3 The company's submission
3.1 The company's systematic review identified 3 phase III randomised controlled trials that evaluated imatinib as adjuvant treatment for gastrointestinal stromal tumours (GISTs): ACOSOG Z9001, the SSGXVIII/AIO study and EORTC 62024. The company identified 12 non‑randomised trials and stated that their results generally supported those from the randomised controlled trials.
3.2 ACOSOG Z9001 was a randomised, double‑blind, placebo‑controlled phase III trial of adjuvant imatinib in adults at any level of risk of recurrence after complete surgical removal of KIT (CD117)‑positive GISTs. The trial compared imatinib 400 mg per day for 1 year (n=359) with placebo for 1 year (n=354). The primary outcome was recurrence‑free survival (changed from overall survival). Secondary outcomes included safety. The company provided intention‑to‑treat analyses for recurrence‑free survival and overall survival in ACOSOG Z9001, for both the full population and a high‑risk population identified retrospectively using the Miettinen 2006 criteria (Miettinen and Lasota 2006; these criteria are based on tumour size, location and mitotic rate). At the time of the primary analysis, the study was unblinded and patients randomised to placebo who had not experienced disease recurrence (n=79) were allowed to crossover to treatment with imatinib for 1 year. Of these patients, 72 opted to crossover to receive 1 year of treatment with imatinib. The company also provided data from a 5‑year follow‑up analysis of ACOSOG Z9001 but noted that because these data were analysed according to intention to treat, they did not take into account the impact of treatment switching.
3.3 The SSGXVIII/AIO study was a randomised, open‑label, prospective phase III trial of adjuvant imatinib in adults with KIT (CD117)‑positive GISTs that had been removed during open surgery, and were classified as being at high risk of recurrence (based on modified US National Institutes of Health consensus criteria). It compared imatinib 400 mg per day for 1 year (n=199) with imatinib 400 mg per day for 3 years (n=198). The primary outcome was recurrence‑free survival and secondary outcomes included safety and overall survival. The company presented results from the SSGXVIII/AIO study based on an analysis of 397 of the 400 randomised patients. Patients in the study were considered to be at high risk of recurrence in line with the modified version of the National Institutes of Health consensus criteria, but the company was also able to present post‑hoc results for people at high risk using the Miettinen 2006 criteria (n=281). The company considered the baseline patient and disease characteristics in ACOSOG Z9001 and the SSGXVIII/AIO study to be well balanced between treatment groups. Neither study recorded health‑related quality of life.
3.4 EORTC 62024 is an ongoing randomised, controlled, open‑label, observational phase III study of adjuvant imatinib in adults with resected localised KIT (CD117)‑positive GISTs who were at intermediate or high risk of recurrence, as defined by the National Institutes of Health consensus criteria. The study is comparing imatinib 400 mg per day for 2 years (n=454) with no other therapy after surgery (n=454). The primary outcome was originally overall survival, but this was later changed to imatinib failure‑free survival (defined as the time at which patients had to change to a different tyrosine kinase inhibitor because of disease relapse or recurrence). Results for EORTC 62024 were presented from an interim analysis but the company stated that it was not possible to determine patients at high risk of recurrence in line with the Miettinen 2006 criteria. The company also stated that imatinib failure‑free survival as a surrogate outcome for overall survival was different from other GIST trials. The company emphasised that robust conclusions on the effect of adjuvant therapy for 2 years on survival cannot be drawn from this interim analysis.
3.5 At the primary outcome analysis of ACOSOG Z9001 (median follow‑up 19.7 months), 1‑year recurrence‑free survival was estimated to be 98% (95% confidence interval [CI] 96 to 100) in the imatinib group and 83% (95% CI 78 to 88) in the placebo group, which was a statistically significant difference (hazard ratio [HR] 0.35 [0.22 to 0.53], p<0.0001). Overall survival at 2 years was estimated to be 98.8% in the imatinib group and 97.6% in the placebo group with a hazard ratio of 0.66 (95% CI 0.22 to 2.03), which was not a statistically significant difference.
3.6 At the 5‑year follow‑up analysis of ACOSOG Z9001, approximately 75% of patients in both groups remained on study. The 5‑year follow‑up data showed that estimated 5‑year recurrence‑free survival (median follow‑up 46.3 months) was 72.8% (95% CI 67.1 to 78.4) in the imatinib group and 68.4% (95% CI 63.0 to 73.8) in the placebo group. Recurrence‑free survival for imatinib was statistically significantly greater in the imatinib group compared with the placebo group during follow‑up (HR 0.718 [95% CI 0.531 to 0.971], p=0.0305). The company advised that the follow‑up analysis was confounded by patients who had been randomised to placebo and were recurrence‑free at the time of study unblinding and had then opted to crossover to active treatment for 1 year. The company reported that a supporting analysis which removed these patients had a hazard ratio of 0.671 (95% CI 0.491 to 0.919, p=0.0123), but did not provide any further details about the methodology.
3.7 The 5‑year follow‑up analysis showed that 5‑year overall survival (median follow‑up of 60.2 months) in ACOSOG Z9001 was 91.3% for the imatinib group and 91.1% for the placebo group. There was no statistically significant difference in overall survival between treatment groups during follow‑up (HR 0.816 [95% CI 0.488 to 1.365], p=0.4385). A sensitivity analysis that censored for patients eligible for crossover to 1 year of imatinib treatment gave a hazard ratio of 0.746 (95% CI 0.441 to 1.262, p=0.2725).
3.8 Median duration of follow‑up was 54 months for the full population. The median time to recurrence was 53.2 months for the 1‑year imatinib group, but it was not reached for the 3‑year imatinib group. Overall, recurrence‑free survival was statistically significantly longer in the 3‑year group than the 1‑year group (HR 0.46 [95% CI 0.32 to 0.65], p<0.0001; 5‑year recurrence‑free survival 65.6% and 47.9% respectively). There was no statistically significant difference in the risk of recurrence or death between the 1‑year and 3‑year treatment groups during the first year of treatment or after 3 years. However, a difference was evident from 1 to 2 years (HR 0.26 [95% CI 0.13 to 0.53]) and from 2 to 3 years (HR 0.17 [95% CI 0.07 to 0.39]) after randomisation.
3.9 Overall survival was statistically significantly greater in the 3‑year imatinib group than in the 1‑year group (HR 0.45 [95% CI 0.22 to 0.89], p=0.019; 5‑year overall survival 92.0% and 81.7% respectively). However, there was no statistically significant difference in 5‑year GIST‑specific survival between the 2 groups (88.5% in the 3‑year group compared with 95.1% in the 1‑year group; HR 0.46 [95% CI 0.19 to 1.14], p=0.09).
3.10 The EORTC 62024 interim analysis was for a median follow‑up of 4.7 years. Of the 908 patients who had been randomised to treatment, 835 were eligible for assessment. Recurrence‑free survival for the total study population was statistically significantly greater in the imatinib group than in the observation group at 3 years (84% compared with 66%, p<0.001) and 5 years (69% compared with 63%, p<0.001). There were no statistically significant differences in 5‑year imatinib failure‑free survival between the imatinib and observation groups (87% compared with 84%; HR 0.80 [98.5% CI 0.51 to 1.26], p=0.23). Overall survival at 5 years was similar in the imatinib and observation groups (100% and 99% respectively).
3.11 Using the ACOSOG Z9001 primary analysis, the company identified 165 patients at high risk of disease recurrence according to the Miettinen 2006 criteria. For these patients, 1‑year recurrence‑free survival was 98.7% in the imatinib group and 56.1% in the placebo group. An analysis of overall survival showed no statistically significant difference between treatment groups overall (p=0.0764). Overall survival at 2 years was 100% in the imatinib group and 94.7% in the placebo group; at 4 years it was 100% and 90.9% respectively.
3.12 At the 5‑year follow‑up analysis, 103 high‑risk patients had been identified in the imatinib group and 98 high‑risk patients in the placebo group. An improvement in recurrence‑free survival was observed in the imatinib group compared with the placebo group (HR 0.608 [95% CI 0.417 to 0.886], p=0.009). The difference between the imatinib and placebo groups was at its greatest 18 months after randomisation (86.7% [95% CI 79.6 to 93.7] compared with 49.9% [95% CI 39.7 to 60.2] respectively), and then decreased over time (reaching 37.9% [95% CI 25.9 to 49.9] and 32.1% [95% CI 21.6 to 42.6] respectively at 5 years). The company noted that, unlike the primary analysis, the 5‑year follow‑up analysis was confounded by placebo patients who were recurrence‑free at the time of study unblinding opting to crossover to active treatment for 1 year.
3.13 At the clarification stage, the company provided a report that used different methods of adjusting for treatment crossover in ACOSOG Z9001. The methods were: a rank‑preserving structural failure time model, the iterative parameter estimation algorithm, inverse probability of censoring weights (IPCW) and per‑protocol analyses that censored crossovers at the time of switching or excluded them altogether. The report concluded that the IPCW method was the most reliable for recurrence‑free survival and overall survival. In patients at high risk of disease recurrence, the IPCW hazard ratio for recurrence‑free survival was 0.50 (95% CI 0.3 to 0.78), which was similar to the simple, unweighted, per‑protocol censoring approach (HR 0.52 [95% CI 0.35 to 0.77]). Both provide a numerically lower hazard ratio for recurrence‑free survival than the intention‑to‑treat analysis (HR 0.61 [95% CI 0.42 to 0.89]). For overall survival, the IPCW hazard ratio in the high‑risk group was 0.76 (95% CI 0.36 to 1.62), which was similar to the simple, unweighted per‑protocol censoring approach (HR 0.79 [95% CI 0.40 to 1.55]). The report stated that these were numerically lower than the hazard ratio for the intention‑to‑treat analysis (0.93 [95% CI 0.47 to 1.83]).
3.14 Overall, 70% of patients were at high risk of recurrence according to the Miettinen 2006 criteria (142 in the 1‑year imatinib group and 139 in the 3‑year imatinib group). At the 5‑year follow‑up, recurrence‑free survival was longer in the 3‑year group than in the 1‑year group (HR 0.43 [95% CI 0.30 to 0.62], p<0.0001). Median time to recurrence was 35.9 months in the 1‑year adjuvant imatinib group and 71.8 months in the 3‑year adjuvant imatinib group.
3.15 Overall survival was greater in the 3‑year imatinib group compared with the 1‑year imatinib group (HR 0.39 [95% CI 0.19 to 0.79], p=0.007). Overall survival rates were higher with 3‑year imatinib than 1‑year imatinib at 4 years (94.5% [95% CI 88.6 to 97.3] compared with 83.0% [95% CI 73.8 to 89.1] respectively) and at 5 years (89.5% compared with 74.2%).
3.16 The EORTC 62024 interim analysis showed that for a subgroup of patients with high‑risk disease according to the National Institutes of Health consensus criteria, there was no statistically significant difference in 5‑year imatinib failure‑free survival between the imatinib and observation groups (77% and 73% respectively, p=0.44).
3.17 The company noted the absence of a head‑to‑head trial comparing no adjuvant treatment with 3 years of adjuvant imatinib, and considered the feasibility of an indirect comparison using a log hazard ratio with pairwise treatment comparisons, which assumes constant proportional hazards. However, after inspecting the Kaplan–Meier curves for recurrence‑free survival in ACOSOG Z9001 and the SSGXVIII/AIO study, the company considered the shapes of the curves to be different for each treatment arm and decided that the assumption of proportional hazards did not hold. From this, it concluded that a simple parametric proportional hazards model fitted to these curves would not accurately estimate mean survival and so did not conduct an indirect comparison using this method. For the purpose of the economic model, an indirect comparison using non‑standard methodology was presented (see section 3.30).
3.18 In ACOSOG Z9001, grade 3 or 4 adverse events occurred in 104 patients (31%) in the imatinib group and 63 patients (18%) in the placebo group. The most common of these were neutropenia (3% in the imatinib group and 1% in the placebo group), abdominal pain (3.6% compared with 1.7%), dermatitis (3% compared with 0%), nausea (2.4% compared with 1.2%) and elevated alanine aminotransferase levels (2.7% compared with 0%). The company reported that the adverse event rate was consistent with imatinib use in chronic myeloid leukaemia and metastatic GISTs.
3.19 In the SSGVXIII/AIO study, the incidence of adverse events was similar in patients receiving imatinib for 3 years (198/198, 100%) and for 1 year (192/194, 99.0%). The incidence of grade 3 or 4 events was 20.1% in the 1‑year group and 32.8% in the 3‑year group, with leukopenia and diarrhoea being the most common. More patients in the 3‑year group (51 patients, 25.8%) discontinued imatinib than in the 1‑year group (25 patients, 12.9%).
3.20 No safety data from EORTC 62024 had been reported before the company provided its evidence submission.
3.21 The ERG stated that the company's submission contained a generally unbiased estimate of imatinib's treatment effect, and noted that the randomised controlled trials had been well conducted (although 2 trials were open label and 2 had experienced a change in the primary outcome). It indicated that the main limitation of the clinical evidence was that the treatment effect for high‑risk patients was based on retrospective subgroup analyses that varied in the proportion of total number of randomised patients (the lowest being 28%), meaning that these are most likely underpowered.
3.22 The ERG stated that differences in baseline patient characteristics between the treatment arms were more pronounced in the Miettinen high‑risk subgroups in than the full trial populations, indicating selection bias. However, the ERG was unclear if the imbalances were statistically significant. Although results were similar for the full population and high‑risk subgroups (in terms of statistically significant recurrence‑free and overall survival differences between trial arms), the ERG concluded that caution was necessary when interpreting the subgroup results.
3.23 The ERG highlighted that there were differences in the results for overall survival between ACOSOG Z9001 and the SSGXVIII/AIO study, and that neither of these trials was statistically powered to detect a difference in this outcome. In ACOSOG Z9001, there were few deaths overall and there was no statistically significant difference in overall survival for imatinib for 1 year compared with placebo at 2 years and 5 years. It further noted that even in the additional analyses that removed patients who had crossed over to active treatment, the difference between trial arms generally remained non‑statistically significant. In the SSGXVIII/AIO study, there were comparatively more deaths and at 5 years there was statistically significantly longer overall survival associated with 3‑year imatinib treatment compared with 1‑year treatment. The ERG stated that although the differences in the overall death rates could potentially be explained by differences in patient characteristics (or other variables) between the 2 trials, the available evidence suggested that extending imatinib treatment to 3 years was associated with longer overall survival than 1‑year treatment.
3.24 The ERG was concerned that the high degree of crossover at study unblinding in ACOSOG Z9001 may have confounded the results and was aware that the company had presented analyses in which patients were censored at the time of crossover. The ERG also reviewed the supplemental report that used various statistical methods to adjust for patient crossover in ACOSOG Z9001, which was provided as part of the company's response to clarification. The ERG agreed that all the methods had advantages and limitations in the assumptions made and their applicability to ACOSOG Z9001, but that the IPCW method appeared to be appropriate. The ERG noted that all methods produced hazard ratios that were lower than the intention‑to‑treat analysis and therefore more favourable to imatinib. It also noted that the IPCW method produced hazard ratios that were similar to a per‑protocol analysis that simply censors switchers at the time of crossover. The ERG considered this to be conservative because both of these approaches gave hazard ratios that were slightly lower (approximately 0.1 to 0.2) than the intention‑to‑treat analysis, compared with bigger differences for some of the other methods.
3.25 The ERG noted that although the company did not present subgroup analyses mentioned in the NICE scope, the SSGXVIII/AIO study reported recurrence‑free survival for predefined exploratory subgroup analyses according to tumour site, tumour size and tumour mutation site for the full population discussed in the company's submission. The ERG noted, however, that no results for the high‑risk group in the SSGXVIII/AIO study had been reported in the journal publication. The results of these subgroup analyses were similar to those of the full population. In the genetic mutational status subgroup, there was a statistically significant treatment effect favouring imatinib for 3 years for patients with the KIT exon 11 mutation, but not for other mutations or for patients with no mutation (but the numbers in these latter groups were smaller). The ERG advised that these analyses were exploratory and were likely to be underpowered.
3.26 The company's cost‑effectiveness analysis included patients at high risk of recurrence based on the Miettinen criteria (that is, a subset of the licensed indication described in section 2.1). The economic model compared adjuvant imatinib (1 or 3 years) after surgery with no adjuvant treatment. The company advised that this model was based on that submitted for NICE technology appraisal guidance 196 (that is, the guidance under review).
3.27 The company's model used a Markov state‑transition approach. During each monthly cycle of the model, patients could:
have a recurrent GIST (first or second recurrence)
have progressive disease (and be treated with best supportive care)
die (from GISTs or other causes).
Transition probabilities between the health states were based on the treatment‑associated probabilities of recurrence or discontinuation. The model had a lifetime time horizon of 50 years, a 1‑month cycle length, and a discount rate of 3.5% was applied to costs and health effects. The analysis was conducted from an NHS and personal social services perspective.
3.28 All patients entering the model were recurrence‑free after surgery. They received either observation or adjuvant imatinib (for 1 or 3 years), and progressed through the model as follows:
Patients who experienced a first recurrence while taking adjuvant imatinib therapy were assumed to then receive sunitinib (90%) or best supportive care (10%).
Patients who experienced a first recurrence after receiving surgery only, or after discontinuing or completing planned adjuvant imatinib treatment, received first‑line imatinib treatment (400 mg per day). Patients who discontinued adjuvant imatinib because of adverse events were assumed to have the same rate of recurrence as patients remaining on adjuvant imatinib because they were not censored when calculating recurrence‑free survival. The company assumed that 15% of patients had further surgery but this was not explicitly modelled (only costs were included, not effectiveness).
After a second progression or recurrence, or discontinuation because of adverse events, most patients (90%) received sunitinib then best supportive care after further progression.
Patients receiving best supportive care were assumed to have progressive disease and remained in this health state until death.
Moving between the different health states was dependent on the probabilities of events (recurrences, adverse events and death), which were taken from the SSGXVIII/AIO study, ACOSOG Z9001 and published sources. Before recurrence, it was assumed that death would be from non‑GIST causes only. After recurrence, the monthly probability of death was 0.043 during or after best supportive care, 0.013 during or after imatinib and 0.040 during or after sunitinib (this was assumed to be independent of adjuvant treatment). Death due to non‑GIST causes was based on published government life tables for England.
3.29 Because there was no head‑to‑head trial directly comparing surgery alone with 3 years of adjuvant imatinib, the company conducted an indirect comparison using data from the SSGXVIII/AIO study and ACOSOG Z9001. It considered that the assumption of constant proportional hazards was not met and so did not use the log hazard ratio with pairwise treatment comparisons. Instead, the baseline risk of recurrence for patients treated with surgical resection only was taken from ACOSOG Z9001. A parametric survival model was fitted to patient‑level data from the placebo arm of the trial using data from the primary analysis (at study unblinding and before cross‑over was allowed) and restricted to patients classified at high risk of recurrence according to the Miettinen criteria. The Miettinen risk group for patients in this trial was derived from the 5‑year follow‑up analysis. The company examined goodness‑of‑fit (visually and using statistical methods), and assessed the extrapolation beyond the trial's duration for validity compared with the published results of other trials.
3.30 The company calculated a treatment effect for imatinib then applied it to the baseline risk of recurrence after surgery, only to estimate the risk of recurrence for patients treated with adjuvant imatinib therapy after surgery. The treatment effect was estimated for 2 distinct periods: during treatment and immediately after stopping treatment, to capture the differences in event rates observed in each period:
During treatment, the same effect for imatinib was assumed regardless of treatment duration (1 or 3 years), with an estimated hazard ratio of 0.111 (95% CI 0.043 to 0.281) for risk of recurrence compared with placebo. This was calculated from the hazard ratio for recurrence from ACOSOG Z9001 using the Cox proportional hazards model with the data truncated at 12 months.
After stopping treatment, when compared with placebo the estimated hazard ratios were 0.519 (95% CI 0.297 to 0.906) for 1 year of adjuvant imatinib and 0.344 (95% CI 0.160 to 0.741) for 3 years of adjuvant imatinib. These were estimated using datasets of patients who had not experienced disease recurrence during adjuvant treatment. The modified dataset from ACOSOG Z9001 was used to calculate the hazard ratio for recurrence for 1‑year adjuvant imatinib compared with surgery only. The modified dataset from SSGXVII/AIO was used to calculate the hazard ratio for recurrence for 3‑year adjuvant imatinib compared with 1‑year adjuvant imatinib. These estimates of treatment effect were then combined using a frequentist indirect comparison using a fixed‑effects model to estimate the hazard ratio for recurrence for 3‑year adjuvant imatinib compared with surgery only.
To estimate the risk of recurrence for patients treated with surgical resection followed by 1 and 3 years of adjuvant imatinib, the company then applied the estimated treatment effect for imatinib (during and after stopping treatment) to the baseline risk of recurrence for patients treated with surgery only. In its clarification response, the company advised that resistance to imatinib was implicitly included in the economic model through the response rates obtained in the clinical trials (in the adjuvant and advanced settings).
3.31 Health‑related quality of life was not recorded in ACOSOG Z9001 and the SSGXVIII/AIO study. The company did a systematic review, which identified 3 potentially relevant health‑related quality‑of‑life publications, to derive the utility values for its economic model:
The company assumed that patients with GISTs who had undergone successful surgery and were recurrence‑free had the same utility as healthy individuals of the same age (0.822).
Patients receiving adjuvant imatinib treatment had a utility value of 0.741 (a utility decrement of 0.081 was applied to all patients in the base case to reflect adverse effects).
Patients receiving first‑line treatment with imatinib or sunitinib (that is, after first recurrence) had a utility value of 0.739.
A utility value of 0.739 was also used for patients taking sunitinib after a second recurrence.
The utility value for the best supportive care health state was 0.577.
3.32 The company's literature review did not identify any primary studies estimating the resource use associated with treating GISTs in the UK. Health state costs were derived from NHS reference costs, UK clinical guidelines and assumptions. The 1‑time onset cost of recurrence was £1430.69 and was assumed to include 1 GP visit, 1 specialist outpatient visit, 1 CT scan and, where appropriate, surgical resection (assumed to be 15% of patients). Annual costs of continuing phase of cancer (defined as the period between the first year after diagnosis and the last year of life) were estimated at £793.50 (an average of 2 GP visits, 5 outpatient visits and 0.5 CT scans). Costs for the last year of life were estimated to be £17,380. Drug costs for imatinib and sunitinib were taken from the British national formulary (October 2013), and the company incorporated the patient access scheme for the second‑line use of sunitinib. Costs for treating adverse events with imatinib were based on the most frequent grade 3 and 4 adverse events in SSGXVIII/AIO (neutropenia, fatigue, nausea and vomiting, and diarrhoea). No costs were assumed for treating adverse effects in patients who received surgical resection only.
3.33 The company's base‑case results showed that adjuvant imatinib treatment (1 year and 3 years) was associated with greater quality‑adjusted life year (QALY) gains and higher costs than no adjuvant treatment. In the company's fully incremental analysis, the incremental cost‑effectiveness ratio (ICER) for 1 year's treatment with imatinib compared with no adjuvant treatment was £3509 per QALY gained (incremental costs £7844; incremental QALYs 2.24). The ICER for 3 years' treatment with imatinib compared with 1 year was £16,006 per QALY gained (incremental costs £22,931; incremental QALYs 1.43).
3.34 At the clarification stage, the company reproduced the base‑case analysis using the 5‑year follow‑up data for recurrence‑free survival for the placebo arm of ACOSOG Z9001, which did not adjust for the crossover of patients from placebo to imatinib. In the company's fully incremental analysis, the ICER for 1 year's treatment with imatinib compared with no adjuvant treatment was £8556 per QALY gained. The ICER for 3 years' treatment with imatinib compared with 1 year was £17,057 per QALY gained (incremental costs and QALYs not provided).
3.35 The company carried out a range of deterministic sensitivity analyses to test the model's structural assumptions and confirmed that the ICERs were insensitive to changes in costs, utility values and most transition probabilities. It reported that varying the 'on‑treatment' and 'off‑treatment' hazard ratios according to their upper and lower confidence limits caused changes in the ICERs, with all except 1 remaining below £30,000 per QALY gained. It confirmed that when the upper limits for both were included, the ICERs increased from £3509 per QALY gained to £30,058 per QALY gained for 1 year's treatment with imatinib compared with no adjuvant treatment, and from £16,006 per QALY gained to £29,162 per QALY gained for 3 years' treatment with imatinib compared with 1 year.
3.36 The company undertook probabilistic sensitivity analyses using 1000 iterations. Like the deterministic base‑case analysis, these showed that adjuvant treatment with imatinib (1 year and 3 years) was associated with greater QALY gains and higher costs than no adjuvant treatment. For the pairwise comparisons with no adjuvant treatment, the ICER for 1 year's treatment with imatinib was £3635 per QALY gained (incremental costs £8375; incremental QALYs 2.30) and the ICER for 3 years' treatment with imatinib was £7950 per QALY gained (incremental costs £30,958; incremental QALYs 3.89). The company also provided a fully incremental analysis. The ICER for 1 year's treatment with imatinib compared with no adjuvant treatment remained at £3635 per QALY gained, and the ICER for 3 years' treatment with imatinib compared with 1 year was £14,205 per QALY gained (incremental costs £22,583; incremental QALYs 1.59). At a maximum acceptable ICER of £20,000 per QALY gained, the probability of imatinib being cost effective was 41.7% for 1 year's treatment and 58.3% for 3 years' treatment. When the maximum acceptable ICER was increased to £30,000 per QALY gained, the probability of 1 year's treatment with imatinib being cost effective decreased to 30.9%, whereas the probability of 3 years' imatinib treatment being cost effective rose to 69.1%.
3.37 The company conducted scenario analyses to further explore uncertainty. It found that there was little impact on the ICERs for any of the following scenarios:
using different parametric distributions for the survival curves
allowing dose escalation of imatinib in the metastatic setting
varying the proportion of patients receiving best supportive care (instead of active treatment) after recurrence
extending survival in the post‑recurrence health states.
The company stated that its sensitivity analyses showed the ICERs were fairly insensitive to changes in parameters and assumptions, with the ICERs generally remaining below £20,000 per QALY gained. It noted that key drivers of the model were treatment effect over time and the time horizon of the analysis, and that changes in these parameters caused some ICERs to exceed £20,000 per QALY gained.
3.38 In response to a request made at the clarification stage, the company provided scenario analyses that assumed imatinib's treatment effect declined over time during the off‑treatment period. The company's ICERs for adjuvant imatinib (1 or 3 years) compared with no adjuvant treatment, and for adjuvant imatinib for 3 years compared with 1 year, increased when the off‑treatment hazard ratio was reduced to 75%, 50% or 25% after 5 years. The ICERs ranged from £4569 per QALY gained to £34,683 per QALY gained.
3.39 The ERG stated that the model structure and methodology used by the company was a reasonable approach to modelling the cost effectiveness of imatinib as adjuvant treatment for GISTs. It observed that the company had made some amendments to the model in response to comments during the original appraisal of Imatinib for the adjuvant treatment of gastrointestinal stromal tumours (NICE technology appraisal guidance 196). However, the ERG raised several concerns with the estimates and assumptions in the current model.
3.40 The ERG noted that the company's model did not explicitly model disease progression and instead defined the health states based on treatment. As a result, the ERG considered that some of the later progressions in the model did not seem appropriate (for example, patients discontinuing treatment because of adverse events may transition to best supportive care without experiencing disease recurrence).
3.41 The ERG had reservations about the validity of some utility values. It stated that although the majority of these were based on the only published set of values for patients with advanced GISTs, insufficient methodological detail had been reported and there was a lack of information about respondents such as baseline characteristics of respondents, sample size, response rate or the valuation method adopted.
3.42 The ERG stated that there was substantial uncertainty over the company's methods used to derive the baseline risk of recurrence and relative treatment effects for adjuvant imatinib in its economic model. The ERG was aware that the company had adopted these methods to avoid confounding by crossover in the placebo arm of ACOSOG Z9001. The ERG agreed that the Kaplan–Meier curves indicated changes in the shape of the survival curves after stopping adjuvant imatinib treatment, but stated that these trends might have been more apparent, and the cut‑points more easily identified and justified, by plotting the hazard function rather than the Kaplan–Meier curves. Regarding the approach to estimating 'on‑treatment' treatment effect, the ERG was concerned that the company had derived the treatment effects using a semi‑parametric model (Cox proportional hazards) then applied these to fully parametric survival functions used to derive baseline risk of recurrence (that is, with surgery alone). The ERG noted that the on‑treatment recurrence‑free hazard ratio for the high‑risk population, using the Cox proportional hazards model with the data truncated at 12 months, was 0.111 (95% CI 0.043 to 0.281). The ERG noted that this was lower than the hazard ratio of 0.265 (95% CI 0.148 to 0.477) that was reported in the clinical‑effectiveness section of the company's submission, and was unclear whether the difference was caused by truncating the data at 12 months or the retrospective reclassification of additional high‑risk patients. The ERG noted that the company's submission does not state clearly the maximum follow‑up for placebo patients before cross‑over (censoring), so it was unable to judge the duration over which the baseline survival function was modelled. The ERG noted that the company's overall approach had required considerable post‑hoc reorganisation of the trial data and was uncertain if this had introduced biases into the estimated effects. The ERG concluded that it may be more appropriate to use the crossover‑adjusted recurrence‑free survival estimates to derive clinical effectiveness parameters from ACOSOG Z9001.
3.43 The ERG also expressed substantial uncertainty about the most appropriate assumptions for extrapolating the effectiveness of adjuvant imatinib beyond the follow‑up period of the randomised controlled trials providing baseline and relative treatment effects for adjuvant imatinib. The maximum follow‑up in the randomised controlled clinical trials was around 9 years, and these effects were extrapolated over a lifetime (40‑year) time horizon in the model. In particular, the ERG was concerned about the face validity of these survival extrapolations based on the Gompertz function, which suggested a long‑term maintenance of recurrence‑free survival in around 30% of patients who received 3 years' adjuvant imatinib treatment. The ERG was concerned that this may not be appropriate in a population initially identified as being at high risk of recurrence. This compares with approximately 20% recurrence‑free survival at 20 years using the log‑logistic model or approximately 5% using the other functions.
3.44 The ERG assessed the validity of the results generated using the company's model compared with the clinical trials. The ERG considered that there was a reasonable fit for recurrence‑free survival compared with the clinical trials at 5 years for patients who had received adjuvant imatinib (for 1 year or 3 years) and at 2 years for patients who had received no adjuvant treatment. However, the ERG considered the fit for overall survival to be poorer and noted that the company's model underestimated overall survival at 5 years for patients who had received 1‑ or 3‑year adjuvant imatinib treatment, and at 2 years for patients who had received no adjuvant treatment. The ERG added that there was uncertainty around estimated long‑term extrapolation of recurrence‑free survival, and noted that long‑term recurrence‑free survival differed widely according to the parametric distribution chosen. The ERG noted that the parametric distribution chosen by the company produced the most favourable ICER for adjuvant imatinib treatment.
3.45 The ERG expressed uncertainty over the company's approach to incorporating costs for sunitinib in the model. It noted that the company estimated sunitinib use by allowing for a 21% probability of discontinuation per month, which resulted in an estimated mean duration of treatment of 3.48 cycles. The ERG considered that the company had over‑estimated sunitinib use because the 2 clinical trials for sunitinib reported a median of 2 cycles of treatment. Based on the clinical trial results, the ERG calculated that 2.89 cycles would be a more appropriate mean estimate for use in the model. Using the ERG's estimate instead of the company's reduced the monthly cost of sunitinib (with a patient access scheme) from £1615.34 to £1231.17.
3.46 The ERG reviewed how the company had explored uncertainty in its economic model. It considered that both the parameters that were varied and the ranges used in the one‑way sensitivity analyses were appropriate and comprehensive. The ERG indicated that the company's probabilistic sensitivity analyses included most of the variables within the model, that the probability distributions had been correctly applied and that the methods used to assess parameter uncertainty were appropriate. Nevertheless, the ERG noted that the sensitivity analyses in the company's submission did not include varying either the cost of imatinib, or the proportion receiving sunitinib or best supportive care after recurrence.
3.47 The ERG identified some errors in the utility values and management costs used in the company's submission and provided corrected base‑case results. In a fully incremental analysis, the ERG's ICER for 1‑year imatinib treatment compared with no adjuvant treatment was £3612 per QALY gained (incremental costs £7819; incremental QALYs 2.16). The ERG's ICER for 3 years' treatment with imatinib compared with 1 year was £16,663 per QALY gained (incremental costs £22,928; incremental QALYs 1.38). These corrected base‑case ICERs were similar to the original base‑case results provided by the company (see section 3.33).
3.48 The ERG explored issues and uncertainties that it had identified in the company's submission, including the assumption of a continuing off‑treatment effect of adjuvant imatinib, the parametric distribution used for modelling recurrence‑free survival, resistance to imatinib and the mortality estimates used for the recurrence health states. The ERG did the following analyses:
It assumed no long‑term off‑treatment benefit after the reported follow‑up, and reported that changing this assumption did not markedly alter the cost‑effectiveness results.
It used the exponential distribution to model recurrence‑free survival and found that the ICERs increased to £9386 per QALY gained for 1‑year adjuvant imatinib treatment compared with no treatment and to £18,741 per QALY gained for 3‑year compared with 1‑year imatinib treatment.
It investigated the effect of varying the off‑treatment hazard ratio for 1‑year imatinib treatment compared with no adjuvant treatment, but maintaining the hazard ratio for 3‑year compared 1‑year imatinib treatment. The ERG reported that the ICERs were very sensitive to changes in the off‑treatment hazard ratio for 1‑year adjuvant imatinib treatment compared with no adjuvant treatment:
The ERG ran the analysis using the 95% upper confidence interval of the 5‑year update unadjusted hazard ratio estimate of 0.727 (provided by the company at the clarification stage).
It found that the ICER for 1‑year adjuvant imatinib treatment compared with no adjuvant treatment increased to £10,489 per QALY gained.
The ERG investigated the effect of assuming resistance to imatinib developing at recurrence in 15% of patients who were initially treated with adjuvant imatinib. Patients who were assumed not to respond to retreatment progressed to sunitinib. This assumption produced marginal changes to the ICERs.
It varied the mortality rate by using the lower confidence interval estimates for GIST mortality in the post‑recurrence health states, which slightly reduced the ICER for 1‑year adjuvant imatinib compared with no treatment (from £3612 to £1595 to per QALY gained). There was little effect on the ICER for 3‑year adjuvant imatinib compared with 1‑year adjuvant imatinib (which decreased from £16,663 to £16,112 per QALY gained).
It assumed that 15% of patients initially treated with adjuvant imatinib and re‑challenged upon recurrence would not respond (based upon SSGXVIII/AIO data), and would then receive sunitinib. This produced marginal changes in the ICERs.
The ERG then ran an analysis that combined several of these factors (no treatment benefit after the end of trial, exponential distribution for recurrence‑free survival and lower mortality rates). This increased the ICERs to £12,122 per QALY gained for 1‑year adjuvant treatment compared with no treatment and £29,966 per QALY gained for 3‑year treatment compared with 1‑year treatment.
3.49 Full details of all the evidence are in the committee papers.