3 The company's submission
3.1 The company's submission focused on the evidence from KEYNOTE‑002, a multicentre randomised controlled trial (including centres in Argentina, US and Europe, although not in the UK). This trial compared pembrolizumab with chemotherapy in people with advanced melanoma who had had at least 2 doses of ipilimumab and whose disease had progressed within 24 weeks of the last ipilimumab dose. People with BRAF mutation‑positive melanomas must have also had treatment with a BRAF inhibitor (vemurafenib or dabrafenib) or a MEK inhibitor (trametinib). People were randomised to chemotherapy chosen by an investigator (paclitaxel plus carboplatin, carboplatin alone, paclitaxel alone, dacarbazine or temozolomide; n=179) according to standard of care or current practice, or pembrolizumab 2 mg/kg (licensed dose; n=180) or 10 mg/kg (unlicensed dose; n=181) given every 3 weeks until disease progression, unacceptable toxicity, withdrawal of consent, physician's decision to stop therapy or study sponsor's decision to stop the study. After week 12, people who had chemotherapy and whose disease progressed were allowed to switch to pembrolizumab. The company reported that people in KEYNOTE‑002 had had several previous treatments for advanced melanoma and that their baseline characteristics were generally balanced between the 3 treatment groups. Seventy seven percent of people in the trial had BRAF wild‑type disease. The company focused on the pembrolizumab 2 mg/kg dose because it was the licensed dose.
3.2 The progression‑free survival results were based on the interim analysis 2 (data cut‑off 12 May 2014). Results based on central review showed that median progression‑free survival was 2.9 months in the pembrolizumab 2 mg/kg group and 2.7 months in the chemotherapy group. The difference in progression‑free survival between the treatment groups was statistically significant (hazard ratio [HR] 0.57, 95% confidence interval [CI] 0.45 to 0.73, p<0.0001). The company noted that the Kaplan–Meier results showed that the progression‑free survival curves for both treatment groups separated from week 12 onwards and showed a substantial separation by month 6. The company also noted that progression‑free survival results based on investigator review were consistent with the results based on central review results (median progression‑free survival was 3.7 months in the pembrolizumab 2 mg/kg group and 2.6 months in the chemotherapy group, HR 0.49, 95% CI 0.38 to 0.62, p<0.0001).
3.3 The overall survival results were also based on the interim analysis 2, at which time 86 out of 179 people (48%) in the chemotherapy group had switched treatment to pembrolizumab. At this time, 215 deaths had occurred. The company reported that there was not a statistically significant difference in overall survival results between pembrolizumab 2 mg/kg and chemotherapy (HR 0.88, 95% CI 0.64 to 1.22, p=0.229). This could be attributed to confounding because people whose disease progressed on the chemotherapy group could switch to pembrolizumab when their disease progressed. The rank‑preserving structural failure time (RPSFT) adjustment method to account for treatment switching was pre‑specified in the trial protocol. Also, the company explored other adjustment methods: 2‑stage and inverse probability of censoring weighting (IPCW). The company stated that the RPSFT method is based on the assumption of a common treatment effect, so this method might not have been appropriate because there may have been a different treatment effect in people who switched to pembrolizumab after having had chemotherapy than in people who had pembrolizumab initially. It also stated that the overall survival results derived when adjusting for treatment switching with the RPSFT method were invalid because the results were similar to the ones before correction, and because the results implied that people having pembrolizumab died more quickly after progression than those having chemotherapy. Because of the small sample size and the high proportion of people switching treatment, the company noted that it was uncertain whether the IPCW method could be considered a valid method. It stated that the 2‑stage method appeared to be the most appropriate because treatment switching occurred after disease progression and the potential relevant confounders were measured until the moment of switching. The company validated the adjusted overall survival results generated with the 2‑stage method for the control group. To do this, it used the predicted overall survival using the algorithm from Korn et al. (2008; a study that evaluated historical data from different trials that included 2100 people with metastatic melanoma in an attempt to develop benchmarks for overall survival and progression‑free survival as reference points for future trials), and reported a high degree of similarity to the adjusted overall survival trial results. For the overall survival analysis, applying the 2‑stage adjustment method, the company presented the results of 2 models. One model adjusted for all relevant covariates (including Eastern Cooperative Oncology Group [ECOG] status, tumour size, lactate dehydrogenase level, BRAF status, melanoma stage and age). The other model only incorporated ECOG. The company noted that both models led to similar results and it focused on the model with all relevant covariates (median overall survival in pembrolizumab group 11.4 months, median overall survival in chemotherapy group 7.9 months, HR 0.63, 95% CI 0.45 to 0.88, p=0.007).
3.4 The most common adverse events in the pembrolizumab 2 mg/kg group were fatigue (38.8%), itching (23.5%), constipation (21.3%), diarrhoea (20.8%), nausea (19.7%), anaemia (17.4%), cough (17.4%), decreased appetite (16.3%) and joint pain (15.2%). The most common adverse events in the chemotherapy group were fatigue (48.0%), nausea (41.5%), anaemia (26.3%), vomiting (22.8%), decreased appetite (22.8%), constipation (20.5%), hair loss (20.5%), diarrhoea (19.9%) and cough (15.8%). The company stated that the results showed that the overall safety profile of pembrolizumab as an immune therapy for advanced melanoma was favourable compared with chemotherapy.
3.5 The company did a de novo economic model to assess the cost effectiveness of pembrolizumab compared with best supportive care in people with unresectable or metastatic melanoma previously treated with ipilimumab and whose disease had progressed within 24 weeks of the last dose. The population in the model differed from the scope in that people with BRAF mutation‑positive disease had also had treatment with a BRAF inhibitor (vemurafenib or dabrafenib) in line with KEYNOTE‑002. Best supportive care included systemic therapies such as dacarbazine, paclitaxel, paclitaxel plus carboplatin, carboplatin or temozolomide. The model structure was a partitioned survival model with 3 states: pre‑progression, post‑progression and death. The cycle length was 1 week, the time horizon was 30 years (assumed to be lifetime), and costs and outcomes were discounted at a 3.5% rate. Data from KEYNOTE‑002 were used to estimate the baseline characteristics, the proportion of people in the different states, the proportion experiencing adverse events, and utility values. The average age of the cohort in the model was 60 years.
3.6 The company assumed that all chemotherapy treatments had equal efficacy in terms of progression‑free survival and overall survival. The company used progression‑free survival results based on central review assessment. It applied standard parametric curve fitting using the Gompertz distribution for extrapolating progression‑free survival in the pembrolizumab group. It stated that, because progression‑free survival results were affected by the fact that the first radiological tumour response assessment was done in week 12, it applied a 2‑part curve fit: Kaplan–Meier curves were used until week 13 and parametric curves were fitted from this point onwards. The company also stated that the proportional hazard assumption could not be rejected so it incorporated it in the extrapolation of the data. For the best supportive care group, the company directly used Kaplan–Meier data until the final date when any patient was seen to still have progression‑free disease (week 62). At this point, all remaining patients were assumed to have died or have disease progression. The company highlighted that using progression‑free survival to represent disease status within the model may have underestimated pre‑progression survival and overestimated post‑progression survival. The company acknowledged that because of the relatively short‑term progression‑free survival data from KEYNOTE‑002, the extrapolation of these results added uncertainty to the cost‑effectiveness results.
3.7 The company reported that, because overall survival data from KEYNOTE‑002 were immature and standard parametric curve fitting resulted in survival estimates that were not clinically plausible, alternative methods were needed to extrapolate survival beyond the trial period. The company used the following sources for the extrapolation of overall survival in its base case:
from 0 to 1 year: KEYNOTE‑002 data
from 1 year to 10 years: ipilimumab (previously treated) survival curve (as published in Schadendorf et al. , a study that included a pooled analysis of long‑term survival data for ipilimumab in unresectable or metastatic melanoma)
from year 10 onwards: Balch et al. (2001) registry data from the American Joint Committee on Cancer registry plus general population mortality.
3.8 The company got the utility values from EQ‑5D questionnaire data from KEYNOTE‑002. The company noted that utility values decreased when patients were closer to the time of death so utility values were calculated based on time to death. The company reported that there were no statistically significant differences in utility values between the pembrolizumab and chemotherapy groups at baseline so it pooled the utility values from both treatment groups in the model. The company also calculated pooled utility values for pre‑progression and post‑progression states and used them in sensitivity analyses.
3.9 The company included costs reflecting the clinical management of unresectable or metastatic melanoma. This included costs of treatment, monitoring and follow‑up, management of complications and adverse events, and terminal care. The incidence of adverse events was based on KEYNOTE‑002, and their associated costs were taken from NICE technology appraisal guidance on ipilimumab for previously untreated advanced melanoma.
3.10 The results from the company's cost‑effectiveness analysis of pembrolizumab compared with best supportive care showed that pembrolizumab provided 1.19 additional quality‑adjusted life years (QALYs) at an additional cost of £50,995 compared with best supportive care. This led to an incremental cost‑effectiveness ratio (ICER) of £42,923 per QALY gained. The company did deterministic sensitivity analyses and found that the variables with the highest impact on the ICER were the curve fit parameters for progression‑free survival data and the HR for overall survival from the 2‑stage treatment switching adjustment method.
3.11 The company did probabilistic sensitivity analyses to assess the uncertainty around the variables included in the model. The results led to a probabilistic ICER of £67,615 per QALY gained for pembrolizumab compared with best supportive care. The company noted that these results were higher than the deterministic results because of the uncertainty in the progression‑free survival data from KEYNOTE‑002 and the fact that, in the model, many patients did not have disease progression and had treatment for life. The cost‑effectiveness acceptability curves showed that there was a probability of about 50% of pembrolizumab being cost effective at a maximum acceptable ICER of £50,000 per QALY gained.
3.12 The ERG considered that KEYNOTE‑002 was generalisable to UK clinical practice even though there were no participating centres in the UK. It noted that the company stated that people in KEYNOTE‑002 had more advanced disease and a worse prognosis than expected in clinical practice in England. However, the ERG considered that it could also be argued that people in KEYNOTE‑002 had a better prognosis because they had ECOG status 0–1 and were considered to be fit enough to have further immunotherapy after treatment with ipilimumab.
3.13 The ERG agreed with the company that the difference in median progression‑free survival between treatment groups in KEYNOTE‑002 could be affected by the timing of the first scheduled response assessment (week 12). The ERG also agreed that it was likely that median progression‑free survival rates underestimated the treatment effect of pembrolizumab compared with chemotherapy. It noted that the company explored different methods to adjust overall survival data from KEYNOTE‑002 for treatment switching, and agreed with the company that the 2‑stage adjustment method was the most appropriate.
3.14 The ERG considered that the exponential distribution provided a better fit to Kaplan–Meier progression‑free survival data from KEYNOTE‑002 than the Gompertz distribution used by the company in its model. The ERG also noted that the Gompertz distribution usually overestimates progression‑free survival results in the long term. It noted that assuming that all patients in the best supportive care group died or had disease progression at week 62 without any projection underestimated the progression‑free survival results in the best supportive care group. The ERG considered that this overestimated the benefit of pembrolizumab in terms of progression‑free survival compared with best supportive care by about 30%. The ERG considered that progression‑free survival by investigator assessment was more representative of clinical practice than progression‑free survival results by central review. Therefore, it applied the progression‑free survival results by investigator assessment using an alternative censoring rule in its exploratory analyses and used exponential models for extrapolating the results in both treatment groups. It found that this still led to a substantial long‑term progression‑free survival benefit for pembrolizumab compared with best supportive care (net extended progression‑free survival benefit with pembrolizumab compared with best supportive care of 4.18 months compared with the company's estimate of 5.35 months). This reduced the ICER for pembrolizumab compared with best supportive care by approximately £6900.
3.15 The ERG stated that the company's approach to modelling overall survival (using 3 different sources of data) led to clinically implausible results such as a 4‑year period of zero mortality risk and a sudden increase in mortality from zero to non‑zero at 10 years. It was concerned that the company applied the overall survival hazard ratio from the trial to the whole time horizon (including to the background mortality from all causes from UK life tables), and that this led to an indefinite overall survival gain in the pembrolizumab group compared with best supportive care from 10 years to 30 years. The ERG applied a different method for extrapolating overall survival data in the model based on a previous approach developed during NICE's technology appraisal guidance on ipilimumab for previously treated advanced (unresectable or metastatic) melanoma. The ERG's method used a mixed exponential model with 2 subgroups of people (1 subgroup had a high risk of mortality, the other subgroup [about 10–15% of the total population] had much longer survival) as seen in clinical practice. The ERG used the American Joint Committee on Cancer registry data. It generated expected survival profiles matched for the subgroups of people with stage IV melanoma (M1a, M1b and M1c) for each treatment group in KEYNOTE‑002. It used the subsequent curves in both extrapolation phases of the company's model (that is, from year 1 onwards) using the point at which the American Joint Committee on Cancer registry data matched profiles corresponded to a common mortality rate in both the KEYNOTE‑002 data and the projection model. This meant that, beyond the observed trial period, most patients having pembrolizumab stopped treatment rapidly because of disease progression or adverse events, so future survival was largely determined by the conventional treatment options covered in the registry data. Using this method for extrapolating overall survival data led to a reduction in the estimated survival gain of about 17%, and increased the ICER for pembrolizumab compared with best supportive care by approximately £8400.
3.16 The ERG applied other amendments to the company's model, including the approach used to incorporate utility values and the way the company included resources use and costs in the model. The ERG noted that, although individual amendments had substantial effects on the ICER, the net effect when implementing all the changes was small. It noted that this led to an overall change in the ICER of less than £4000, so the ICER for pembrolizumab compared with best supportive care including all of the ERG's preferred amendments was £46,662 per QALY gained.
3.17 Full details of all the evidence are in available.