are sufficiently similar to warrant comparison.

Identification of studies

The search strategy comprised of the following main elements:

• searching of electronic databases

• scrutiny of bibliographies of retrieved papers and manufacturer submissions.

The following databases were searched: MEDLINE (Ovid), EMBASE, The Cochrane Library (including the CDSR, CCTR, DARE, NHS EED and HTA databases), NRR, Web of Science (including Conference Proceedings); CCT; ClinicalTrials.gov, FDA website and the EMA website. These were searched from search end date of the last technology appraisal report on this topic of October 2002. 65

The searches were written by CC with advice from TP, RA, RT, OC and RG. The surrogate terms circulated were cross-checked against a previous review of surrogate outcomes and CR for sensitivity and inclusion. 90

Inclusion and exclusion criteria

Data extraction strategy

Study characteristics and surrogate/final outcome data were extracted by one reviewer (OC) using a standardised data extraction form and independently checked by a second (TP or RT). Data digitalisation software (WinDIG, version 2.5 WinDIG, Geneva, Switzerland) was used to extract data from Kaplan–Meier survival curves. Disagreements were resolved by discussion, with involvement of a third reviewer if necessary. Data extraction forms for each included study are included in Appendix 4.

Quality assessment strategy

The methodological quality of included studies was assessed according to a modified list of criteria specified by the CRD. Quality was assessed by one reviewer (OC) and judgements were checked by a second (TP or RT).

Methods of data synthesis

An initial review of included studies revealed two key limitations. First, there was a lack of data reported to assess the trial level association between TKI treatment effects on CCyR and TKI treatment effect on patient-relevant outcome. This would be needed for high-level evidence of surrogacy. Second, there was no presentation of data of the association of CCyR or MMR and HRQoL. It was therefore decided to focus on studies that reported OS and/or PFS stratified by either CCyR or MMR. 37,89

For each study, levels of OS and PFS were extracted by response stratum at each year following trial recruitment (or randomisation) up to the latest follow-up point reported. In most studies OS and PFS data were reported in Kaplan–Meier curves using landmark analysis to evaluate differences in the final patient-relevant outcomes between responder and non-responders. The landmark method determines each patient’s response at a fixed time point, with survival estimates calculated from that time point and associated statistical tests being conditional on patients’ landmark responses. (In the included papers, the survival probabilities referred to the starting point of the treatment rather than to the time of response.) Note that in this method, patients who die before the landmark time point are excluded from the analysis. 91

We selected 12 months after the start of first-line TKI therapy as the landmark for our analysis, as the DASISION29 and ENESTnd20 trials consider, respectively, the rate of MMR and confirmed CCyR at 12 months after randomisation as primary end points. 20,29 A weighted average of the OS and PFS at different yearly intervals was estimated for both the responders and non-responders by taking into account the initial number of patients in the two groups. Wilson 95% CIs were derived for each point estimate assuming binomial distributed variables and no censoring of data. 92 Analyses were carried out using Stata^© version 11.2 (StataCorp, TX, USA).

Identification of evidence

The electronic searches retrieved a total of 5033 titles and abstracts. Two papers were found by updated databases searches, and three were identified by reviewers through hand-searching and/or referenced in industry submissions. These papers were then excluded based on title and abstract because the population was treated with IFN or they were conference abstracts. 93–95

After de-duplication, 3555 papers were screened and the majority of them were excluded on title and abstract. Full text of the remaining 63 papers was requested for more in-depth screening. The process is illustrated in detail in Figure 9. The last step of the process was the inclusion of selected papers in the quantitative analysis for the assessment of both CCyR and MMR as surrogate measure for OS and PFS. Where a study had been reported in several publications, it was considered only once according to the type of relationship reported (CCyR and/or MMR vs OS and/or PFS) and the paper reporting maximum follow-up was used. One study from India was deemed not externally valid in portraying the UK patients with CML population and treatment response. 96 Although reporting on OS and PFS stratified by level of CyR, this study was excluded from the quantitative analysis. The details of studies excluded on full review, along with the reasons for their exclusion, are detailed in Appendix 5.

FIGURE 9.

Flow diagram study inclusion process surrogate outcomes.

Assessment of surrogate evidence

Survival by level of cytogenetic response

Figure 10 shows the weighted pooled OS (95% CI) at yearly intervals after the initiation of imatinib treatment by CyR. Three publications provided data for the estimates. 97,102,104 The impact of failing to achieve a CCyR at 12 months becomes increasingly apparent over time, with increasing differences in OS between those who respond and those who do not. No non-responder group data at 48 months are reported. It was decided not to include the non-responder group data from Kantarjian et al. ,102 because they included five patients who developed a minor CyR at 12 months.

FIGURE 10.

Pooled weighted average (95% CI) of OS by level of CyR at yearly intervals after first-line imatinib initiation.

The weighted average of the PFS by CCyR at 12 months at yearly intervals after the initiation of imatinib therapy is shown in Figure 11. The estimates were obtained by the three papers that reported PFS across groups with different level of CyR. 97,99,102 The plotted values and the uncertainty around the estimates of OS and PFS by CyR are given in Table 24.

TABLE 24 - Pooled weighted average of overall and PFS (95% CI) by level of CyR at 12 months after the starting of imatinib therapy

Time (months)	OS% (95% CI)		PFS% (95% CI)
	CCyR	No CCyR	CCyR	No CCyR
12	100 (100 to 99.3)	100 (100 to 98.1)	100 (100 to 99.3)	98.9 (99.8 to 94)
24	98.1 (98.9 to 96.5)	94 (96.5 to 89.7)	98.8 (99.4 to 97.4)	94.3 (97.6 to 87.7)
36	97.5 (98.5 to 95.9)	89 (92.6 to 83.8)	97.6 (98.5 to 95.9)	85.5 (91.4 to 77.1)
48	98 (99.3 to 95.3)	Not reported	97.6 (98.5 to 95.9)	85.5 (91.4 to 77.1)
60	97.4 (98.6 to 94.9)	74.1 (82.4 to 62.4)	96.8 (97.9 to 95)	75.2 (82.5 to 64.9)
72	Not reported	Not reported	95.5 (97.0 to 93.1)	80 (91.5 to 56.7)

FIGURE 11.

Pooled weighted average (95% CI) of PFS by level of CyR at yearly intervals after first-line imatinib initiation.

Survival by level of molecular response

Figure 12 shows the weighted average OS at yearly intervals after the start of the first-line imatinib therapy for CP-CML by level of molecular response. Three publications provided data for the estimates. 98,100,102 It is worth highlighting Hehlmenn et al. ,105 considered in the OS curves by landmark analysis of MMR at 12 months for the whole study population (n = 848) because, independent of the treatment approach (imatinib 400 mg/day, imatinib 800 mg/day, imatinib 400 mg/day + IFN-α), they found that MMR compared with no MMR at 12 months was associated with better PFS (99% vs 95%, p = 0.0143 at 3 years) and OS (99% vs 95%, p = 0.0156 at 3 years). Consistent with the weighting approach used in this report, the number of units involved in the construction of Kaplan–Meier curves, the overall sample size population in this case, was considered.

FIGURE 12.

Pooled weighted average (95% CI) of OS by level of molecular response at yearly intervals after first-line imatinib initiation.

The pooled PFS by MMR at 12 months after the starting of imatinib therapy for CP-CML is shown in Figure 13. The estimates are derived from three publications that reported on PFS for groups of patients presenting different levels of molecular response. 98,100,102

FIGURE 13.

Pooled weighted average (95% CI) of PFS by level of molecular response at yearly intervals after first-line imatinib initiation.

No non-responder PFS estimate at 72 months after therapy initiation was reported. The IRIS report26 by Hughes et al. 100 shows PCR curves by BCR–ABL transcript levels at 12 months converted to the International Scale (IS) (i.e. ≤ 0.1%, > 0.1% to ≤ 1%, > 1% to ≤ 10%, >10%) up to 84 months’ follow-up. These curves provide data for the PFS for patients achieving MMR at 12 months, defined as ≤ 0.1% IS, but not for the cumulative group of patients who do not achieve MMR at 12 months. The same authors give a tabulated value for the 7-year PFS in patients with no MMR at 12 months’ landmark time. The plotted values and the uncertainty around the estimates of OS and PFS by level of molecular response are given in Table 25.

TABLE 25 - Pooled weighted average of overall and PFS (95% CI) by level of molecular response at 12 months after the starting of imatinib therapy

Time (months)	OS% (95% CI)		PFS% (95% CI)
	MMR	No MMR	MMR	No MMR
12	100 (100 to 99.1)	100 (100 to 99.4)	100 (100 to 98.5)	99.6 (99.9 to 97.8)
24	100 (100 to 99.1)	96.7 (97.9 to 95)	99.2 (99.8 to 97.1)	94 (97.3 to 87.9)
36	99.2 (99.8 to 97.9)	95.7 (97.1 to 93.8)	98.6 (99.3 to 97.3)	94.3 (95.8 to 92.1)
48	96.7 (97.9 to 94.4)	93.3 (95.0 to 91.0)	96.6 (98.3 to 93.7)	91 (95.4 to 84.4)
60	96.6 (97.9 to 94.9)	91.2 (93.2 to 88.6)	95.8 (97.8 to 92.7)	89 (93.9 to 82.0)
72	92.5 (95.9 to 87.6)	90 (92.3 to 87.0)	99 (99.6 to 95.3)	Not reported
84	96 (97.5 to 93.2)	89.2 (93.4 to 83.5)	99 (99.6 to 95.3)	89.9 (93.9 to 84.2)

Summary of surrogate outcomes

In summary, there is observational association evidence supporting the use of CCyR and MMR at 12 months as surrogates for OS and PFS in patients with CML in CP. This is based entirely on imatinib treatment studies. In the absence of evidence of adequacy of these surrogates for dasatinib and nilotinib as first-line therapies for CP-CML, assuming a TKIs class-specific relationship between the surrogate outcomes and the patient-relevant outcomes, these results can be potentially applied to other drugs in the same class.

Scope of the submission

The submission from BMS considers the use of dasatinib for the first-line treatment of people with CML as an alternative to the standard dose of imatinib (400 mg daily) or nilotinib (600 mg daily).

The clinical effectiveness outcomes considered are:

• OS

• PFS

• response rates

• adverse effects of treatment

• HRQoL.

The outcomes for the economic analysis are:

• incremental cost per QALY

• incremental cost per life-year gained.

In order to derive these outcomes, the following costs have been considered:

• cost of first- and second-line TKIs

• cost of second- or third-line treatment post TKI failure

• the cost of treating serious AEs.

The time horizon for the economic analysis is between 46 and 86 years, and costs are considered from an NHS perspective. No subgroup analysis is conducted for the economic evaluation.

Summary of submitted cost-effectiveness evidence

The manufacturer uses a ‘time in state’ (area under the curve) model extrapolating CML-related survival and PCR data. The health states represent the CP and AP/blast phases, as well as death. Within the CP, patients may also be in first-, second- or third-line treatment, whereas in the AP/blast phase they may be receiving either third-line treatments or palliative care. Time is modelled in blocks of 1 month (Figure 14).

FIGURE 14.

Bristol-Myers Squibb model structure. Source: Figure 5, p. 40 of BMS submission.

Bristol-Myers Squibb have modelled one scenario with three different comparators. The interventions and sequence of treatments are summarised in Table 27.

The BMS base-case analysis produces incremental cost-effectiveness ratios (ICERs) of (Table 28):

• £26,000 per QALY for dasatinib in comparison with imatinib as first-line TKI, and

• £145,000 per QALY for nilotinib compared with dasatinib (nilotinib provides more benefit at greater cost than dasatinib) as a first-line TKI.

The sensitivity analysis shows the key parameters to which the model is sensitive:

• drug costs

• OS

• the cost of SCT.

The BMS model contained a number of formula errors. After correcting for these errors the model predicts ICERs of:

• £36,000 per QALY for first-line dasatinib compared with first-line imatinib, and

• £103,000 per QALY for dasatinib compared with nilotinib (dasatinib provides more benefit at greater cost than nilotinib).

In the original model, the cost of nilotinib used by BMS does not account for the Patient Access Scheme (PAS) discount applied to nilotinib.

[Academic-in-confidence (AiC) information has been removed.]

Including this change, the BMS model predicts an ICER of £45,600 per QALY for dasatinib compared with imatinib. When comparing dasatinib with nilotinib, the model predicts that nilotinib is more effective and less costly.

Furthermore, BMS assume that dasatinib is taken as a third-line treatment in all treatment arms. However, in the NICE draft guidance final appraisal determination, dasatinib was not recommended (the draft guidance FAD for second-line, high-dose imatinib, dasatinib and nilotinib for CML is available on the NICE website at http://guidance.nice.org.uk/TA/WaveR/99). (In the draft guidance on 18 August 2011, NICE has recommended nilotinib for the treatment of the chronic and APs of CML that is resistant or intolerant to standard-dose imatinib. Dasatinib and high-dose imatinib are not recommended in the draft guidance. Consultees have the opportunity to appeal against the draft guidance. Until NICE issues final guidance, NHS bodies should make decisions locally on the funding of specific treatments. This draft guidance does not mean that people currently taking dasatinib or high-dose imatinib will stop receiving them. They have the option to continue treatment until they and their clinicians consider it appropriate to stop.) When the BMS model is adjusted so that dasatinib is not taken third line, the ICER of dasatinib compared with imatinib increases further, from £45,600 to £64,000 per QALY, and nilotinib is still more effective and less costly than dasatinib.

Finally, BMS assumes that half of all patients in the imatinib and nilotinib treatment arms who are eligible for second-line treatment take dasatinib. Again, in the NICE draft guidance final appraisal determination, dasatinib was not recommended (the draft guidance FAD for second-line, high-dose imatinib, dasatinib and nilotinib for CML is available on the NICE website at http://guidance.nice.org.uk/TA/WaveR/99). When the BMS model is adjusted so that dasatinib is not taken second line, and instead when we assume that all second-line patients in the imatinib arm take nilotinib second line, the ICER of dasatinib compared with imatinib increases further, from £64,000 to £96,000 per QALY. There appears to be no simple way to adjust the BMS model to disallow for patients taking dasatinib second line.

In summary, the BMS adjusted model yields an ICER for dasatinib compared with imatinib of £96,000 per QALY. Furthermore, nilotinib is more effective and less costly than dasatinib.

Commentary on the robustness of the submitted evidence

Areas of uncertainty

The BMS model does not provide the raw data that were used to fit the OS and time to treatment discontinuation curves. However, the choice of distribution and coefficients of the distribution appear to be correct on the basis of graphs showing the observed data and the fitted curves.

A considerable area of uncertainty is the chosen sequence of second-line TKI treatments that might follow failure of different first-line TKIs. This is partly because the submission was prepared before NICE’s draft guidance FAD on the use of dasatinib, nilotinib or high-dose imatinib as second-line treatments (the draft guidance FAD for second-line, high-dose imatinib, dasatinib and nilotinib for CML is available on the NICE website at http://guidance.nice.org.uk/TA/WaveR/99). However, uncertainty also results from the fact that data on the effectiveness of second-line TKI treatments are only available following the use of imatinib as first-line treatment.

Key issues

• The BMS model does not use the cost of nilotinib agreed under the PAS in the submission. However, it is acknowledged that BMS was unable to account for the discount, as it did not have knowledge of the PAS discount at the time of its submission.

• The BMS model is structured in such a way that it would require significant changes to run it without second-line treatment, should this be required by NICE.

• The time horizon chosen by the BMS model does not reflect the lifetime of a patient with CML. In the model, nearly 20% of the population is still alive in the last cycle (86 years old), suggesting that the model overestimates the period that those with CML will survive.

• The BMS model has a number of formula errors, correcting for which impacts on the ICER.

• The cost and proportions of patients who receive SCT have a significant impact on ICERs, but the source of the BMS estimates of these parameters is unclear. Clinical opinion is required to assess whether or not the BMS assumption on the provision and costing of SCT is appropriate.

Scope of the submissions

The submission from Novartis considers the use of nilotinib for the first-line treatment of people with CML as an alternative to the standard dose of imatinib (400 mg daily). In one analysis, dasatinib is used in the cost-effectiveness model as second-line treatment when first-line treatment with imatinib or nilotinib fails. In the other analysis, no second-line TKIs are assumed.

The clinical effectiveness outcomes considered are:

• PFS

• time to discontinuation

• adverse effects of treatment

• HRQoL.

The outcomes for the economic analysis were:

• incremental cost per QALY

• incremental cost per life-year gained.

In order to derive these outcomes the following costs were estimated in the model:

• cost of first- and second-line TKIs,

• cost of post-TKI failure second- or third-line treatment

• the cost of treating AEs.

The time horizon for the economic analysis is lifetime and costs are considered from the NHS perspective.

The Novartis cost-effectiveness modelling reflects a cost discount (PAS for the cost of first-line nilotinib) (CiC information has been removed). Its cost of second-line nilotinib also reflects this cost discount (also a PAS). No subgroup analyses are conducted for the economic evaluation, although a policy scenario without the use of second-generation TKIs is simulated.

Summary of submitted cost-effectiveness evidence

The manufacturer uses a Markov approach to model the cost-effectiveness of nilotinib compared with the current standard of care (imatinib 400 mg daily). This model has nine states. Patients enter the model in the CP. The model estimates when one treatment fails and hence the patient is switched to an alternative treatment. At the end of each cycle, patients have a probability of remaining on current treatment, progressing to an alternative treatment or dying (Figure 15).

FIGURE 15.

Novartis model structure. Source: Novartis industry submission (p. 82).

Novartis modelled two different scenarios to reflect the availability or not of second-generation TKIs as second-line treatment. The interventions and sequence of treatment are summarised in Table 29.

The Novartis model predicts that nilotinib is both more effective and less costly than imatinib (dominates) when followed by dasatinib as second-line treatment. In a scenario analysis without dasatinib as second-line treatment, the model predicts an ICER of £5908 per QALY for nilotinib in comparison with imatinib (Table 30).

The sensitivity analysis shows the key parameters to which the cost-effectiveness results are sensitive to are:

• drug costs (i.e. without PAS)

• time to discontinuation of first-line TKI.

No major formula errors have been identified in the Novartis model.

Commentary on the robustness of submitted evidence

Areas of uncertainty

The Novartis model does not provide the raw data which were used to fit the OS and time-to-treatment discontinuation curves. However, the choice of distribution and coefficients of the distribution appear to be correct on the basis of graphs showing the observed data and the fitted curves.

Another area of uncertainty is the chosen sequence of second-line TKI treatments that might follow the failure of different first-line TKIs. This is partly because this submission was prepared before NICE’s draft guidance FAD on the use of dasatinib, nilotinib or high-dose imatinib as second-line treatments (the draft guidance FAD for second-line, high-dose imatinib, dasatinib and nilotinib for CML is available on the NICE website at http://guidance.nice.org.uk/TA/WaveR/99). However, uncertainty also results from the fact that data on the effectiveness of second-line TKI treatments are only available following the use of imatinib as first-line treatment.

Another area of uncertainty is regarding the cost and utility of stem cell patients. Assumptions around SCT have a significant impact on the model. Novartis uses a one-off cost of £99,224 for each transplant with a post-transplant utility for survivors of 0.813.

Key issues

• Novartis uses a PAS for pricing nilotinib as first-line treatment. This has a significant impact on the results.

• Novartis makes no use of the major molecular and CCyR rates from the RCT of nilotinib compared with imatinib, both of which are important indicators of clinical effectiveness.

• The cost and the proportions of patients who receive SCT differ between the Novartis and BMS models and they have a significant impact on ICERs. Clinical opinion is required to assess whether or not the BMS assumption on the provision and costing of SCT is appropriate.

Peninsula Technology Assessment Group model structure

The PenTAG cost-effectiveness model is a state-transition model with states for the main disease phases, and for the different possible treatments within the CP. It is very similar to the Novartis model in terms of states and allowable transitions.

Patients enter the model in the CP. During each model cycle, a patient is assumed to be in one of the health states. In Figure 16, arrows represent possible transitions between health states. At the end of each cycle, patients have a probability of remaining on their health state (shown by circular arrows), progressing to an alternative state or dying (see Figure 16). In scenarios 3 and 4, after first-line treatment failure, patients in the imatinib and dasatinib treatment arms progress to second-line nilotinib, as shown in Figure 16 by the dotted ellipse. Patients in the nilotinib arm progress directly to hydroxycarbamide or SCT. In scenarios 1 and 2, all patients progress directly to hydroxycarbamide or SCT after first-line TKI.

FIGURE 16.

Structure of PenTAG cost-effectiveness model.

Cumulative survival approach

In this approach, OS for each treatment arm is estimated in scenarios 1 and 2 (see Table 31) as the sum of time on first-line treatment and OS following either hydroxycarbamide or SCT. For scenarios 3 and 4, OS for the nilotinib comparator is as for scenarios 1 and 2, whereas OS for the imatinib and dasatinib arms equals the sum of time on first-line treatment, time on second-line nilotinib and OS following either hydroxycarbamide or SCT. This general approach is the same as that used by Novartis (see Comparison of PenTAG model with Novartis model, below). The method ignores the CCyR and MMR response rates from the two main RCTs of first-line dasatinib compared with imatinib and first-line nilotinib compared with imatinib.

An important assumption in this approach is that OS after second-line nilotinib and OS after hydroxycarbamide or SCT are independent of previous treatment. In the cumulative survival method, the time-independent annual transition probability from CP to AP, while people take hydroxycarbamide, was assumed to be the same for all three treatment arms.

Surrogate-predicted survival approach

In this approach, OS for all three treatment comparators is estimated using a surrogate relationship based on MMR at 12 months on first-line TKI (scenarios 1a and 2a), and in a separate analysis, on CCyR at 12 months (scenarios 1b and 2b). The methods of estimating OS based on the surrogate relationships with MMR and CCyR are described below [see Surrogate-predicted overall survival (for surrogate survival only)] and are based on the results of our clinical effectiveness systematic review and meta-analysis of surrogate outcomes (see Chapter 4, Analysis of overall survival and progression-free survival by cytogenetic and molecular response).

Modelling for these scenarios uses only the proportion of patients with or without a response at 12 months. We also assume that, for a given response rate, OS is independent of first-line treatment comparator.

The model does not reflect possible differences in the depth, speed of achieving or duration of a response. Given that (1) dasatinib and nilotinib are believed to be superior to imatinib in all these respects (see Novartis and RCP submissions and our clinical effectiveness systematic review, see Chapter 3, Overall clinical effectiveness conclusions) and (2) the historical surrogate data are all based on OS for patients taking imatinib, it is possible that this method underestimates OS for dasatinib and nilotinib but the extent of this is unquantifiable.

The BMS modelling also predicts OS by a surrogate relationship based on CCyR, but not on MMR. Novartis does not model OS by a surrogate method, instead using only the cumulative survival method (see Comparison of PenTAG model with Novartis model and Comparison of PenTAG model with BMS model, below, comparing the PenTAG with the manufacturers’ economic analyses). Our scenario analyses, which make use of the surrogate-based survival relationships, do so by adjusting the analyses based on the cumulative survival approach. In the following paragraphs, we describe the adjustments to the cumulative survival model that are needed to reflect the surrogate OS for the three treatment arms estimated below [see Surrogate-predicted overall survival (for surrogate survival only)].

It is not surprising that OS for each treatment arm under the cumulative survival method is different to OS as predicted by the MMR and CCyR surrogate relationships, given that the cumulative survival method relies on numerous assumptions that have a cumulative impact. Specifically, in the Results section (see PenTAG cost-effectiveness results, below) we show that OS under the cumulative survival method is far shorter, at approximately 16–18 years, than under the surrogate survival methods, at approximately 21–23 years. We are then faced with the decision of how to adjust the model, which is based on the cumulative survival method, so that it predicts OS specific to each treatment as estimated by the surrogate survival method. BMS achieved this by leaving unaltered the transition probabilities under the cumulative survival method, but setting the transition probabilities that determine the times in AP and BC as the ‘balancing items’, so as to achieve the surrogate OS experienced in historical trials of imatinib.

We ruled out this approach because this would result in unrealistically long mean times in AP plus BC of approximately 5–8 years, the difference between typical OS predicted under the surrogate relationship and typical OS under the cumulative survival method. In practice, typical times in these advanced disease states are believed to be 6 months to 1 year (see Time on hydroxycarbamide in chronic phase, and time in accelerated phase and blast crisis, below).

For the transition probabilities in the PenTAG model, the mean times corresponding to first-line TKIs and second-line nilotinib were not altered (from their cumulative survival model values) because they are informed by good evidence from high-quality trials. This left three choices:

1. adjust the annual transition probability from CP to AP while people take hydroxycarbamide

2. adjust mortality after the SCT operation

3. or some combination of the above.

These choices seemed plausible given that the corresponding transition probabilities are informed by poorer-quality evidence. The third option was ruled out as too complex. The second option was ruled out because even if we assumed that all patients are completely cured after SCT then the modelled OS is still shorter than OS predicted from the surrogate relationships. The first option was selected as it was possible to model OS from the surrogate relationship. Under the surrogate survival method, this probability was unique for each treatment arm.

A pair of analyses was performed for each of scenarios 1a, 1b, 2a and 2b (see Table 31). First, the transition probability from CP to AP while people take hydroxycarbamide, unique to each treatment arm, was set to precisely match the mean OS from the appropriate surrogate relationship. Second, the transition probability from CP to AP for the imatinib treatment arm was left unadjusted (as in the cumulative survival method) but the transition probabilities from CP to AP for the nilotinib and dasatinib treatment arms were adjusted so as to model the differences in OS between treatment arms from the surrogate OS. These adjustments are shown graphically in the Results section (see Figure 42). The purpose of the second analysis was to capture the essence of OS estimated by the historical surrogate data, which is the magnitude of the difference in OS according to response (see Chapter 4).

Simplified method

In this simplified approach (used in scenarios 2, 2a, 2b and 4 in Table 31), the post-TKI (first-line TKIs and second-line nilotinib) per-patient costs and QALYs are set to be equal across treatment arms. The costs and QALYs while patients are on TKIs are modelled specific to each treatment arm, exactly as normal. However, because slightly different proportions of patients will have died during the time when they are taking first- or second-line TKIs, there will still be small differences in the total costs and QALYs accrued after this time point between the treatments compared.

Specifically, suppose the total discounted per-patient post-TKI treatment cost in the imatinib treatment arm is given as C_im, and suppose the proportion of patients who are still alive and start second- or third-line treatment on hydroxycarbamide or SCT in the imatinib and nilotinib treatment arms are P_im and P_nil, respectively (which are calculated from scenarios 1 or 3; see Table 31). Then we estimate the total discounted per-patient post-TKI treatment cost in the nilotinib treatment arm as:

and similarly for the dasatinib treatment arm. The total discounted per-patient post-TKI treatment QALYs in the nilotinib and dasatinib treatment arms are calculated similarly. In the Results section, we show that the proportions still alive and starting second- or third-line treatment on hydroxycarbamide or SCT are similar across the three treatment arms, as the durations of TKI treatments are similar across treatments. Therefore, this method largely equalises all post-TKI costs and QALYs between treatment arms.

One further adjustment is performed when using the simplified analysis method in combination with the surrogate survival method (scenarios 1a, 1b, 2a, 2b; see Table 31). In these scenarios, relative survival between treatment arms is modelled by setting the time on hydroxycarbamide in CP as a function of treatment arm in order to recreate the modelled OS based on surrogate data (see Surrogate-predicted survival approach, above). In this case, if denoting T_im and T_nil as the mean times on hydroxycarbamide in CP for those patients who receive this treatment in the imatinib and nilotinib treatment arms, respectively, then we estimate the total discounted per-patient post-TKI treatment cost in the nilotinib treatment arm as:

and similarly for the dasatinib arm, and for the QALYs in the dasatinib and nilotinib treatment arms.

This simplified method clearly does not represent our best estimate of the courses of treatments after resistance or intolerance to TKIs. However, we include this scenario to represent largely the cost-effectiveness of the treatment arms allowing for the ‘pure’ cost-effectiveness of first-line TKIs and second-line nilotinib. Also, we believe this analysis may be useful given the substantial uncertainty in the nature and costs of subsequent lines of treatment. This is especially true given that we predict that patients will take first-line TKIs for many years (between 7 and 9 years, see Results). Therefore, in the simplified method analysis the results should not reflect the treatments post TKIs (which remain uncertain, and also start about 8 years from diagnosis) and their associated costs and QALYs.

Perspective, discounting and time horizon

The model cycle length is 3 months, and the model time horizon is 50 years, or age 107 years, at which time all people have died. A model half-cycle correction is applied.

Future costs and benefits (QALYs) are discounted at 3.5% per year, and the perspective is that of the NHS and PSS, in accordance with the NICE reference case.

Modelled treatment sequences post first-line treatment

Summary of scenario analyses

The relative merits of our scenario analyses are presented in Table 32.

In the context of a MTA, a secondary purpose of the economic model produced by the independent technology assessment/review group is to enable consideration and comparison of the similarities and implications of the modelling approaches used by the manufacturers.

Table 33 shows the scenarios analysed with the PenTAG model and how they relate to the model-based cost-effectiveness analyses provided by Novartis and BMS.

Note that the Novartis analysis in the last row of this table (their ‘base-case scenario’) is probably no longer valid, given that in the NICE draft guidance FAD, dasatinib as second-line treatment for CML was not recommended (the draft guidance FAD for second-line, high-dose imatinib, dasatinib and nilotinib for CML is available on the NICE website at http://guidance.nice.org.uk/TA/WaveR/99). It is acknowledged that this was unknown to Novartis at the time of their submission. However, the Novartis ‘scenario A’ analysis in the first row of this table, where no TKI is assumed second line, is still valid and therefore of interest (see the comparison of Novartis results with PenTAG scenario 1 analysis at the end of the cost-effectiveness section Comparison of PenTAG model with the Novartis model). BMS model only a single scenario.

Surrogate-predicted overall survival (for surrogate survival only)

Overall survival for all three treatment arms was estimated using a surrogate relationship based on CCyR at 12 months and, in a separate analysis, on MMR at 12 months. In each case, OS was estimated in four stages.

Stage 1 Overall survival for responders, and separately for non-responders, was estimated as a function of time using imatinib-arm data from a meta-analysis of trials of imatinib first-line (including the IRIS RCT26). The estimates of OS for responders and non-responders, separately for CyR and molecular response, using a meta-analysis are given in Chapter 4 (see Analysis of overall survival and progression-free survival by cytogenetic and molecular response).

Stage 2 Mortality due to CML was estimated from these historical imatinib trial data. Mortality was assumed to occur due to CML-related causes and non-CML causes. Given limited historical data, the probability of CML-related death was assumed to be constant over time. Non-CML mortality was taken from UK life tables,112 and the age at diagnosis was estimated as the average age at diagnosis across all historical trials, weighted by the number of responders or non-responders in each trial, as appropriate. The probability of CML-related death was estimated using the Microsoft Excel ‘Solver’ function (Microsoft Corporation, Redmond, WA, USA) in such as way that the sum of squares of differences between the actual historical OS and modelled OS at each year was minimised. The actual historical OS and fitted OS are shown in Figure 17.

FIGURE 17.

Historical vs fitted OS for patients (a) with and without a CCyR and (b) with and without a MMR.

Stage 3 Overall survival was estimated separately for responders and non-responders given a cohort of patients starting first-line treatment at the age of 57 years (the mean age at diagnosis for our modelling, and at present in the UK). OS was estimated by applying mortality from the general population with starting age of 57 years and the appropriate estimate of CML-related mortality from Stage 2.

Stage 4 Overall survival was estimated for each treatment arm (imatinib, dasatinib, nilotinib) by averaging the responder and non-responder OS, estimated in Stage 3, weighted by the proportion of patients who did and did not achieve a response to first-line treatment at 12 months (Figure 18). Our estimates of mean OS based on these estimation methods are given in Figure 19.

FIGURE 18.

Overall survival for each treatment arm estimated by surrogate relationship based on CCyR and, separately, MMR.

FIGURE 19.

Estimated mean OS as a function of surrogate measure used and treatment arm; treatment started at age 57 years.

Finally, we compare our estimates of expected OS with the actual 24 month OS from the RCT ENESTnd,20 the RCT DASISION,29 and with the longer-term imatinib survival data from the IRIS trial26 of imatinib compared with IFN-α.

Comparison of actual and expected overall survival

Given that the OS from the RCTs of first-line dasatinib and nilotinib is very immature, it is difficult to gauge the accuracy of the modelled OS shown in Figure 18. Nonetheless, it appears that the modelled OS is consistent with these data. At 2 years’ follow-up:

• dasatinib empirical OS was 95% based on DASISION trial data,29 compared with 97% in the model based on either the CCyR or MMR surrogate relationships75

• nilotinib empirical OS was 97% based on ENESTnd trial data,20 compared with 97% in the model based on either the CCyR or MMR surrogate relationships85

• imatinib empirical OS was 95% and 96% in the dasatinib and nilotinib RCTs, respectively, compared with 96% in the model based on the CCyR surrogate relationship and 97% based on the MMR surrogate relationship. 75,85

In addition, the estimated OS for the imatinib arm closely predicts the actual OS in the imatinib arm of the IRIS RCT26 (Figure 20). This is not a completely independent verification of OS by this method, because some of the data on OS for imatinib responders and non-responders from the IRIS RCT26 were also used to estimate the OS surrogate relationships. However, other historical data also heavily influenced the surrogate OS estimates so it is a useful calibration of the model’s survival outputs using this method (see Chapter 4, Study and population characteristics).

FIGURE 20.

Overall survival for the imatinib treatment arm estimated by surrogate relationship compared with actual OS from the IRIS RCT. 26

Overall survival by subsequent treatment and disease phase

Duration of first-line tyrosine kinase inhibitor treatment

The mean times on first-line treatment for imatinib, dasatinib and nilotinib are very important quantities in the estimation of the cost-effectiveness of first-line nilotinib and dasatinib compared with imatinib. We used the following sources of data:

• For nilotinib, we used treatment duration data up to 2.5 years’ follow-up from the RCT ENESTnd. 20

• For dasatinib, we used treatment duration data up to 2 years’ follow-up from the RCT DASISION. 29

• For imatinib, we used three sources of data from ENESTnd,20 DASISION,29 and the IRIS RCT,26 which has up to 8 years’ follow-up.

Follow-up was limited to just 2 or 2.5 years in the RCTs of first-line nilotinib and dasatinib. Given that a large proportion of patients were still on treatment at this time (0.65–0.80), it was necessary to extrapolate these proportions. This was achieved by using data from the IRIS RCT26 of first-line imatinib compared with IFN-α, with follow-up extending to 8 years. We deemed this trial as appropriate because these are the longest follow-up TKI treatment duration data that currently exist.

Treatment duration for all three drugs was estimated in the following three stages, described below, and which assumes that temporal pattern of treatment discontinuation in the new drugs would be broadly similar.

• In Stage 1, we modelled the treatment duration of imatinib in the IRIS RCT. 26

• In Stage 2, we modelled the treatment duration of dasatinib from DASISION,29 the treatment duration of nilotinib from ENESTnd,20 and the treatment duration of imatinib from DASISION29 and ENESTnd. 20

• In Stage 3, we synthesised these quantities to estimate the modelled treatment durations for imatinib, dasatinib and nilotinib.

Stage 1

First, we fitted a curve to the proportion of patients on imatinib treatment in the IRIS RCT. 26 The empirical data were taken from four publications. 27,46,99,113 Treatment cessation due to non-CML mortality was modelled independently of treatment cessation due to any other causes. Non-CML mortality was modelled by using mortality of the general population in England and Wales, with starting age of 50 years, the median starting age in the IRIS RCT. 26 We modelled treatment cessation due to any other causes as a Weibull distribution, which is most commonly used in survival analysis (Figure 21). Fitting was achieved by minimising the sums of squares of differences between actual and modelled treatment duration. The resulting parameters of the Weibull distribution, which modelled treatment cessation due to any causes except non-CML mortality, were lambda = 0.093, gamma = 0.861. Including all causes of treatment cessation (‘fitted with general mortality’ in Figure 21), the mean treatment duration was 13.0 years.

FIGURE 21.

Actual vs modelled imatinib treatment duration in IRIS RCT. 26

Stage 2

Next, the proportion of patients on nilotinib treatment in the nilotinib compared with imatinib (ENESTnd20) RCT was modelled, again splitting out non-CML mortality and modelling the remaining causes of treatment cessation as a Weibull distribution. Given such short follow-up in the nilotinib compared with imatinib RCT, it was not reasonable to estimate a shape parameter from the data from this trial. Instead, we assumed the same shape parameter of the Weibull distribution (of gamma = 0.861) as estimated in the longer-duration IRIS RCT for imatinib (Stage 1). This strongly impacts the extrapolated nilotinib treatment duration. When modelling non-CML mortality, a starting age of 57 years for first-line treatment of CML was assumed, our estimate for general age at diagnosis of the patient population in the UK [see Cohort starting age (age at diagnosis with chronic-phase chronic myeloid leukaemia)].

In the ENESTnd RCT20 of nilotinib compared with imatinib, 12-month data on nilotinib and imatinib treatment discontinuation is given in the online appendix of the paper describing this RCT. However, the most up-to-date data on treatment discontinuation are the 24-month data, which is provided by Novartis in its report to NICE and used in its model. Therefore, in common with Novartis, we based our estimate of nilotinib and imatinib treatment duration in this RCT on the Kaplan–Meier data provided by Novartis (Figure 22). This yielded parameter lambda = 0.144, and a mean treatment duration of 8.5 years (Figure 23). If, instead, we had modelled treatment cessation due to causes other than non-CML mortality as an exponential distribution, i.e. not using the shape parameter from the IRIS RCT,26 then the mean nilotinib treatment duration would have been 6.9 years. However, we repeat that we believe this method is less sound because it does not use the valuable information on treatment duration at longer follow-up from the IRIS RCT. 26

FIGURE 22.

Kaplan–Meier estimates of time on treatment in first-line nilotinib vs imatinib RCT. b.i.d., twice a day; q.d., once a day. Source: Figure 5 of Novartis’s manufacturer’s submission to NICE.

FIGURE 23.

Actual vs modelled nilotinib treatment duration in nilotinib vs imatinib RCT.

Data for treatment duration in the RCT of dasatinib compared with imatinib were taken from Kantarjian et al. 29 for 12 months’ data, Shah et al. 73 for 18 months’ data, and from conference slides by Kantarjian et al.75 for 24 months’ data.

Exactly the same procedure was then followed for dasatinib, imatinib in the dasatinib RCT and imatinib in the nilotinib RCT (Figures 24–26), with the corresponding lambda parameters equal to 0.150, 0.166 and 0.190, and corresponding mean treatment duration values of 8.2, 7.5 and 6.6 years. If instead we had modelled treatment cessation due to causes other than non-CML mortality as an exponential distribution – i.e. not using the shape parameter from the IRIS RCT26 – then the mean treatment durations would have been much shorter: 6.6, 6.0 and 5.4 years.

FIGURE 24.

Actual vs modelled dasatinib treatment duration in dasatinib vs imatinib RCT.

FIGURE 25.

Actual vs modelled imatinib treatment duration in dasatinib vs imatinib RCT.

FIGURE 26.

Actual vs modelled imatinib treatment duration in nilotinib vs imatinib RCT.

All the fitted curves are shown together in Figure 27. This clearly shows the model’s prediction that the treatment duration of dasatinib and nilotinib will be considerably shorter than treatment duration extrapolated from the IRIS RCT. 26 One possible reason may be the lack of any second-line TKI treatment following imatinib during the IRIS trial,26 so participants stayed on imatinib longer.

FIGURE 27.

First-line treatment durations before adjustment for indirect comparison.

Stage 3

In this final stage, the treatment duration curves were adjusted for the purposes of the indirect comparison between the three first-line treatments (Figure 28 and Table 38). The treatment duration for imatinib was estimated in such a way that the mean treatment duration (excluding non-CML mortality) was set to the average of the mean imatinib treatment duration (excluding non-CML mortality) in the RCT DASISION29 and the mean imatinib treatment duration (excluding non-CML mortality) in the RCT ENESTnd. 20 It was appropriate to take the average duration from the two RCTs, as this is consistent with our estimate of average response rates for patients taking all first-line drugs in our MTC (see Estimated complete cytogenetic response and major molecular response at 12 months by first-line treatment, above). In addition, the shape parameter gamma was unchanged at 0.861, the value estimated from the IRIS RCT. 26 Given that treatment cessation due to non-CML mortality is a relatively minor cause of cessation, the modelled mean duration for imatinib is approximately equal to the modelled mean duration for imatinib in the ENESTnd20 and DASISION29 RCTs (see Figure 28 and Table 38).

TABLE 38 - Estimated mean first-line treatment duration (years) for model

Trial data used for estimation	Imatinib	Nilotinib	Dasatinib
DASISION29	7.5	–	8.2
ENESTnd20	6.6	8.5	–
IRIS RCT26	13.0	–	–
Modelled for indirect comparison	7.1	9.0	7.8

FIGURE 28.

First-line treatment durations after adjustment for indirect comparison.

Next, parameter lambda for the treatment duration of nilotinib was adjusted by the ratio of lambda for imatinib for the indirect comparison, and lambda for imatinib from the nilotinib compared with imatinib RCT. The shape parameter gamma was unchanged at 0.861. This adjustment follows that suggested by Bucher et al. 67 Treatment duration for dasatinib was similarly adjusted for the indirect comparison.

Overall survival on hydroxycarbamide following tyrosine kinase inhibitor failure

Given a lack of evidence for survival on hydroxycarbamide following TKI therapy, we have adopted the same strategy as the Novartis (2009) submission to NICE for second-line nilotinib for patients with CP-CML who are resistant or intolerant to imatinib (Novartis 2009 submission, p. 36) to estimate survival on hydroxycarbamide following TKI failure. 114 This is based on a cohort study by Kantarjian et al. ,115 who present combined results for a subgroup of ‘other’ patients who are resistant or intolerant to imatinib. The 61 patients of this subgroup received a range of treatments including tipifarnib, lonafarnib, decitabine, cytarabine, homoharringtonine and IFN, with 12 receiving hydroxycarbamide. Survival when taking hydroxycarbamide is assumed to be the same as that of the ‘other’ treatment arm for imatinib-intolerant patients, even though, as acknowledged by Novartis in its 2009 report, some of the non-hydroxycarbamide treatments in this treatment group may prolong survival compared with hydroxycarbamide. However, the Novartis consultation with clinical experts for its second-line submission also suggested that this was a reasonable assumption given the lack of available relevant data on hydroxycarbamide in this setting. 114 Given the lack of relevant data, and in common with Novartis in its current submission, we assume that OS on hydroxycarbamide is independent of previous treatment. Given these limitations, our estimate of OS hydroxycarbamide following TKI failure is uncertain.

The Kaplan–Meier estimates of OS on hydroxycarbamide were read off at yearly intervals from figure 2a of Kantarjian et al. 115 As previously, two sources of mortality were modelled: CML-specific and non-CML. Non-CML mortality was modelled assuming the median initial age in the Kantarjian et al. 115 study of 54 years. The probability of CML-specific mortality was assumed to be constant over time. This probability was adjusted in such a way as to minimise the sums of squares of differences between the actual and modelled survival (Figure 29). This yielded a mean OS of 7.0 years for this trial, and a 5-year survival of 50%. Note that OS on hydroxycarbamide is lower in our model, because we assume that patients start first-line treatment at the age of 57 years, and these first-line treatments are taken for 7–9 years, so patients start second-line hydroxycarbamide aged approximately 65 years.

FIGURE 29.

Actual vs modelled OS for hydroxycarbamide following TKI failure. Kaplan–Meier data are from Kantarjian et al. 115

The derivation of our estimated time on hydroxycarbamide in CP, shown in Figure 29, is explained in the next section.

Proportion of patients receiving stem cell transplant post tyrosine kinase inhibitor failure

In all scenarios, some patients are assumed to receive SCT and some to receive hydroxycarbamide, as second- or third-line treatment after TKI failure (see Table 31). We did not identify any published evidence on the proportion of patients receiving SCT after first-line TKI failure. We therefore asked three of our clinical experts the proportion of patients that they believed would receive SCT, specifically after failure of TKIs, and at different ages. The similarity of their responses was notable, particularly the steep drop in the estimated percentage of patients who would receive a SCT in the CP after the age of 65 years. Over the age of 75 years, all three clinicians said that no patients with CP-CML would be likely to receive SCT. To approximate the responses that we received from our clinical experts, we first estimated the percentage of patients who receive a SCT for each of a range of ages (Table 39). Because of both the high cost of SCT and its important impact on life expectancy for those that survive to 5 years or more, these key assumptions will be varied in sensitivity analysis. For ease of modelling, we then estimated the proportion of patients receiving SCT as a simple linear function of time by least squares (Figure 30).

FIGURE 30.

Proportion of patients receiving SCT as a function of age in the PenTAG model.

The corresponding equation is:

Overall survival following stem cell transplant

We reviewed a number of potential published sources for estimating OS following SCT, including those used in the manufacturers’ models. The Novartis source relates to a cohort of European patients in which only 30% were patients with CML, and the data manipulation to estimate the likely survival of the CML subgroup involves the assumption that the mean pre-transplant eligibility/European Group for Blood and Marrow Transplantation (EBMT) risk score for patients with CML is 4. 119 The whole data set, as used by Novartis, is also from patients transplanted between 1980 and 2005. Graphs in the Gratwohl paper (their figure 2A and B) further indicate that patients with CML, and patients transplanted in the most recent period (2001–5), have the greatest survival compared with other diseases and the whole cohort as used by Novartis. 119

This produces an estimate of survival at 5 years (34% for those with a risk score of 4), which seems far lower than other published estimates in patients with CML we identified and reviewed. 120–122 EBMT registry data were as cited in Pavlu et al. 122 – 61% survival at 2 years, and higher (66–70%) in those patients with CML transplanted in the first CP. 123

We therefore used an alternative UK-based source. This is a review and analysis of 173 patients with CML who received SCTs in CP at Hammersmith Hospital, London, between 2000 and 2010. 124 Of these patients, 74% survived to 3 years and 72% to 6 years. This is also very similar to the survival estimate from the US Center for Bone Marrow & Transplant Research, which analysed data from 1309 patients transplanted between 1999 and 2004; its 3-year survival estimate for patients with CP-CML undergoing transplant post imatinib was 72%. 120

Similar to Novartis, we modelled OS following SCT by assuming that patients fall in to one of two distinct groups – those who have high mortality soon after transplant and those who have low mortality. Investigation of possible fits to the two Hammersmith Hospital data points revealed a plausible solution that:

1. The high-risk group has constant probability of death of 0.55 (lower i.e. dashed line in Figure 31).

2. The low-risk group has mortality equal to that of the general England and Wales population (upper dashed line in Figure 31).

3. Twenty-five per cent of patients are in the high-risk group.

The survival of the resulting total population then closely matched the empirical survival data (continuous line in Figure 31). Note that it is not important whether or not these two groups are clinically plausible. Instead, it is the survival function for all patients combined (continuous line in Figure 31) that should appear plausible compared with available empirical estimates. (A further substantial advantage of modelling survival according to the weighted average of two cohorts is that it greatly simplifies modelling, and bypasses the need for transition probabilities related to the time spent in the SCT health state.)

FIGURE 31.

Modelled OS following SCT vs actual OS from Hammersmith Hospital, 2000–10.

The life expectancy for patients having SCT at the age of 60 years is then 17.4 years, compared with 22.8 years in the general England and Wales population.

Duration of second-line nilotinib treatment

Second-line nilotinib is modelled in scenarios 3 and 4 only (see Table 31). In these scenarios, we assume that all patients initially randomised to imatinib or dasatinib take nilotinib after first-line treatment failure. Patients randomised to nilotinib take either hydroxycarbamide or SCT after nilotinib failure. We are aware of no clinical evidence of nilotinib after dasatinib failure. However, there has been a Phase II single-arm trial of nilotinib after imatinib failure. 124 In the absence of further data, we assume the duration of second-line nilotinib is independent of whether dasatinib or imatinib was taken first line.

The proportion of all patients (70% imatinib resistant and 30% imatinib intolerant) still on second-line nilotinib treatment from the single-arm trial is reported as 0.47 at 2 years. 125 But the proportion is most recently reported as 0.39 at 2 years in an update publication of Kantarjian et al. 126 Also, in their model, Novartis presents the Kaplan–Meier data for the proportion of imatinib-resistant (not imatinib-intolerant) patients still on second-line nilotinib treatment from 0 to 2 years. We assume that it is merely coincidental that its proportion at 2 years is the same as the published value of 0.47 from the 2008 abstract for both imatinib-resistant and imatinib-intolerant patients combined. Therefore, we used the Kaplan–Meier data presented by Novartis for imatinib-resistant patients only. We fitted an exponential curve to these data, and this yielded a mean time on treatment of 2.4 years. Given that the duration of treatment was short, it was not necessary to model treatment cessation due to non-CML mortality.

Health-related quality-of-life literature

We undertook a systematic literature search to identify studies which had either directly or indirectly elicited social preference weights or ‘utilities’ for different CML health states. Appendix 1 shows the full search strategy used and databases searched. Manufacturer submissions to NICE were also reviewed to identify any additional studies.

Because this was to update the search previously conducted in 2009 for our technology assessment of the same drugs for second-line treatment of CML, titles and abstracts from the bibliographic searches were examined only for the years 2009–11. The review was carried out by one researcher (RA).

Our searches identified only one new study of relevance, the 2010 study by Szabo et al. , which used the time trade-off (TTO) technique to elicit valuations of seven CML health state descriptions from 339 members of the public in the USA, Canada, UK (UK n = 97) and Australia. 127 This study, and the EQ-5D-derived utility values from the IRIS trial26 reported by Reed et al. 116 and previously supplied in a Novartis submission, appear to be the only two research-based sources of utility values for HRQoL in people with CML (excluding those based on clinicians’ estimates). 114

Utilities in Peninsula Technology Assessment Group model

Our choice of utilities is given in Table 40.

In our cost-effectiveness model we chose to use the EQ-5D-based valuations of CML health states previously reported by Reed et al. ,116 and supplemented by the unpublished data from the same trial (IRIS) for the AP and BC phases of the disease (Dr Shelby Reed, Duke University Medical Center, Durham, NC, USA, 5 July 2011, personal communication; Novartis data cited in Dalziel et al. 65). These data were collected for patients taking imatinib during the IRIS trial,26 as reported by Reed et al. 116 and used by Dalziel et al. 65 in a previous HTA of imatinib for CML. These data are drawn from a large sample of patients, using the EQ-5D, which is preferred in the NICE reference case. 129

It was necessary to estimate utility values for people taking dasatinib and nilotinib in CP, because no values are cited in the literature. In common with BMS and Novartis, we set these values equal to the value for imatinib in CP, based on clinical opinion and the similarity of the incidence of AEs by treatment.

The utilities for AP and blast phase reported by Reed et al. 116 are slightly different from those quoted by Dalziel et al. ,65 although both are taken from the IRIS trial originally. 26,116,130 In the Reed et al. analysis,116 no difference was assumed between AP and blast phase, as the observed difference in values was not statistically significant. We have therefore used the utility values cited by Dalziel et al. 65 in our model.

Utilities decrease with increasing age. 131 In common with Novartis (but not BMS), we adjusted the utilities in the model for age using the following equation (Ara and Brazier 2010):131 utility = 0.951 + 0.021 × male – 0.000259 × age – 0.000033 × age2.

The estimated utility of people taking imatinib in CP-CML of 0.854 from Reed et al. 116 is for patients of mean age of 50 years in the IRIS trial. 116 Given the formula from Ara and Brazier,131 the mean utility of a member of the general population aged 50 years, assuming that 58% are male (our assumption for CML population), is 0.867. This implies a disutility of 0.867 – 0.854 = 0.013 for patients taking imatinib in CP-CML compared with the general population. Therefore, for patients taking either imatinib, dasatinib or nilotinib in CP, we assumed that general population utilities as a decreasing function of age with a disutility of 0.013 at all ages.

We assume that the utility for patients taking hydroxycarbamide in CP equals that for the TKIs, and we further assume the same decrease in utility over time.

We do not model additional utility decrements associated with AEs in the base case.

Utility: after stem cell transplant

Although patients who survive SCT can, in most cases, be regarded as ‘cured’, in the early years following the transplant many patients will experience complications such as graft-versus-host disease (GvHD) and serious infections (due to the after-effects of myeloablation and the use of immunosuppressive agents). 132 As well as increasing mortality risk, these complications have inevitable quality-of-life impacts. 133,134

Novartis assumes that 52% of those receiving SCT experience a 0.079 utility decrement compared with patients in CP taking dasatinib, nilotinib or imatinib for the rest of their lives, while the other 48% experience the same utilities as patients in CP taking TKIs. This disutility is based on the quality-of-life impact of chronic GvHD in a 1997 health state preference elicitation study by Lee et al. (conducted with 12 US clinicians familiar with bone marrow transplant patients). 128

In the absence of other research evidence, our assumption for the utility of survivors after SCT is similar to that of Novartis. We modelled OS following SCT by assuming that patients are in one of two groups: a high-risk group with a constant high probability of death and a low-risk group with mortality equal to that of the general England and Wales population (see Overall survival following stem cell transplant, above). In the low-risk group, we assume that 52% of patients, those with chronic GvHD, have a disutility of 0.079 compared with the general population of England and Wales (not versus people in CP on TKIs as Novartis assumes). The remaining 48% of patients are effectively cured of CML, and therefore experience the age-related utility of the general population of England and Wales. Hence, on average, patients in the low-risk group have a disutility of 52% × 0.079 = 0.041 compared with the general population of England and Wales. We further assume that patients in the high-risk group have a disutility of 0.079 compared with patients taking TKIs in CP. This reflects the earlier impact of many of the post-transplant complications, and that chronic GvHD is one of a number of possible serious complications.

Drug acquisition costs

Table 41 and Figure 32 present the drug prices. The prices of dasatinib and hydroxycarbamide were taken from BNF 61 (2010), (CiC information has been removed) and the price of imatinib was taken from MIMS, which is more up to date than the lower price in BNF 61 (2010). 45,55 NICE have agreed these sources of price information.

In common with Novartis, we assume that the main alternative treatment to SCT after TKI failure is hydroxycarbamide. This drug costs only £36 per 3 months (see Table 41). However, hydroxycarbamide may not be the only, or even the main, treatment for patients post TKI failure who do not receive SCT (and we have taken our survival estimates for patients on hydroxycarbamide for second- or third-line treatment from a study population in which patients were taking several other treatments such as IFN-α, homoharringtonine and cytarabine). Therefore, to reflect the broader mix of drugs that such patients might receive, the cost of treatment with hydroxycarbamide is varied widely in the sensitivity analyses.

Cost of serious adverse events

We included an estimate of the cost of treating selected serious (grade 3 or 4) AEs while on first-line or second-line TKIs. Based on the reported rates of different AEs during the first 12 months of treatment, we decided to include only the cost of treating neutropenia, thrombocytopenia and anaemia. No other types of grade 3 or 4 AE were experienced by > 1% of patients in either of the main RCTs (i.e. DASISION29 and ENESTnd20). Although there were very few patients experiencing grade 3 or 4 pleural effusions with dasatinib, because this complication is quite common at lower grades and specific to this TKI, we also estimated the cost of these. The number of additional AEs from months 13 to 24 was so small that we chose to model only AEs during the first year of treatment with TKIs. Rates of AEs costed in the model are shown in Table 43.

The cost of treating each of these four types of AE was taken from the Oxford Outcomes study, and used a weighted average of the cost of treating a patient experiencing these complications when hospitalised or not hospitalised. They were also inflated from the 2008 to the 2011 values (Table 44).

Table 45 shows the resulting annual cost of treating the main grade 3/4 AEs, and the cost of treating pleural effusions in those taking imatinib. The incidence rates of neutropenia, thrombocytopenia and anaemia in those taking nilotinib as second-line treatment are 29%, 29% and 3.2%, respectively (sourced from Kantarjian et al. ,124 except for anaemia, which is assumed to be the same as for first-line treatment with nilotinib).

Cost of other medical management and monitoring

We based our medical management and monitoring costs on the mean frequency of hospital outpatient appointments and tests reported by the Oxford Outcomes 2009 survey of six UK-based CML clinicians. As with the estimates used in the BMS cost-effectiveness model, these were based on the frequency of routine appointments and tests after the first 3 months of treatment, and separately for patients in the chronic and advanced phases. They were also inflated to 2011 prices and adjusted for some tests when our clinical expert believed the frequency from the Oxford Outcomes survey was unrealistic or illogical (e.g. having a frequency of mutation analysis, when only one such test per patient would be conducted; C Rudin, Royal Devon and Exeter Hospital, UK, 31 July 2011, personal communication). Following comments from Novartis, with which our clinical expert also agreed, we also took out those costs in CP for repeat bone marrow aspirations (original monthly frequency of 0.3) and hospital nurse consultations (monthly frequency of 0.4), and reduced the frequency of haematology consultant outpatient appointments to quarterly if on a TKI and every 6 weeks if on hydroxycarbamide. [Although there would, in reality, be a higher frequency of visits when patients start taking TKIs, these costs would be equal (i.e. cancel out) between treatment arms.]

Note that, unlike the BMS modelling analyses, we did not include different costs for patients responding and not responding to treatment. This is for simplicity, and because the time that most patients are not responding to treatment, and before they are switched to a new treatment, should be relatively small relative to overall time in first- or second-line treatment. Also, in the questions that distinguished patients as either responding or not responding to treatment in the Oxford Outcomes cost survey (used by BMS), response was not defined; so it is wholly unclear whether this related to cytogenetic, molecular or some other type or level of treatment response for those who answered this survey.

Table 46 shows the resulting estimates of the medical management costs per month for patients in the chronic and advanced phases (AP and BC).

Cost of care post tyrosine kinase inhibitor: failure

Cost of care in advanced phases

Summary of cost-effectiveness results

Table 49 shows the cost-effectiveness results for scenarios 1–4, conventionally with comparators in order of increasing effectiveness, and the ICERs representing the incremental costs and QALYs gained by moving to the next most effective non-dominated. In the more detailed results tables in the rest of the chapter the ICERs are calculated relative to the current best clinical practice in the NHS (imatinib as first line) and then between nilotinib and dasatinib.

The variation in cost-effectiveness results across the four scenarios is considerable, with the ICERs for nilotinib compared with imatinib being either slightly above (scenario 1) or on (scenario 2) the £20,000 cost-effectiveness threshold, or – in scenarios 3 and 4 – generating slightly fewer lifetime QALYs than imatinib followed by nilotinib, but yielding significant cost savings. However, in all scenarios dasatinib is shown to be either dominated (by nilotinib) or have an ICER relative to imatinib of over £250,000 per QALY gained. These widely different cost-effectiveness results again reinforce the significance of structural uncertainty in the modelling of CML, including the substantial impact of assumptions regarding second- and third-line treatment sequences after first-line TKI failure.

The interpretation of the ICERs for scenarios 3 and 4 is unusual, because having nilotinib as second-line treatment after imatinib or dasatinib – but of course not after nilotinib failure – results in the nilotinib comparator being both less effective and more costly over patients’ lifetimes than the current best practice treatment of imatinib. Depending on which modelling assumptions are used, this means that adopting nilotinib as first-line treatment yields considerable cost savings for relatively modest QALY losses per patient (either £192,000 or £46,000 of savings yielded per QALY lost, in scenarios 3 and 4). This is discussed further in the following sections and in the Discussion (see Chapter 8).

Results: scenario 1 – cumulative survival method without second-line nilotinib

Table 50 presents the cost-effectiveness results for scenario 1.