Dabrafenib for treating unresectable or metastatic BRAF V600 mutation‑positive melanoma

The key clinical evidence came from the BREAK‑3 trial. This was an international, multi‑centre, randomised, open‑label, active‑controlled trial comparing dabrafenib with dacarbazine in people with previously untreated, unresectable, advanced or metastatic BRAF V600 mutation‑positive melanoma. The company also provided supportive evidence from 4 other trials of dabrafenib, including 2 randomised‑controlled trials (BRF113220 and Combi‑d) and 2 single arm trials (BREAK‑2 and BREAK‑MB). However, none of these trials included comparators relevant to the decision problem; therefore they were not included in a quantitative analysis.

Patients in BREAK‑3 were randomly assigned in a 3:1 ratio to receive either 150 mg of dabrafenib twice daily orally (n=187) or 1,000 mg/m² of body‑surface area of dacarbazine by intravenous infusion every 3 weeks (n=63). Baseline patient characteristics were generally similar between the treatment groups. The mean age was 53.5 years in the dabrafenib arm of the trial and 51.6 years in the dacarbazine arm. Approximately 67% of patients had an Eastern Cooperative Oncology Group (ECOG) performance status of 0. The company stated that most patients had undergone surgery previously, and adjuvant immunotherapy with interferon was the most common previous anti‑cancer therapy.

The company conducted 2 separate analyses based on 2 different cut‑off dates (December 2011 and June 2012) for the primary outcome of investigator‑assessed progression‑free survival and all secondary outcomes. For the outcome of overall survival, the company conducted a third analysis based on a later cut‑off date of December 2012. In response to a request for clarification from NICE, the company presented more recent overall survival data with a cut‑off date of January 2014.

Results from the June 2012 analysis of BREAK‑3 showed that there was a 63% reduction in disease progression for patients receiving dabrafenib compared with patients receiving dacarbazine. The median progression‑free survival was 6.9 months for dabrafenib compared with 2.7 months for dacarbazine, a statistically significant difference of 4.2 months (hazard ratio [HR] 0.37, 95% confidence interval [CI] 0.24 to 0.58, p<0.0001). The company reported results from a range of pre‑specified subgroups including age, sex, ECOG performance status, disease stage, lactate dehydrogenase levels and number of disease sites. The results showed that the progression‑free survival benefit with dabrafenib treatment was generally maintained across each subgroup.

Results from the December 2012 analysis of BREAK‑3 showed an improvement of 2.6 months in the key secondary outcome of median overall survival, which was 18.2 months in the dabrafenib group compared with 15.6 months in the dacarbazine group. However, the difference between the treatments was not statistically significant (HR 0.76, 95% CI 0.48 to 1.21, p value not presented).The company explained that approximately 57% of patients assigned to receive dacarbazine had crossed over to the dabrafenib arm by the time the analysis was conducted in December 2012. Therefore, the company adjusted the overall survival results using the 'rank preserving structural failure time' (RPSFT) model. The RPSFT‑adjusted analysis resulted in a hazard ratio of 0.55 (95% CI 0.21 to 1.43). The analysis based on the January 2014 data was not adjusted for crossover, and the results were designated academic in confidence by the company. The company specified that a crossover‑adjusted analysis of the final data from BREAK‑3 would be completed at a later date.

Health‑related quality of life (HRQoL) was assessed in BREAK‑3 using the EuroQoL (EQ‑5D) utility index. EQ‑5D data were collected for all the participants in the trial at screening, week 6, week 12, week 15, at disease progression, and approximately 30 days after disease progression. The mean change in EQ‑5D utility index score from baseline to week 15 was lower in the dabrafenib group (+0.053) than in the dacarbazine group (+0.128).

The data from December 2012 showed that the most commonly reported grade 3 adverse events with dabrafenib were pyrexia, back pain, squamous cell carcinoma and hyperglycaemia; whereas neutropenia, decreased appetite and leukopenia had a higher incidence in the dacarbazine arm. Treatment‑related adverse events, adverse events leading to discontinuation of trial treatment, and adverse events leading to dose reduction or interruption were not reported for the December 2012 analysis. For the December 2011 analysis, treatment‑related adverse events occurred with a higher frequency in the dabrafenib arm (88%) than in the dacarbazine arm (73%), whereas the incidence of adverse events leading to discontinuation of trial treatment or dose reduction was similar between the treatment arms.

The company submitted a de novo '3‑state partitioned survival' model comparing dabrafenib with dacarbazine and vemurafenib for previously untreated, unresectable or metastatic BRAF V600 mutation‑positive melanoma. The company considered a partitioned survival model to be more appropriate than a Markov model because it uses distributions of progression‑free survival and overall survival as model inputs and therefore ensures that the model results match those observed in the trial. Patients were assumed to enter the model in the 'progression‑free' health state, and in each cycle could either remain in that state or progress to a worse state; that is, the 'post‑progression' health state or 'death' state. The model had a lifetime horizon of 30 years, consisting of weekly cycles and no half‑cycle correction. The company based the analysis on an NHS and personal social services perspective, and costs and benefits were discounted at an annual rate of 3.5%. The company explained that there were not enough clinical data available for it to estimate the cost effectiveness of dabrafenib in adults who had received previous treatment. No subgroup analyses were conducted by the company.

For the comparison with dacarbazine, the proportion of people in each health state in the company's model was estimated based on survival functions for investigator‑assessed progression‑free survival and overall survival using the June 2012 and December 2012 data respectively from BREAK‑3. The company fitted independent log‑normal distributions to individual patient data for progression‑free survival for both treatment arms from time 0, using accelerated failure time (AFT) regression. The fitted curves were then extrapolated beyond the trial period (defined as 53.1 weeks for dacarbazine and 71.1 weeks for dabrafenib based on the last censor or observed failure time for progression‑free survival) to 30 years. The company stated that the log‑normal distribution provided the best fit to the data based on goodness‑of‑fit statistics and comparisons of the area under the curve.

The company modelled overall survival in the dabrafenib arm in 3 phases:

• Phase 1: The company fitted the log‑normal distribution to individual patient data from the dabrafenib arm of BREAK‑3 for the trial period only (defined as 96 weeks based on maximum censor or failure time for overall survival) using AFT regression.

• Phase 2: From the end of trial follow‑up and up to 10 years, the company applied hazard rates obtained by fitting the log‑logistic distribution to the overall survival data from the American Joint Committee on Cancer (AJCC) registry reported by Balch et al. (2009), which was weighted by the relative proportions of patients according to disease stage in BREAK‑3.

• Phase 3: For the remaining duration of the model, the company modelled survival by applying mortality rates obtained from UK general population life tables.

The company stated that there are uncertainties associated with fitting parametric curves to the data from the RPSFT analysis for the dacarbazine arm because of the small number of patient in the analysis. Therefore, it modelled overall survival for the dacarbazine arm as proportional hazards compared with dabrafenib using the RPSFT‑adjusted hazard ratio from BREAK‑3. This hazard ratio was applied for the trial period of 96 weeks, after which no further treatment effect was assumed.

For the comparison with vemurafenib, the company applied the progression‑free survival and overall survival hazard ratios from the indirect treatment comparison to the parametric survival curves used to model the dabrafenib arm. Proportional hazards were assumed throughout the entire model timeframe for progression‑free survival, whereas for overall survival it was assumed for the trial period of 96 weeks only.

Treatment‑specific EQ‑5D utility data for pre‑progression and post‑progression, derived directly from BREAK‑3, were used in the model for dabrafenib and dacarbazine. The utility values for the progression‑free health state were 0.77 for dabrafenib and 0.75 for dacarbazine. The value for the post‑progression state was 0.68 for dabrafenib. The company did not calculate the decrement in post‑progression utility for dacarbazine because of potential confounding from crossover in BREAK‑3. Given that there was no rationale to assume that health‑related quality of life after progression would differ between treatments, the company assumed that the post‑progression utility value for dacarbazine would be the same as that for dabrafenib. In the absence of comparable EQ‑5D utility data for vemurafenib, the company assumed that the vemurafenib utility values would be the same as those for dabrafenib. The company did not include disutilities associated with adverse events in the model because it considered the effect of adverse events on quality of life to be captured in the progression‑free and post‑progression health state utility values. Health related quality of life was assumed to be constant over time in each health state.

Costs incorporated in the company's model included drug costs, dispensing costs for dabrafenib and vemurafenib, administration cost for dacarbazine, BRAF testing, anti‑cancer therapy after the study, costs associated with health states and adverse events and one‑off costs for starting treatment and for death. The costs of dabrafenib and vemurafenib were estimated using the patient access schemes provided by the companies; the cost of dacarbazine was based on the list price. Estimates of the costs of treating adverse events were based on the results of a cost‑of‑illness study commissioned by the company and the incidence of adverse events from BREAK‑3 and BRIM‑3. The anti‑cancer therapy costs after the study were estimated to be £3,013 for dabrafenib and vemurafenib and £6,044 for dacarbazine.

The base‑case incremental cost‑effectiveness ratio (ICER) was £49,019 per quality‑adjusted life year (QALY) gained for dabrafenib compared with dacarbazine, and £11,028 per QALY gained for dabrafenib compared with vemurafenib.

The company's probabilistic analysis showed that at a maximum acceptable ICER of £30,000 per QALY, dabrafenib would have a 6% probability of being cost effective compared with dacarbazine and a 56% probability compared with vemurafenib. At a maximum acceptable ICER of £50,000 per QALY dabrafenib would have a 43.5% probability of being cost effective compared with dacarbazine.

The company conducted a series of deterministic sensitivity analyses. For the comparison with dacarbazine, overall survival was the key driver of the cost‑effectiveness results. When the RPSFT‑adjusted hazard ratio for overall survival was varied using the 95% confidence interval, the ICERs ranged from £26,470 per QALY gained to dacarbazine dominating dabrafenib (that is, dabrafenib was more expensive and less effective). The key driver of the cost‑effectiveness result for the comparison with vemurafenib was the progression‑free survival assumption; varying the hazard ratio using the 95% confidence interval resulted in ICERs ranging from £67,220 per QALY gained to a scenario where dabrafenib dominated vemurafenib. The cost effectiveness estimate for dabrafenib compared with vemurafenib was also sensitive to the overall survival hazard ratio used in the model, assuming a class effect between dabrafenib and vemurafenib and using alternative distributions to model survival. Changes to the time horizon, cost parameters and utility values did not have a large effect on the base‑case ICER.

The ERG was satisfied that all relevant studies with the appropriate comparisons were included in the company's analysis. In general, it concluded that BREAK‑3 was a good quality trial. The ERG agreed that evidence from the BREAK‑3 trial shows that treatment with dabrafenib was associated with progression‑free survival benefits compared with dacarbazine. However, it did not agree with the company that crossover in BREAK‑3 explained the lack of an overall survival benefit with dabrafenib, because an analysis of the survival data showed that patients in the dacarbazine arm who crossed over did not gain any significant benefit over patients that did not cross over.

The ERG did not consider the RPSFT method used to adjust for cross‑over in BREAK‑3 to be appropriate for the following reasons:

• survival data was not based on the final trial data, but on an interim analysis with few deaths

• the assumption in the RPSFT method of a 'common treatment effect' (that is, that the effect of dabrafenib is the same whether treatment starts at diagnosis or after the disease has progressed) is questionable

• patients in the dabrafenib arm and dacarbazine arm did not receive similar treatments at the time of disease progression, and some received treatments that are not used in routine clinical practice in the UK.
  
  The ERG stated that it might have been more appropriate to use the 'inverse probability of censoring weighting' (IPCW) method to adjust for crossover. However, it acknowledged that the lack of complete overall survival data and the small number of deaths to date may invalidate the use of the IPCW method at present.

Regarding the company's indirect treatment comparison, the ERG was satisfied that the BREAK‑3 and BRIM‑3 trials were broadly similar in terms of patient population and eligibility criteria. However, it noted that a greater proportion of patients in BRIM‑3 had lactate dehydrogenase levels above the upper limit of the normal range than patients in BREAK‑3. The ERG commented that this may have a negative effect on the prognosis of patients in BRIM‑3 compared with that for patients in BREAK‑3. The ERG questioned the validity of the approach used to conduct the indirect comparison because the assumption about constant hazard ratios for progression‑free survival and overall survival within BREAK‑3 and the assumption about constant proportional hazards for dacarbazine overall survival between the 2 trials were not met. The ERG also commented that because of the problem of adjusting for crossover in the individual trials it is more appropriate to use the unadjusted hazard ratios in the indirect treatment comparison. The ERG also noted that the most recent overall survival data from BREAK‑3 trial (January 2014 data) had not been used in the company's indirect treatment comparison. In view of these issues, the ERG stated that the evidence from the company's indirect treatment comparison was not robust. It was therefore unable to comment on the clinical effectiveness of dabrafenib compared with vemurafenib.

The ERG stated that the assumptions about the clinical effectiveness model inputs were the main drivers of the company's cost effectiveness estimate. The ERG noted that the company censored progression‑free and overall survival data at the date of last observation, which could misrepresent survival projections when used to calibrate parametric survival functions. It considered that, to remove any bias, censoring survival data at the date of data‑cut would have been more appropriate.

The ERG noted that the company modelled overall survival in 3 phases and it considered this approach to be complex and associated with several limitations, as follows:

• Phase 1 from randomisation until 96 weeks (1.8 years): The ERG stated that the log normal distribution used to represent the whole period of the trial follow‑up was questionable because the log‑normal distribution is known to have a long tail, and therefore overestimates overall survival gain. The ERG also considered that it would have been more appropriate to use the most recent overall survival data available from BREAK‑3 (January 2014).

• Phase 2 data from 1.8 to 10 years: The ERG noted that the company did not present evidence to support the clinical or biological plausibility of the log‑logistic distribution fitted to the AJCC registry data used in this phase. It also noted that the registry data are based on a North American population and therefore may not be representative of the UK population.

• Phase 3 data from 10 to 30 years: The ERG noted that this approach assumes, without any supporting evidence, that long‑term survivors are effectively cured of metastatic disease.

The ERG noted that the dabrafenib parametric model was used as a basis for modelling survival for the first 10 years in the dacarbazine and vemurafenib arms and was concerned that the company's assumption of proportional hazards across the 3 treatments may not be valid. The ERG indicated that the reliability of this approach was strongly affected by the assumptions about the RPSFT‑adjusted hazard ratio from BREAK‑3 used to model the dacarbazine arm.

The ERG noted that the majority of the estimated QALY gain in the company's model arises from the estimated life‑years after disease progression (62% for the comparison with dacarbazine and 93% for the comparison with vemurafenib). It also noted that the estimated mean survival in the company's model is much larger than is normally observed in clinical trials or in registry data. The ERG explained that 2 aspects of the model contributed to these results:

• the use of the RPSFT‑adjusted survival data from BREAK‑3 resulted in a large survival difference between dabrafenib and dacarbazine after 96 months

• the structure of the model in 3 phases involves applying mortality rates from different sources that are not compatible, and that seriously underestimate death rates, so extending estimated survival times over many years. The ERG drew attention to the sudden changes in mortality rates at 96 months and 10 years in the company's model, which lack any clinical or epidemiological justification.
  
  These 2 features of the model serve to establish a large early survival advantage for dabrafenib at 96 months, which is then extended over 3 decades by the overoptimistic modelling of mortality.

The ERG noted that a comparison of the cost effectiveness of dabrafenib and vemurafenib depends on an indirect treatment comparison between the BREAK‑3 and BRIM‑3 trials, in order to generate progression‑free survival and overall survival hazard ratios. The ERG examined the trial results and found that the proportional hazards ratio assumption was not valid within each trial or between the dacarbazine arms of the trials. Therefore, robust hazard ratios for dabrafenib compared with dacarbazine could not be obtained. The ERG concluded that any estimated ICERs generated by this indirect comparison would be unreliable and likely to be misleading. The ERG therefore restricted any detailed numerical comparisons to the direct evidence from the BREAK‑3 trial for dabrafenib compared with dacarbazine.

For the comparison with dacarbazine, the ERG used alternative methods to derive the clinical effectiveness model inputs. The ERG revised the censoring of progression‑free survival data for some patients still at risk at the end of the trial to reflect the date of data‑cut rather than the date of last observation. It then used the area under the curve approach to model progression‑free survival in the early variable segment and an exponential projective function for the latter period. This change increased the ICER from £49,019 to £52,035 per QALY gained.

The ERG analysed the latest overall survival data from the January 2014 data and concluded that the risk of bias introduced by the choice of censoring method is small, and could only possibly affect the last few recorded events in each trial arm. Therefore the analysis was carried out without any post‑hoc adjustments. The ERG was aware that the latest overall survival data were unadjusted for cross over. However, it noted its earlier analysis, which showed no statistically significant survival difference between the patients in the dacarbazine group who crossed over and those who did not cross over (see section 3.21). The ERG then modelled overall survival using the area under the curve survival data until a long term trend was established and then applying the long‑term exponential function throughout the rest of the trial period, that is, 1.8 years. It continued to apply the long‑term exponential projection derived from BREAK‑3 data throughout the company's phase 2 time period, that is, up to 10 years. The ERG noted that all patients had died at the end of 10 years; therefore it restricted the modelling to 10 years rather than the 30 years assumed by the company. The ERG's analysis resulted in a considerably lower overall survival gain than the company's base case, and the resulting ICER was £99,560 per QALY gained.

The ERG made several amendments to the dacarbazine acquisition and administration cost estimates, which increased the ICER for dabrafenib compared with dacarbazine by £87 per QALY gained. The ERG applied equal costs for anti‑cancer treatment after the study to the dabrafenib and dacarbazine groups because the difference between the post‑study treatments received in BREAK‑3 were not statistically significant. This amendment increased the ICER by £3,047 per QALY gained for dabrafenib compared with dacarbazine. Combining all the ERG's exploratory analyses and model amendments resulted in an ICER of £112,727 per QALY gained for the comparison with dacarbazine.

Full details of all the evidence are in the committee papers.