Everolimus, lutetium-177 DOTATATE and sunitinib for advanced, unresectable or metastatic neuroendocrine tumours with disease progression: a systematic review and cost-effectiveness analysis

Location

Most NETs have been generally classified as foregut (including those in the lungs), midgut or hindgut, as it was thought that they were derived from embryonic neural crest cells. However, this theory is not now accepted and classification should be based on the site of origin of the tumour, that is, pancreas, lung, stomach, small bowel or large bowel (colon). The term ‘carcinoid’ is outdated but colloquially refers to NETs of the small bowel that secrete 5-hydroxytryptamine; the term ‘carcinoid’ is also still in common usage for NETs of the lung. ‘NET’ is the preferred term for all of these tumours. NET tumours may be grouped together as gastroenteropancreatic (GEP) NETs. Typically, the locations of these tumours are as follows:1

• foregut tumours: develop in the bronchi, stomach, gallbladder, duodenum and pancreas

• midgut tumours: develop in the jejunum, ileum, appendix and right colon

• hindgut tumours: develop in the left colon and rectum.

Prognosis can be dependent on where a tumour is located. An analysis of 13,715 carcinoid tumours over a 5-decade period in the USA found that the best 5-year survival rates were found in patients with rectal (88.3%), bronchopulmonary (73.5%) and appendicle (71.0%) NETs. 3 The lowest 5-year survival rates were found in patients with pancreatic NETs (pNETs) (37.5%). 3

Tumour grade/degree of differentiation

A NET can be defined as grade 1, 2 or 3. The grade relates to an estimation of how fast the cells are dividing to form new cells and is based on the histological assessment and the mitotic count of the tumour. The grade of a tumour is also related to its differentiation. Differentiation relates to how well/little the tumour looks like the normal tissue/tissue of origin. Well-differentiated and low-grade cancer cells look more like normal cells and tend to grow and spread more slowly than poorly differentiated cells. High-grade tumours have cells that look very abnormal and are likely to grow and spread rapidly.

In 2010, the World Health Organization (WHO) introduced a new system for grading cancer tumours. 4 This grading system is also endorsed by the European Neuroendocrine Tumor Society (ENETS) grading schemes. 5,6 The grading scheme is as follows:

• NET grade 1 (low grade)

  • well-differentiated tumour with a low number of cells actively dividing

  • Ki-67 index of ≤ 2%

  • mitotic count of < 2 per 10 high-power field (HPF)

• NET grade 2 (intermediate grade)

  • well-differentiated tumour, but with a higher number of cells actively dividing

  • Ki-67 index of 3–20%

  • mitotic count of 2–20 per 10 HPF

• neuroendocrine carcinoma (NEC) grade 3 (high grade)

  • poorly differentiated, malignant carcinoma (most aggressive form of NET)

  • Ki-67 index of > 20%

  • mitotic count of > 20 per 10 HPF.

Stage of tumour (see Appendix 20)

Determination of the size of a tumour and whether or not it has spread beyond its original site is known as staging of the tumour. Tumour staging is performed according to a system of site-specific criteria. There are two main systems for staging NETs: the Union for International Cancer Control’s (UICC’s) TNM Classification of Malignant Tumours, Seventh Edition7 (see Appendix 19, Table 133) and the ENETS staging system5,6 (see Appendix 20, Table 134). The Royal College of Pathologists8 recommended both the WHO and the ENETS systems for assessing the grade/stage of a NET. In current practice, both systems are used together with the UICC grading system above. The difference between the UICC grading system and the ENETS grading systems is not great and would not affect outcomes relating to this report.

Secretory profile

A tumour that is releasing above-typical levels of hormones is known as a functioning tumour. The increase in hormone release will often cause symptoms that may themselves need treating in addition to treating the cancer. For example, hormones released by a pNET include insulin, glucagon and pancreatic polypeptide, whereas hormones released by an appendix NET include serotonin and somatostatin. Tumours that are not releasing hormones, and, therefore, that have no hormone-related clinical features, are known as non-functioning tumours.

Neuroendocrine Tumors Therapy trial

The NETTER-1 (Neuroendocrine Tumors Therapy) trial102 was identified as an includable trial through four published abstracts103–106 identified in the systematic review in accordance with the original NICE scope. However, it was not included in this systematic review as it did not meet the revised inclusion criteria of the updated scope issued by NICE on 18 August 2016. 30

The NETTER-1 trial is a RCT that compares 177Lu-DOTATATE plus 30 mg of octreotide long-acting release (LAR) with 60 mg of octreotide LAR. Whether or not octreotide LAR could be deemed a concomitant treatment, as the doses were different in each treatment arm, was explored.

The AG sought consultation from our clinicians, who were unable to confirm whether or not the different dosing of octreotide LAR would result in different clinical effectiveness results. Therefore, the AG searched for RCT dosing studies (see Appendix 5 for the search strategy) to ascertain whether or not 30 mg of octreotide LAR had the same clinical effectiveness as 60 mg of octreotide LAR in the NETs population. Following screening of 180 citations, no studies were identified.

As the AG could not verify with any certainty that 30 mg of octreotide LAR had the same clinical effectiveness as 60 mg of octreotide LAR, and octreotide LAR was not a comparator within the scope, this study was excluded from the review.

Taken from the company submission,31 AAA reports that the rationale for treating the comparator arm with a high dose of octreotide (60 mg) was as follows:

A higher dose was required by the regulatory authorities at the time of the parallel scientific advice meeting with the FDA and EMA considering that the patients enrolled in the trial had have progressive disease following 20 or 30 mg octreotide LAR, and it was not ethical to maintain them on the same dose regimen. Consequently, 60 mg octreotide LAR at 4-week intervals dose was agreed for the control arm in the absence of an alternative efficacious treatment approved for this type of tumour.

p. 4431

The AG appreciate that, as the only RCT of 177Lu-DOTATATE identified, this trial may be of interest to the committee and so the main outcomes are presented in Appendix 6 along with the results of an ITC with everolimus from RADIANT-4. 35

The study selection process is outlined in Figure 1.

FIGURE 1.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram.

Progression-free survival

In RADIANT-334 and A618111181 the primary outcome was PFS. Disease progression was defined by both trials as the time from randomisation to the first evidence of progression or death from any cause. 34,81 Both trials used the Response Evaluation Criteria in Solid Tumours (RECIST) version 1.0107 to define disease progression. In RADIANT-3,34 PFS was obtained from central radiology review and also local investigator review, whereas in A618111181 PFS was obtained only from local investigator review in the published paper. 81 PFS from the assessment of an independent review was available from the company submission. 33

In RADIANT-3,34 median PFS as assessed by central review was 11.4 months (95% CI 10.8 to 14.8 months) in the everolimus plus BSC arm and 5.4 months (95% CI 4.3 to 5.6 months) in the placebo plus BSC arm. Everolimus was associated with a 66% reduction in the risk of disease progression or death for people with pNETs compared with placebo [hazard ratio (HR) 0.34, 95% CI 0.26 to 0.44; Table 5].

In the submission from Pfizer,32 PFS from the assessment of an independent review was 12.6 months (95% CI 11.1 to 20.6 months) for sunitinib plus BSC and 5.8 months (95% CI 3.8 to 7.2 months) for placebo plus BSC. Sunitinib was associated with a 68% reduction in the risk of disease progression or death for people with pNETs compared with placebo (HR 0.32, 95% CI 0.18 to 0.55; see Table 5).

Locally assessed PFS in RADIANT-334 was 11.0 months (95% CI 8.4 to 13.9 months) in the everolimus plus BSC arm compared with 4.6 months (95% CI 3.1 to 5.4 months) in the placebo plus BSC arm. Everolimus was associated with a reduction (65%) in the risk of disease progression or death for people with pNETs compared with placebo (HR 0.35, 95% CI 0.27 to 0.45; see Table 5). The A618111181 trial reported locally assessed PFS to be 11.4 months (95% CI 7.4 to 19.8 months) in the sunitinib plus BSC arm and 5.5 months (95% CI 3.6 to 7.4 months) in the placebo plus BSC arm. Sunitinib was associated with a 58% reduction in the risk of disease progression or death for people with pNETs compared with placebo (HR 0.42, 95% CI 0.26 to 0.66; see Table 5). Both trials reported a shorter time for PFS in both arms for locally assessed PFS than for central/independent review.

Kaplan–Meier curves for progression free survival

Kaplan–Meier curves are presented for RADIANT-334 (local and central review) and A618111181 (local review) in Figures 2 and 3 respectively.

FIGURE 2.

Kaplan–Meier plots for PFS in RADIANT-3. 34 (a) PFS assessed by local review – Kaplan–Meier medians: everolimus 11.0 months, placebo 4.6 months; HR 0.35 (95% CI 0.27 to 0.45; p < 0.001 by one-sided log-rank test); (b) central review – Kaplan–Meier median: everolimus 11.4 months, placebo 5.4 months; HR 0.34 (95% CI 0.26 to 0.44; p < 0.001 by one-sided log-rank test). Source: figure 4.3 (p. 39) of the Novartis submission. 33

FIGURE 3.

Kaplan–Meier plots for PFS (local review) in A6181111. 81 HR 0.42 (95% CI 0.26 to 0.66; p < 0.001). Note: ITT population. Source: figure 4 (p. 45) of the Pfizer submission. 32

Overall survival

Both of the pNET studies (RADIANT-334 and A618111181) reported some data relating to OS.

It was reported for RADIANT-334 that the OS data were immature: ‘median overall survival was not reached at the time of this analysis . . . final analysis of overall survival will be performed once approximately 250 deaths have occurred’. 34 In addition, of the 203 people initially assigned to receive placebo in RADIANT-3,34 172 (85%) received open-label everolimus and 148 (73%) crossed over from placebo to everolimus following disease progression. By individuals crossing over from placebo to everolimus, the detection of a treatment-related survival benefit is confounded in ITT analysis. In RADIANT-334 the HR for OS was 1.05 (95% CI 0.71 to 1.55; see Table 5).

As it had not been reached, median OS was not reported for A6181111;81 instead, survival probability at month 6 was reported. Survival was predicted to be higher in the sunitinib arm (92.6%, 95% CI 86.3% to 98.9%) than in the placebo arm (85.2%, 95% CI 77.1% to 93.3%). Survival was improved by 59% following sunitinib treatment compared with placebo (HR 0.41, 95% CI 0.19 to 0.89; see Table 5).

Both companies [Novartis33 for everolimus (RADIANT-334) and Pfizer32 for sunitinib (A618111181)] presented updated OS data in their submission.

For everolimus, the initial data presented in Yao et al. 34 were analysed on 28 February 2010. In its submission, Novartis33 presented interim OS analysis from 23 February 2011 and final OS analysis from 5 March 5. The final OS data are also available in the published paper by Yao et al. 48 At the interim time point, median OS was still not reached in the everolimus plus BSC arm, but it was 36.63 months for the placebo plus BSC arm (HR 0.89, 95% CI 0.64 to 1.23). At the final OS time point, the median OS for everolimus plus BSC was 44.0 months (95% CI 35.6 to 51.8 months) and for placebo plus BSC was 37.68 months (95% CI 29.1 to 45.8 months), indicating an overall improvement in median OS of 6.3 months (HR 0.94, 95% CI 0.73 to 1.20; p = 0.30; see Table 5). In its submission, Novartis33 commented that ‘the results may be confounded due to the high level of treatment switching from placebo to everolimus and the receipt of subsequent anti-neoplastic therapies’ (p. 43). Novartis accounted for the treatment switching from placebo to everolimus using the rank-preserving structural failure time (RPSFT) model. The RPSFT results are shown in Appendix 7 (see Table 49) and suggest a 40% reduction in OS with everolimus compared with placebo (HR 0.60, 95% CI 0.09 to 3.95).

Additional OS data were also available in the submission from Pfizer32 (pp. 45–9). Pfizer performed final OS analysis in A618111181 after 5 years of follow-up post study closure. From the ITT population, median OS was 38.6 months (range 25.6–56.4 months; n = 55 deaths) in the sunitinib plus BSC arm and 29.1 months (range 16.4–36.8 months; n = 58 deaths) in the placebo plus BSC arm, an improvement of 9.5 months (HR 0.73, 95% CI 0.50 to 1.06; p = 0.094; see Table 5). After accounting for treatment switching using the RPSFT method, median OS in the placebo group (if the 69% of patients who crossed over to sunitinib had remained on placebo) was estimated as 13.2 months (range 11.3–16.5 months) (HR 0.34, 95% CI 0.14 to 1.28; p = 0.094) (it should be noted that the 95% CI was also reported as 0.15 to 1.28 in the company submission).

Kaplan–Meier curves for overall survival

Kaplan–Meier curves for OS and RPSFT-adjusted OS were presented for both RADIANT-334 (see Appendix 7, Figure 29, and Figure 4 respectively) and A618111181 (see Appendix 7, Figure 30, and Figure 5 respectively).

FIGURE 4.

Kaplan–Meier plot of the final OS analysis from RADIANT-334 adjusted using RPSFT (everolimus vs. placebo). Kaplan–Meier medians (95% CI): everolimus 44.02 (35.61 to 51.57) months, placebo 37.68 (29.14 to 45.77) months, placebo RPSFT (reconstructed placebo data as if the treatment crossover to open-label everolimus had not occurred): not available (20.61 to not available) months. Source: figure 4.7 (p. 45) of the Novartis submission. 33

FIGURE 5.

Kaplan–Meier estimates of OS from A618111181 with and without adjustment for treatment switching: final analysis, ITT population (sunitinib vs. placebo). mOS, median overall survival. Source: figure 6 (pp. 48–9) of the Pfizer submission. 32

Response rate

Both studies used RECIST version 1.0107 to assess tumour response. Response rate was assessed by local investigators in RADIANT-3. 34 It was unclear whether response rate was assessed locally or centrally in A6181111;81 however, as PFS was assessed only locally,81 it might be assumed that the response rate was also assessed locally. In A618111,81 clinical assessments were performed at screening, during weeks 5 and 9 and every 8 weeks thereafter, whenever progression was suspected and at the end of treatment or withdrawal from the study, whereas in RADIANT-334 assessments were performed at baseline and every 12 weeks thereafter.

Complete response, partial response, stable disease and progressive disease were reported by both studies (see Table 5). In RADIANT-334 tumour shrinkage was also reported, whereas in A618111181 the proportion of individuals who could not be evaluated and ORR were also reported (see Table 5). Complete response was achieved by only two individuals in A618111181 following treatment with sunitinib and BSC. Complete response was not achieved by any participants receiving placebo (both trials), nor following treatment with everolimus (RADIANT-334). The numbers of individuals achieving a partial response or stable disease were higher in the treatment arms (everolimus and sunitinib) than in the placebo arms in both trials. Likewise, there were higher proportions of individuals with progressive disease in the placebo arms than in the treatment arms.

Novartis33 reported in its submission that (confidential information has been removed). Novartis33 also reported central investigator RRs and adjudicated central investigator RRs from RADIANT-3. 34

Although relevant outcome data were reported, this information was designated as commercial-in-confidence.

Adverse events

In both RADIANT-334 and A618111181 AEs were assessed in accordance with the National Cancer Institute Common Terminology Criteria for Adverse Events, version 3.0. 108

Treatment-related AEs (all grades and grades 3 and 4 combined) are reported for both trials in Table 6. In RADIANT-3,34 AEs were reported that occurred in at least 10% of the safety population, whereas, in A6181111,81 AEs were reported that occurred in > 15% of the safety population. AEs were more commonly reported following treatment with everolimus and sunitinib than following treatment with placebo. The five most common all-grade AEs following treatment with everolimus (RADIANT-334) were stomatitis (64%), rashes (49%), diarrhoea (34%), fatigue (31%) and infections (23%). Following treatment with sunitinib (A618111181), the five most common all-grade AEs were diarrhoea (59%), nausea (45%), vomiting (34%), asthenia (34%) and fatigue (32%).

Treatment-related AEs occurring in ≥ 20% of patients in RADIANT-334 at the latest cut-off (5 March) were presented in the company submission from Novartis33 (see Appendix 7, Table 50). These AE rates are different (predominantly higher) from the rates published by Yao et al. 34 The AEs published in the paper by Yao et al. 34 were coded using the Medical Dictionary for Regulatory Activities, version 16.1. 109

Health-related quality of life

IN A618111181 the European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire (EORTC QLQ-C30) version 3.0 was used to measure HRQoL. The EORTC QLQ-C30 was available for 73 out of 86 (85%) individuals treated with sunitinib and 71 out of 85 (84%) individuals treated with placebo. The EORTC QLQ-C30 includes one global health scale, five functional scales (physical, role, emotional, cognitive and social), three symptom scales (fatigue, nausea/vomiting and pain) and six single-item scales (dyspnoea, insomnia, appetite loss, constipation, diarrhoea and financial difficulties). High scores are better for the global and functioning scales, whereas low scores are better for the symptom items/scales. The questionnaire was administered at baseline, at every cycle (4 weeks) and at the end of treatment. There were no overall differences observed between study groups for any of these measures except for diarrhoea (21.4-point difference; p < 0.001) and insomnia (7.8-point difference; p = 0.04), which were higher in the sunitinib arm than in the placebo arm.

More detailed results were available for HRQoL from Pfizer’s submission. 32 The mean baseline global HRQoL score was 67.0 (95% CI 62.0 to 72.0) in the sunitinib plus BSC arm compared with 64.0 (95% CI 58.4 to 69.6) in the placebo plus BSC arm. Post-baseline scores were 62.44 for sunitinib and 61.28 for placebo. Overall differences in the EORTC QLQ-C30 for each item on the scale between the two arms are shown in Table 53 (see Appendix 7). Changes in global HRQoL scores over time in the two arms are shown in Figure 31 (see Appendix 7).

Subgroup analysis

Progression-free survival

Table 7 shows the evidence used from RADIANT-334 and A618111181 to inform the indirect comparison between everolimus and sunitinib for PFS assessed by local review. The analysis suggests that everolimus is associated with a 17% decrease in disease progression or death compared with sunitinib (HR 0.83, 95% CI 0.49 to 1.42). The 95% CI is wide and includes the null hypothesis that there is no difference in PFS effectiveness between everolimus and sunitinib.

Further data available from the company submissions32,33 enabled an indirect comparison for PFS assessed by central radiology review to be carried out. The ITC between everolimus and sunitinib for PFS based on central radiology review suggests no difference between the treatments (see Table 7).

Overall survival

Table 7 shows the evidence used to inform the indirect comparison between everolimus and sunitinib for OS. The analysis suggests that there is a 2.56 times greater hazard of dying associated with treatment with everolimus than treatment with sunitinib, which is statistically significant. However, as these analyses are based on published HRs from RADIANT-334 and A6181111,81 which were not adjusted for treatment switching after disease progression, these results should not be relied on.

Final follow-up data for OS available from the company submissions32,33 enabled an indirect comparison between everolimus and sunitinib for this outcome to be carried out, which suggests a lower hazard of death associated with sunitinib than with everolimus. However, the 95% CI includes the null effect, suggesting that this is not a statistically significant effect (see Table 7). This analysis does not account for the fact that approximately 70% of participants in the placebo plus BSC arms of these two trials switched to receive the active treatment after disease progression and so it should be interpreted with caution.

Response rate

All ORs reported in Table 7 for the intervention compared with placebo were calculated by the AG, based on the numbers of individuals reported to have experienced the different outcomes in RADIANT-334 and A6181111. 81 The ITCs for everolimus and sunitinib were based on the AG-calculated ORs from RADIANT-334 and A6181111. 81 The indirect analysis suggests that there is an 82% increase in the odds of a partial response in individuals treated with sunitinib compared with everolimus. However, sunitinib was associated with a 52% increase in the odds of progressive disease compared with everolimus. Everolimus was associated with a 2.3 times greater odds for disease stability than sunitinib. However, all of these ITCs were associated with wide 95% CIs, suggesting that there is little evidence of a difference in RRs between everolimus and sunitinib.

Adverse events

An ITC was completed only for those AEs for which data were available from both trials. All ORs reported in Appendix 7 (see Table 57 for all grades of AE and Table 58 for grade 3/4 AEs) were calculated by the AG based on the numbers of participants experiencing these AEs reported in A618111181 and RADIANT-3. 34 For all grades of AE, the ITC suggests that there is a 19% increase in the odds of experiencing stomatitis and a 42% increase in the odds of experiencing nausea associated with sunitinib compared with everolimus. For the other AEs (all grades), the evidence suggests an increase in the odds of experiencing the AE with everolimus compared with sunitinib. However, except for decreased appetite, all of these ITCs were associated with wide 95% CIs that included the null hypothesis of no difference, suggesting that there is little evidence of a difference in AEs between everolimus and sunitinib. For all grades of decreased appetite, there was a statistically significant increase in the odds of experiencing the event with everolimus compared with sunitinib.

For the grade 3/4 AEs, the ITC could consider only seven AEs based on the data available from the two trials. The evidence suggests an increased odds of experiencing grade 3/4 stomatitis, fatigue, diarrhoea, nausea and thrombocytopenia with everolimus compared with sunitinib and an increased odds of experiencing decreased appetite and asthenia with sunitinib compared with everolimus. However, all of the ITCs for grade 3/4 AEs were associated with wide 95% CIs that included the null hypothesis of no difference, suggesting that there is little evidence of a difference in AEs between everolimus and sunitinib.

Subgroup analysis

Subgroup analysis based on whether or not participants had previously received SSAs suggests very little difference in time to disease progression or death for everolimus compared with sunitinib (see Appendix 7, Table 59).

Progression-free survival

In RADIANT-4,35 PFS was the primary outcome, with disease progression defined as ‘the time from randomisation to death or progression as per modified RECIST version 1.0 criteria’. 35 In RADIANT-4,35 PFS was reported according to both central radiology review and local investigator review.

Median PFS in RADIANT-4,35 assessed by central review, was 11.0 months (95% CI 9.2 to 13.3 months) in the everolimus plus BSC arm and 3.9 months (95% CI 3.6 to 7.4 months) in the placebo plus BSC arm. Everolimus was associated with a 52% reduction in the risk of disease progression or death for people with lung and GI NETs compared with placebo (HR 0.48, 95% CI 0.35 to 0.67; Table 8).

In RADIANT-4,35 locally assessed PFS was longer in duration in both arms than PFS assessed by central review, at 14.0 months (95% CI 11.2 to 17.7 months) in the everolimus plus BSC arm compared with 5.5 months (95% CI 3.7 to 7.4 months) in the placebo plus BSC arm. Everolimus was associated with a reduction (61%) in the risk of disease progression or death for people with lung and GI NETs compared with placebo (HR 0.39, 95% CI 0.28 to 0.54; see Table 8).

The company submission from Novartis33 made available secondary analysis of PFS (1 year and 2 days after the PFS analysis presented in the published paper35). Median PFS from central review was 14.39 months (95% CI 11.24 to 17.97 months) for the everolimus plus BSC arm and 5.45 months (95% CI 3.71 to 7.39 months) for the placebo plus BSC arm. Everolimus was associated with a 59% reduction in the risk of disease progression or death for people with lung and GI NETs compared with placebo (HR 0.41, 95% CI 0.30 to 0.56).

Kaplan–Meier curves were produced for PFS in RADIANT-435 from both central review and local review (Figure 7) with data from the primary cut-off point.

FIGURE 7.

Kaplan–Meier plots for PFS (primary cut-off point) in RADIANT-435 (everolimus vs. placebo). (a) PFS assessed by central review – Kaplan–Meier medians: everolimus 11.0 (95% CI 9.2 to 13.3) months, placebo 3.9 (95% CI 3.6 to 7.4) months; HR 0.48 (95% CI 0.35 to 0.67; p < 0.00001 by stratified one-sided log-rank test); and (b) PFS assessed by local review – Kaplan–Meier medians: everolimus 14.0 (95% CI 11.2 to 17.7) months, placebo 5.5 (95% CI 3.7 to 7.4) months; HR 0.39 (95% CI 0.28 to 0.54; p < 0.00001 by stratified one-sided log-rank test). Source: figures 5.3 (p. 67) and 5.4 (p. 68) of the Novartis submission. 33

Overall survival

Interim OS analysis was carried out for RADIANT-435 once 70 deaths had been reached. Data were not sufficiently mature to provide an estimation of median OS. In individuals with lung and GI NETs, Kaplan–Meier estimates for OS at the 25th percentile – 25% of individuals having died – were 23.7 months (95% CI 17.6 to 27.3 months) for everolimus and 16.5 months (95% CI 9.0 to 21.0 months) for placebo. Everolimus was associated with a 36% improvement in OS for individuals with lung and GI NETs compared with placebo (HR 0.64, 95% CI 0.40 to 1.05; see Table 8).

In its company submission, Novartis33 presented secondary analysis of OS from RADIANT-4,35 which was performed 1 year and 2 days after the published analysis presented by Yao et al. 35 This analysis was based on 101 deaths, corresponding to a 52.9% information fraction, and the median duration of follow-up was 33.4 months. Everolimus was associated with a 27% improvement in OS for individuals with lung and GI NETs compared with placebo (HR 0.73, 95% CI 0.48 to 1.11).

A Kaplan–Meier plot was produced for OS in RADIANT-435 at both the primary data cut-off point (see Appendix 7, Figure 32) and the secondary data cut-off point (Figure 8).

FIGURE 8.

Kaplan–Meier plot for OS estimates in RADIANT-4:35 secondary data cut-off point. Kaplan–Meier medians: everolimus + BSC 37.16 (95% CI 35.35 to not evaluable) months, placebo + BSC 39.56 (95% CI 23.46 to not evaluable) months; HR 0.73 (95% CI 0.48 to 1.11; log-rank p-value = 0.071). p-value is obtained from the stratified log-rank test. Source: figure 5.11 (p. 75) of the Novartis submission. 33

Response rate

In RADIANT-4,35 a modified version of RECIST version 1.0107 was used to assess tumour response by central radiology review. Efficacy was assessed every 8 weeks following randomisation for the first 12 months and then every 12 weeks thereafter.

Complete response, partial response, stable disease, progressive disease, ORR, disease control rate and tumour shrinkage, following central radiology review, were reported (see Table 8). For all RR outcomes, treatment with everolimus for lung and GI NETs resulted in a favourable response compared with treatment with placebo, except for complete response, which was not achieved in either arm.

Adverse events

In RADIANT-435 AEs were assessed according to the National Cancer Institute Common Terminology Criteria for Adverse Events, version 4.03. 111 Treatment-related AEs (all grades and grades 3 and 4 combined), reported in ≥ 10% of the safety population, are presented in Table 9. AEs were more commonly reported following treatment with everolimus than following treatment with placebo. The five most common all-grade AEs following treatment with everolimus were stomatitis (63%), diarrhoea (31%), fatigue (31%), infections (29%) and rash (27%).

In its submission, Novartis33 also reported AEs in patients regardless of the study drug relationship of the safety population. This table can be found in Appendix 7 (see Table 51).

Health-related quality of life

In its submission, Novartis33 presented data on HRQoL from RADIANT-435 using the Functional Assessment of Cancer Therapy – General (FACT-G) questionnaire. 112 The FACT-G is based on 27 items in four domains: physical well-being, social/family well-being, emotional well-being and functional well-being. Participant completion rates for the FACT-G questionnaire in RADIANT-4 are provided in Table 60 (see Appendix 7).

Mean total score over time for the FACT-G questionnaire is presented in Figure 34 (see Appendix 7). The HR for definitive deterioration of the total FACT-G score was (confidential information has been removed) Figure 34 (see Appendix 7).

In its submission, Novartis33 reported that the ‘scores were well-balanced between the two arms and never exceeded the threshold of 7 points, defined as the minimal clinically important difference between treatment arms’ (p. 77).

Subgroup analysis

Progression-free survival (based on central review) in RADIANT-435 was reported for everolimus compared with placebo based on treatment naivety, previous chemotherapy use and previous long-acting SSA use (see Appendix 7, Table 61). There is little evidence of a difference in PFS within subgroups.

Baseline characteristics

Baseline characteristics for individuals with GI NETs only are presented in Table 62 (see Appendix 7).

Progression-free survival

Novartis provided data for PFS from the RADIANT-435 trial for patients with GI NETs. Median PFS assessed by central review was 13.1 months (95% CI 9.2 to 17.3 months) for the everolimus plus BSC arm and 5.4 months (95% CI 3.6 to 9.3 months) for the placebo plus BSC arm. Everolimus was associated with a 44% reduction in the risk of disease progression or death for people with GI NETs compared with placebo (HR 0.56, 95% CI 0.37 to 0.84; see Table 8).

Overall survival

Novartis provided data for OS and Kaplan–Meier estimates of OS at the 25th percentile from the RADIANT-435 trial for patients with GI NETs (see Table 8).

Response rate

Novartis provided data for RRs from RADIANT-435 for patients with GI NETs (see Table 8).

For all RR outcomes, treatment with everolimus plus BSC for GI NETs resulted in a favourable response compared with treatment with placebo plus BSC.

Adverse events

Novartis provided data for AEs from RADAINT-4 for individuals with GI NETs. AEs were more commonly reported following treatment with everolimus than following treatment with placebo (Table 10). The five most common all-grade AEs following treatment with everolimus were stomatitis (71.8%), infections (59%), diarrhoea (44.4%), peripheral oedema (40.2%) and fatigue (36.8%).

Baseline characteristics

Baseline characteristics for individuals with lung NETs only are reported in Table 63 (see Appendix 7).

Progression-free survival

Novartis provided data for PFS from the RADIANT-435 trial for patients with lung NETs. There were 42 progression events for 63 individuals assigned to everolimus plus BSC compared with 18 events for 27 individuals in the placebo plus BSC arm. Everolimus was associated with a 50% reduction in the risk of disease progression or death for people with lung NETs compared with placebo (HR 0.50, 95% CI 0.28 to 0.88; see Table 8).

Overall survival

Novartis provided data for OS from the RADIANT-435 trial for patients with lung NETs (see Table 8). HRs were obtained from the unstratified Cox model.

Response rate

Novartis provided data for RRs from RADIANT-435 for patients with lung NETs (see Table 8).

For all RR outcomes, treatment with everolimus plus BSC for lung NETs resulted in a favourable response compared with treatment with placebo plus BSC.

Adverse events

Novartis provided data for AEs from RADIANT-4 for individuals with lung NETs. AEs were more commonly reported following treatment with everolimus than following treatment with placebo (see Appendix 7, Table 64). The five most common all-grade AEs following treatment with everolimus (RADIANT-435) (confidential information has been removed).

Results

In the advanced pNETs patients with progressive disease, the base-case analysis found a life expectancy of 4.17 years over a 20-year time horizon after the start of treatment with everolimus or sunitinib in patients with a median age of 58 years (as in the everolimus arm of RADIANT-334). Everolimus produced more QALYs than sunitinib (2.73 vs. 2.71, a difference of 0.02 QALYs discounted at 3.5%) (Table 11). As this analysis assumed equal PFS and OS outcomes between the two treatments, the QALY difference was purely the result of the impact on HRQoL of treatment differences in AEs.

Everolimus was associated with discounted health-care costs of £36,933 per patient, that is, an overall saving of £1636 relative to sunitinib, which had discounted health-care costs of £38,569 per patient. With a (confidential information has been removed) Patient Access Scheme (PAS) discount applied to the list price of everolimus and reimbursement of the first cycle of sunitinib, the total cost to the NHS of using everolimus was reduced to (confidential information has been removed) and of using sunitinib was reduced to £36,247, that is, a saving of (confidential information has been removed) per patient for everolimus. As PFS, OS and costs in the progressive disease state were the same across the two arms, the cost differences were the result of differences in the drug acquisition costs of the targeted therapies in the stable disease period. As a result, everolimus was found to be dominant over sunitinib.

Novartis [table 6.18 ( p. 112) of the Novartis submission33] reports that the total cumulative time spent per patient in the stable disease health state is 0.899 years under everolimus and 0.878 years under sunitinib. This must be an error as the base-case model assumes equal PFS outcomes for both treatments and there is no accounting in the model for differences in grade 5 AEs (deaths).

Probabilistic sensitivity analyses (PSAs) of drug acquisition prices unadjusted for the PAS produced a mean cost saving estimate for everolimus of £2055 per patient and a QALY gain of 0.002. With the PAS, the mean cost saving with everolimus was increased to (confidential information has been removed). This result suggests that the main determining factor is costs, as the mean QALY difference of 0.002 is not considered to be clinically significant. 115 However, these PSAs by Novartis were not adequately performed as the PFS and OS and SSA use (set at zero) parameter values were assumed to be the same rather than differ between the treatments according to their mean estimates in the trial data and their associated sampling uncertainty. Therefore, the claim by the company that ‘at the £30,000/QALY threshold, there is (confidential information has been removed) probability of everolimus being cost-effective when compared to sunitinib at their respective PAS prices’ (Novartis submission, p. 11433) should be considered with these reservations in mind.

Consequently, deterministic sensitivity analysis performed by Novartis showed that the most influential parameters are the relative treatment effects on PFS and OS, and the treatment duration, RDI and costs of AEs. However, the company states that the incremental difference in terms of the incremental cost-effectiveness ratio (ICER) is ‘so marginal, it does not materially affect the model results’ (Novartis submission, p. 11633). Similarly, positive rates of SSA use, use of PFS data assessed by local experts compared with centrally assessed data and choice of PFS distribution had a marginal impact. In scenario analyses, exploring the effects of changing clinical effectiveness, utility and cost parameter values, as well as structural assumptions about OS and PFS outcomes, everolimus was still dominant, except for the case in which the relative treatment effect on PFS was set to favour sunitinib according to estimates derived from the Bucher indirect comparison (HR: PFS 0.93, OS 0.72). In this instance, everolimus had lower costs and produced fewer QALYs than sunitinib and the ICER was found to be (confidential information has been removed). The interpretation of the ICER in this case reverses, that is, the higher the value the more cost-effective everolimus is; it may be reasonable to consider (confidential information has been removed) in this patient population. It must be noted that this is the result of everolimus saving £5963 (confidential information has been removed) at the expense of having 0.685 fewer QALYs per patient [table 6.23 (p. 117) of the Novartis submission33].

The scenarios explored by Novartis do not cover some major areas of uncertainty, such as variation in the relative probability of AEs, nor structural uncertainty associated with different functional forms for the different treatment arms in each of the progressive disease and stable disease phases. Nevertheless, given the possibly clinically insignificant utility differences discussed earlier, or that Novartis may produce fewer QALYs than sunitinib in exchange for a small reduction in NHS costs, the dominance of everolimus over sunitinib in the Novartis base-case analysis is not robust to the sources of uncertainty investigated by Novartis.

Strengths and weaknesses of Novartis’s evaluation of pancreatic neuroendocrine tumours

The model by Novartis follows the NICE reference case37 (see Appendix 9) except for one major aspect, which was the lack of inclusion of BSC as a comparator. This omission is at odds with the fact that the two RCTs from which the effectiveness data were obtained for the model compared targeted therapy (everolimus or sunitinib) plus BSC with placebo plus BSC. There was no justification given by Novartis for omitting the BSC-only treatment option in its cost-effectiveness analysis. However, in the opinion of our clinical experts, BSC is a relevant initial treatment option for patients with advanced, progressive pNETs and small or asymptomatic tumours, in whom active treatment may be considered on disease progression.

This was a complex area to analyse because of limited data on advanced pNETs patients with progressive disease, which is a natural result of the small incidence of this disease. Evidence on resource use was particularly limited, especially for the progressive disease phase of the model, for which the quantities used in the model were based on expert opinion. Another major uncertainty in the evidence base is the lack of HRQoL data for everolimus in the patient population of interest here. This ultimately led Novartis to base the comparison of HRQoL and, given the base-case assumptions of equal OS and PFS outcomes between the treatments, QALY outcomes between the two treatments on their relative impact on AE incidence and severity. Thus, the difference in QALYs was based on values derived from actual patient AE outcomes complemented by clinical experts’ views of quality of life in stable disease under different AEs.

A critical feature of the OS data used in the Novartis economic model was the adjustment for crossover from placebo to active treatment in the RCT data. In particular, the method used for such adjustment, the RPSFT model, relied on the assumption that the benefit derived by patients from receiving targeted treatment was the same whether they were given it as initial treatment or subsequently on disease progression. This assumption may be questionable and it is therefore natural to expect that, in the present case, sensitivity analyses allowing for a reduction in the benefit conferred by targeted treatment received after disease progression should have been performed. Although other methods are available to adjust for treatment crossover, such as inverse probability of censoring weight116 and censoring at crossover,117 they are clearly inferior as the majority of cases of crossover in RADIANT-334 and A618111181 occurred following study termination, making the key assumption underlying these methods, either that crossover is random or that patients who did not cross over may be representative of those who did, unlikely to hold true.

The Bucher indirect comparison from which Novartis derived estimates of relative effectiveness and side effects for populating its economic model was based on data that appear to be outdated on three fronts. First, Novartis estimated relative OS effectiveness using an indirect comparison (Bucher method) of placebo-adjusted HRs that corrected for crossover from the placebo to the active treatment arms in the two RCTs used in the indirect comparison. The crossover-adjusted OS HR estimate used for sunitinib compared with placebo in the indirect comparison by Novartis is lower than that in the final published results for OS for A6181111. 81 With these new data, we obtained a crossover-adjusted OS HR for everolimus compared with sunitinib of 0.51, instead of the OS HR of 0.72 derived by Novartis. Second, our updated searches of the literature conducted in September 2016 identified a forthcoming publication48 providing grade 3/4 AE counts that differ in some instances from those used for the everolimus and placebo arms in the Bucher indirect comparison submitted as evidence to NICE by Novartis. Third, some of the grade 3/4 AE data used in the Novartis Bucher indirect comparison for the sunitinib and placebo arms were different from the corresponding data submitted by Pfizer32 to NICE. When the Pfizer data are used instead, the pooled AE OR estimated by Novartis as 4.47 changes to 1.37, thus reducing the differences in costs and disutilities of AEs between sunitinib and everolimus.

The way that Novartis synthesised the effectiveness and safety evidence in its cost-effectiveness model inadequately reflected the available information. The company’s base-case assumption was that the PFS and OS outcomes of sunitinib and everolimus were the same, on the basis of wide CIs around the point estimates of relative effectiveness. This practice is clearly inadequate because it misrepresents the level of uncertainty in the data as evidence of lack of effect. This issue is made more serious in view of the direction and extent of possible bias resulting from the use of inadequate data, discussed in the previous paragraph.

In terms of model implementation, the limitations of the Novartis model analysis include the assumption of the same treatment duration for everolimus and sunitinib. As pointed out earlier, the mean number of treatment cycles of sunitinib is likely to be lower than that of everolimus, based on the Pfizer data32 submitted to NICE and comparison of median treatment durations in the main publications of A618111181 and RADIANT-3. 34 On the other hand, Novartis did not account for the fact that the mean PFS (area under the Kaplan–Meier curve) in the placebo arm of the A618111181 trial of sunitinib was lower than the mean PFS of the placebo arm of the RADIANT-334 trial of everolimus. If treatment duration is proportional to PFS, a fair indirect comparison with everolimus would require an increase in sunitinib treatment duration in proportion to the magnitudes of PFS in the placebo arm of RADIANT-334 and PFS of the placebo arm of A6181111. 81

Another limitation was the implementation of subsequent treatment costs. In the partitioned survival model used by Novartis the number of people who transitioned into progressive disease and who were eligible to receive subsequent treatment was not obtainable from the model output and had to be approximated using summary information reported in the trial about the number of people who were censored, died before progression and experienced a PFS event. This approximation involved the strong assumption of a constant relative frequency of these events throughout the PFS horizon.

Despite its limitations, the evidence presented by Novartis suggests that, with the currently available information, the choice between sunitinib and everolimus hinges on their relative effects on PFS and OS and drug acquisition costs and is subject to high levels of uncertainty related to clinical effectiveness. The disutility of AEs is unlikely to be a significant factor in that choice and determining its importance is hampered by the lack of data of sufficient quality for meaningful assessment.

Results

Novartis reports that the life expectancy over a 30-year period for patients with advanced, progressive, well-differentiated, non-functional GI and lung NETs is 5.79 life-years for 10 mg of everolimus daily plus BSC compared with 4.77 life-years for BSC alone (Table 12). The respective total discounted costs are expected to be £59,720 and £25,817 and the total QALYs produced are 4.28 and 3.51. This results in an ICER of £43,642 per QALY gained for everolimus plus BSC compared with BSC alone. Under the (confidential information has been removed) PAS discount on the everolimus price, the cost per QALY gained with everolimus would be reduced to (confidential information has been removed). In total, 71% of the QALY gain for everolimus plus BSC compared with BSC takes place in the stable disease state and the share of the cost falling during the stable disease state is 98% at list prices. The costs of active initial treatment with everolimus and BSC costs represent 79% and 8%, respectively, of the total incremental costs of treatment with everolimus plus BSC. The mean PSA ICER for everolimus plus BSC compared with BSC was £45,385 without the PAS and (confidential information has been removed) with the PAS.

Univariate deterministic sensitivity analyses found that the ICER was most sensitive to the choice of distribution for extrapolating OS. The ICER varied from £39,571 to £59,832 with different OS distributions and (confidential information has been removed) with the PAS. When treatment-specific utility values were applied, the ICER increased to £56,385 and (confidential information has been removed) under the PAS. The results were also sensitive to the RDI. Novartis also reports how extending the lifetime horizon and extrapolating outcomes beyond the trial period improved the cost-effectiveness of everolimus.

From the results reported by Novartis it is evident that, in the analysis of the PAS, the (confidential information has been removed) PAS discount was applied only to everolimus given as initial treatment, not as subsequent treatment. In principle, the discount should have been applied to subsequent treatment too.

The company concluded that the ICER (confidential information has been removed). The company states that the results should be considered within the context of unmet medical need for effective treatment options in this heterogeneous and small patient population, with there being approximately 936 patients in England across the two indications (pNETs and GI/lung) [table 3.1 (p. 27) of the Novartis submission33].

Strengths and weaknesses of Novartis’s evaluation of gastrointestinal and lung neuroendocrine tumours

The economic evaluation by Novartis of everolimus in GI and lung NET patients relies on the quality of RADIANT-4,35 which was the source of the effectiveness, AE and treatment duration and intensity data for the model and rates of subsequent treatment use. A major limitation is the omission of relevant active comparators, such as 177Lu-DOTATATE, in the analysis. Another limitation of the study design is the lack of a separate analysis of lung and GI patients.

In terms of data, the main limitation is the lack of resource use data measured in a sample of patients. It is not clear how robust the estimated costs of subsequent treatment use are likely to be, with issues such as administrative censoring (i.e. from termination of the study), and, indeed, whether or not the differences captured may have been an artefact of the length of follow-up.

In terms of evidence synthesis, the decision analysis relied on applying the same parametric survival distributions for extrapolation to both arms, which may have unnecessarily restricted the modelling capabilities in this study. Another issue was that, although crossover from everolimus to placebo was not permitted in the trial, 10% (10/97) of patients in the placebo arm did cross over (four before and six after unmasking). In spite of this, the analysis of OS data in RADIANT-435 did not adjust for such treatment crossover. This limitation may slightly bias the results if the analysis is intended to inform the evaluation of two alternative states of the world, one in which everolimus is provided as initial treatment and another in which it is not provided at all. If instead the NICE decision is between choosing everolimus as initial treatment or everolimus at the discretion of the physician as a potential subsequent treatment, the lack of adjustment for crossover by Novartis may not need to be a source of bias per se.

Another minor limitation is the inaccuracy in estimating the costs of subsequent treatments in the Novartis GI and lung model. This issue is the same as that discussed in the previous section for the Novartis model in pNETs and is not repeated here.

On the other hand, data on BSC treatment use, everolimus treatment duration and intensity and the incidence of grade 3/4 AEs in RADIANT-435 are detailed and Novartis describes important sources of uncertainty in the evidence base.

Mean age

We assumed that all patients were aged 60 years at the start of treatment. The estimate of the mean age of patients in the model population was based on the patient characteristics from the clinical trials used in our analysis. In the model, this affects only age-related utilities and background mortality. Mean age estimates from the companies’ models varied from 61.7 to 63.7 years.

Sex composition

All analyses were performed assuming that the proportion of male patients is 53% (as in RADIANT-334), which affects only background mortality (see Appendix 11, Table 98). The models in the company submissions used proportions of 50–51%.

Baseline randomised controlled trials

RADIANT-334 was chosen as the baseline trial for the pNETs population, whereas RADIANT-435 was chosen as the baseline trial for the GI (midgut) NETs and GI/lung NETs analyses. The size of the study populations in these trials was larger and the data were more mature than in A618111181 and NETTER-1. 102 In addition, the control arm in NETTER-1102 received 60 mg of octreotide daily plus BSC, which is outside the NICE scope for this review.

Modelled progression-free survival and overall survival

A partitioned survival approach was used to populate the parameters of the semi-Markov model. PFS Kaplan–Meier curves for the trial arms of the main RCTs informing the company submissions on pNETs (including A6181111 and RADIANT-3), GI/lung NETs (RADIANT-4) and GI (midgut) NETs (RADIANT-4 midgut subgroup and NETTER-1) were extracted from graphs in the latest available source for each trial [peer-reviewed publications for A6181111,81 RADIANT-334 and RADIANT-4;35 a published conference abstract for the RADIANT-4 midgut subgroup;74 and an industry submission (NETTER-1)] using digitising software (DigitizeIt version 2.2.2; Braunschweig, Germany). The extracted data were used to recreate the associated original individual patient data using the Guyot algorithm126 implemented in R (The R Foundation for Statistical Computing, Vienna, Austria).

A range of parametric curves from the proportional hazards (Weibull, exponential and Gompertz), piecewise proportional hazards (restricted cubic splines with five pieces or knots) and accelerated failure time (AFT) (log-normal, log-logistic and generalised gamma) families were fit to the recreated individual patient data for each arm separately and evaluated for use in the base-case analysis according to goodness of fit criteria [Akaike information criterion (AIC) and Bayesian information criterion (BIC)], visual fit to the empirical data (i.e. the instantaneous probability of event occurrence and Kaplan–Meier curve), plausibility of long-term extrapolation and consistency between PFS and OS (i.e. no crossing of PFS and OS curves of the same trial arm). We also consulted our clinical experts for their opinion about the plausibility of the long-term extrapolations associated with candidate functions. Finally, for our base-case analysis, we adopted the recommended practice that the same parametric function be used to extrapolate trial data for all arms in a comparison, to avoid introducing subjective assumptions in the long-term effectiveness estimates. 127 We relaxed this restriction in scenario analyses. The following is a summary of the main results of the two time-to-event outcomes in each of the three locations analysed.

In the RCT sources of effectiveness data for the pNETs model, treatment switching from the placebo arm to the active treatment arm was observed after disease progression (89% in RADIANT-334 and 65% in A618111181). Therefore, the following analyses for pNETs are based on Kaplan–Meier OS curves adjusted for crossover, using the RPSFT method (Robins and Tsiatis128). This approach affects only the placebo arm of each trial and produces a counterfactual placebo Kaplan–Meier curve, that is, the curve that would have occurred had no patient switched to the active arm.

In contrast, the analysis of GI or lung NETs is based on RADIANT-4,35 in which 6% of placebo patients switched to the active arm after disease progression. As no RADIANT-4 OS data were identified in which treatment switching in the placebo arm was adjusted for, the following analyses are based on the most recent ITT OS curves reported for RADIANT-435 (cut-off date 20 November 2015). As for the PFS Kaplan–Meier curves used in our analysis, switching to new active antineoplastic therapy before disease progression occurred in (confidential information has been removed) of patients in the everolimus arm and (confidential information has been removed) in the placebo arm (central radiological review), and these cases were censored from the analysis at the time of switch [RADIANT-435 final clinical study report (CSR), p. 71].

The statistical analysis used to select parametric survival functions for modelling PFS and OS outcomes and to extrapolate beyond the trial follow-up periods is presented next.

Everolimus plus best supportive care

The log-logistic model has the most favourable goodness-of-fit results (i.e. lowest value of information criteria) for the data from the everolimus arm of RADIANT-3,34 although its advantage over the log-normal and Weibull models is not statistically significant (see Appendix 12, Figure 38). The log-logistic model also follows the shape of the instantaneous risk (hazard) of progression or death (see Appendix 12, Figure 38).

The log-logistic and log-normal models perform similarly to each other in fitting the Kaplan–Meier PFS curve, with the Weibull model fitting the data almost as well as the log-logistic and log-normal models. However, by the end of the observed follow-up period, the risk of progression or death with the Weibull model is increasing, whereas the risk of progression or death with the log-normal and log-logistic models is declining (see Appendix 12, Figure 38).

By the end of the observation period, almost 2 years after randomisation, 20% of patients in the everolimus arm are alive and their disease has not progressed (Figure 12). Thus, adopting the log-logistic or log-normal model has noticeable implications for the long-term modelling of life free of disease progression. By the end of a 10-year follow-up, 3.5% of patients would be alive and progression free with the log-logistic or log-normal (not shown) model, whereas according to the Weibull model all patients would have progressed or died by 6 years after randomisation (Figure 13).

FIGURE 12.

Everolimus arm in RADIANT-3:34 Kaplan–Meier and best-fitting parametric PFS curves.

FIGURE 13.

Progression-free survival in the everolimus arm in RADIANT-3:34 extrapolation to 10 years.

Best supportive care alone

The BSC-only arm was modelled from data from the placebo arm in RADIANT-3. 34 The information criteria favour the log-normal, log-logistic and generalised gamma models over the rest (see Table 21). The AIC and BIC statistics, however, do not discriminate between the favoured models (their magnitudes differ by < 5 points).

The hazard function is non-constant and non-monotonic, which suggests that the Weibull and exponential models are inappropriate models for these data (see Appendix 12, Figure 39). The information criteria statistics are consistent with this observation and suggest that the log-normal or gamma model fit the data best.

Consistent with the model diagnostics of Table 21, the generalised gamma model is a closer match to the hazard function (see Appendix 12, Figure 39). The log-normal model approximates the smoothed hazard in Figure 39 (see Appendix 12), except for the drop between week 20 and week 40.

As illustrated in Figure 14, the Weibull model underestimates PFS early on, overestimates it in the medium term and underestimates it in the final part of the analytical horizon. The log-normal and generalised gamma models differ only in the final part; the log-normal model appears to fit better the final part of the Kaplan–Meier curve. Nevertheless, the choice of curve has little impact on mean PFS in this case.

FIGURE 14.

Placebo arm in RADIANT-3:34 Kaplan–Meier and best-fitting parametric PFS curves.

Sunitinib plus best supportive care

As presented in Figure 15, for the PFS data from the sunitinib arm of A6181111,81 the generalised gamma function provides the best diagnostic results, although the differences between the generalised gamma model and the exponential and log-normal models are not significant.

FIGURE 15.

Sunitinib arm in A6181111:81 Kaplan–Meier and best-fitting parametric PFS curves.

As depicted in Figure 40 (see Appendix 12), the generalised gamma model consistently underestimates the risk of progression, whereas the exponential model, with its constant risk, fits the pattern of risk up to approximately week 30 and underestimates it thereafter. The log-logistic model seems to follow the shape of the hazard function for longer periods than the other models.

The implications of adopting one of these curves to extrapolate outcomes for cost-effectiveness analysis is illustrated in Figure 16. The generalised gamma model’s parametric flexibility appears to produce an overly optimistic forecast of approximately 35% of patients still alive and not experiencing disease progression after 10 years. The exponential model, in contrast, predicts that by 5 years 95% of people have experienced progression or died. The predictions of the log-logistic model fall in between the predictions of the generalised gamma model and the exponential model, but much closer to the exponential forecast than the generalised gamma forecast.

FIGURE 16.

Progression-free survival in the sunitinib arm: extrapolation to 10 years.

The apparent contradiction between the diagnostic results, which suggest that the generalised gamma function fits the data best, and the hazard and survival function fits, which suggest that the log-logistic form is superior, appears to be determined by the ability of the gamma form to fit the data better in the early follow-up period, when more observations are available (e.g. number at risk: 85 at baseline vs. 39 at 22 weeks). Its poor ability to match the risk (hazard) profile and the latter part of the survival curve suggests, however, that it overfitted the data. The exponential curve was thus selected for the base-case analysis.

Adjustment for indirect comparisons

To derive estimates of PFS time for the sunitinib arm that were comparable to the PFS estimates in RADIANT-3,34 the sunitinib parametric PFS distribution was adjusted by the ratio of the area under the non-parametric Kaplan–Meier curve of the placebo arm in A618111181 to the area under the non-parametric Kaplan–Meier curve of the placebo arm in RADIANT-334 at the shortest of the maximum follow-up times across the two placebo arms. This method was preferred to the common alternative approach of using the extrapolated means, which by definition are affected by the choice of parametric function as opposed to being determined solely by the observed data, as in our ‘restricted means’ approach. Thus, in the base case, in which the sunitinib PFS parametric distribution was the exponential, the sunitinib PFS parametric distribution function was adjusted according to the equation:

where λ^s is the adjusted hazard function of the exponential time-to-disease progression or death distribution, λ_s is the hazard function of the exponential distribution estimated from the sunitinib arm of A618111181 and the AUC(tmin{TAp,TRp)Ap and AUC(tmin{TAp,TRp)Rp functions are the area under the Kaplan–Meier curve of the placebo arms of A618111181 and RADIANT-3,34 respectively, evaluated at the shortest of the maximum observation times in the placebo arms of the two trials (i.e. 65 weeks). This is illustrated in Figure 17, where the vertical discontinuous line denotes the 65th-week time point. At this point, the mean PFS in the placebo arm is 25.69 weeks in A618111181 and 28.85 weeks in RADIANT-3. 34 Thus, the PFS distribution of the sunitinib arm in the base case is (t)=exp(−λ^t)=exp(−0.01197*t). The same approach was used for OS.

FIGURE 17.

Progression-free survival in the placebo + BSC arms of pNETs trials.

Individual patient data provided by Pfizer to the AG for this assessment allowed us to investigate the robustness of indirect comparisons of PFS outcomes between the sunitinib arm in A618111181 and the everolimus and placebo arms in RADIANT-3. 34 We conducted a MAIC following the methods described in a previous study by Signorovitch et al. 129 We obtained similar results (which are academic-in-confidence) to those reported here.

Everolimus plus best supportive care

As illustrated in Figure 18, according to the exponential function, by 15 years (180 months) 4% of patients initially treated with everolimus remain alive, whereas according to the log-normal survival curve, 10% of patients would be alive at that time point. In its submission, Novartis33 cites an estimated 15-year survival rate of 6% for patients with advanced pNETs from the Surveillance, Epidemiology, and End Results Program (SEER) database, and considers as plausible extrapolating only those curves predicting survival rates above that level. However, there is the caveat that:

the RADIANT-3 trial only included patients with well- or moderately-differentiated tumours with radiologic progression within the 12 months prior to entry in the study. (Confidential information has been removed.) In contrast, patients in the SEER registry are followed from diagnosis, and are therefore likely to be treatment-naïve. Also SEER includes all advanced patients regardless of tumour grade and SEER data were derived over a relatively long period beginning in 1973 whereas follow-up in RADIANT-3 began in 2009.

Novartis submission, p. 9333

FIGURE 18.

Overall survival in the everolimus arm of RADIANT-3:34 extrapolation to 20 years.

[In English life tables, the 15-year survival rate in the general population aged 60 years increased from 56% in males and 74% in females in 1980–2, to 75% and 83%, respectively, in 2006–8 (Office for National Statistics130)]. The company states that the net impact of these differences in patient characteristics is unknown. It concludes that ‘it is reasonable to assume that survival in the everolimus arm of the RADIANT-334 trial would not be substantially less than that for patients with advanced pNETs in the SEER registry’ (Novartis submission, p. 9333).

Best supportive care alone

The log-normal function had the best fit to the data of the placebo plus BSC arm in RADIANT-3. 34 The generalised gamma function had a good visual fit to the risk of death observed in the trial period (see Appendix 12, Figure 42) but an overly optimistic 20-year OS rate of 20% (Figure 19). The exponential function underestimated the hazard risk throughout the trial period (see Appendix 12, Figure 42), but had a 20-year PFS in the middle of those depicted in Figure 19.

FIGURE 19.

Rank-preserving structural failure time (RPSFT)-adjusted OS in the placebo arm of RADIANT-3:34 extrapolation to 20 years.

Sunitinib plus best supportive care

The log-normal and exponential functions had the best fit to the OS data for the sunitinib arm in A618111188 (Table 21), although the log-normal function tracked the risk of death (hazard) observed in the trial better than the exponential function (see Appendix 12, Figure 43). The projected 15-year survival rate with the log-normal function is > 10%, whereas with the exponential function it is 4.7% (see Figure 20). However, the 15-year OS rate used in the model was higher than that seen with the exponential function as, after adjusting the exponential function for the difference in the placebo arms of the A618111181 and RADIANT-334 trials (see the following section), the OS rate was 9.7%.

FIGURE 20.

Overall survival in the sunitinib arm of A6181111:81 extrapolation to 20 years.

Adjustment for indirect comparison

As before, to derive comparable OS estimates for the treatments in RADIANT-3,34 the OS exponential curve for sunitinib was adjusted to reflect the difference in RPSFT-adjusted OS between the placebo arm of A618111181 and the placebo arm of RADIANT-3. 34 Figure 21 illustrates the difference, and the vertical discontinuous line marks the point at which the area under the curve calculation was restricted for both arms (24 months).

FIGURE 21.

Kaplan–Meier RPSFT-adjusted OS in the placebo + BSC arms of the pNETs trials.

Progression-free survival

In the base-case analysis, the Weibull function was chosen for the everolimus plus BSC arm and the BSC-only arm based on data from RADIANT-4. 35 In scenario analyses, the generalised gamma function was used instead for both model arms. See Appendix 13 for details.

Overall survival

The base-case analysis adopted exponential distributions separately fitted to OS data in the everolimus arm and the placebo arm of RADIANT-4. 35 The Gompertz and Weibull distributions had the best goodness of fit statistics and seemed to provide the best fit to the everolimus hazard rates (see Appendix 12, Figure 46) and Kaplan–Meier curves (Figure 22), whereas the exponential function seemed to be the best fit to the placebo data (see Figures 23 and 47). However, only the extrapolations of the exponential and log-logistic distributions seemed plausible, as discussed in the following two sections. In scenario analyses, log-logistic distributions separately estimated for the two trial arms were adopted.

Everolimus plus best supportive care

To reflect the uncertainty resulting from immature data, Figure 22 presents the OS extrapolations for all available parametric curves. The exponential function appears to overestimate the risk of death (see Appendix 12, Figure 46) and underestimate the Kaplan–Meier OS curve (see Figure 22) in the early part of the trial observation period35 (data cut-off 30 November 2015), although the discrepancy is within the sampling error (95% CI, not presented). The exponential curve crosses the log-logistic curve twice, once during the interpolation (within-trial) period and once in the late extrapolation (beyond-trial) period. In its submission to NICE, Novartis33 turns to external data to inform its choice of survival curves. In particular, it states that:

Analysis of distant NET cases diagnosed between 1997 and 2012 in the SEER database (a large population-based registry in the USA) suggests that survival for patients with distant disease at diagnosis at 15 years is approximately 10% (Data unpublished). Although it is difficult to make comparisons between the RADIANT-4 trial population and the available SEER data (see Appendix 9 of Novartis submission for further details), it is reasonable to assume that survival in the placebo plus BSC arm of the RADIANT-4 trial is likely to be no less than that of patients with distant disease in the SEER database given improvements in survival over time in patients with NET.

Novartis submission (p. 137)33

FIGURE 22.

Overall survival in the everolimus arm of RADIANT-4:35 extrapolation to 20 years.

In appendix 9 of the Novartis submission, Novartis33 reports the data and methods used to obtain its 10% survival benchmark at 15 years; Kaplan–Meier survival curves by location from SEER were weighted according to the distribution of patients by location in RADIANT-4. 35 Novartis also acknowledges the limitations of these SEER data, as discussed earlier for pNETs.

Best supportive care alone

High degrees of uncertainty are similarly present in the later parts of the follow-up period for patients in the placebo arm of RADIANT-4,35 as evidenced by the large steps observed after approximately 2 years of follow-up (Figure 20).

FIGURE 23.

Overall survival in the placebo arm of RADIANT-4:35 extrapolation to 20 years.

Progression-free survival

In the base-case analysis the exponential distribution was adopted to model PFS outcomes in the everolimus arm and the placebo arm of RADIANT-4. 73 PFS in the 177Lu-DOTATATE arm of NETTER-1102 was also modelled with an exponential distribution (see Appendix 11 for scenario analyses).

Everolimus plus best supportive care

The exponential function, which was the function with the best statistical fit (Figure 24), appeared to have a poor fit to the hazard rates for everolimus in RADIANT-473 (see Appendix 12, Figure 48). However, this is caused by the small sample available in the latter part of the RADIANT-473 follow-up period. Of the candidate functions, the log-normal function has the longest PFS duration, followed by the exponential and the Weibull functions (see Figure 24).

FIGURE 24.

Progression-free survival in the everolimus arm of RADIANT-435 [GI (midgut)]: extrapolation to 20 years.

Best supportive care alone

The diagnostic statistics (see Table 20) did not discriminate between the [accelerated failure time (AFT) or proportional hazards] models available to represent the PFS data in the placebo arm of RADIANT-4,73 although the generalised gamma and log-normal functions displayed PFS hazard rates that were similar to those observed during the trial (see Appendix 12, Figure 49). When turning to the extrapolation of PFS, the generalised gamma function seems to result in overly optimistic disease PFS rates (see Figure 25). In scenario analyses, the log-normal distribution was used instead of the exponential to model the PFS experience of patients in the placebo arm.

FIGURE 25.

Progression-free survival in the placebo arm of RADIANT-435 [GI (midgut)]: extrapolation to 20 years.

Overall survival

In the absence of data, the base-case analysis currently assumes that the OS curve for everolimus in the GI (midgut)-only location is the exponential OS curve estimated in the GI or lung patient population in the everolimus arm of RADIANT-4,35 discussed earlier, adjusted by the proportional difference in mean PFS in the everolimus arm of RADIANT-435 between the overall GI/lung patient population and the subgroup of GI (midgut) patients. Likewise, the OS curve for the BSC-only arm of RADIANT-435 in the GI (midgut)-only population is derived from adjusting the exponential function fitted to the OS data from the everolimus arm of RADIANT-435 in the GI/lung population considered above by the proportional difference in mean PFS between the overall GI/lung and the GI (midgut) patient groups. These derivations follow the same steps as in Equation 1. We requested OS data for the GI (midgut)-only location from Novartis, but the company did not provide them to us.

Adverse events

The probabilities of AEs were used to estimate costs in the stable disease state. For the economic evaluation of treatments for pNETs, probabilities of AEs were derived from rates estimated from our indirect comparison of treatment-related grade 3/4 AEs of ≥ 2% incidence in any active treatment arm (see Chapter 4, Indirect treatment comparison: pancreatic neuroendocrine tumours; Appendix 21, Table 136 and Appendix 9). We updated these analyses with data provided in the Pfizer submission to NICE. 32 For the GI/lung analysis, the AG model adopted the probabilities in the Novartis model submitted to NICE,33 as these were calculated with individual patient data not available to the AG. For everolimus plus BSC and BSC only in the GI (midgut) evaluation, we adopted the grade 3/4 AE rates for the everolimus and placebo arms reported in a recent conference poster by RADIANT-4 investigators;73 for 177Lu-DOTATATE we used the grade 3/4 AE rates reported in the AAA submission to NICE. 31

The AE probabilities for the pNETs model were obtained by assuming that patients had no multiple instances of the same AE type and that AEs lasted for only one cycle. This seems a reasonable assumption in the light of the evidence in the CSR for A6181111,81 which reports the actual duration of the grade 3/4 AEs recorded in the trial (tables 5, 24, 26 and 29 in the full CSR of A61811181). For GI/lung and GI (midgut) NETs the same assumption was adopted.

Based on calculations by the AG the measured differences in grade 3/4 AEs between everolimus and sunitinib in patients with pNETs considered by Novartis in its submission,33 and for which disutility values are available (see Utility values for the Peninsula Technology Assessment Group model for details), are associated with negligible differences in utility, equal to a 0.002 quality-adjusted months, and were therefore not used in calculating utility values for stable disease in the pNETs model. For GI/lung and GI (midgut) analyses, the available utility values, which were derived from patient-reported outcomes in RADIANT-435 and evidence from the ERASMUS study31 submitted to NICE by AAA, were assumed to capture the impact of AEs.

Modelling post progression

Based on data from RADIANT-334 for pNETs and RADIANT-435 for GI and lung NETs, which we assumed applied to GI (midgut) NETs, in the base-case analysis we assumed that all patients have BSC after progression on initial treatment. This consists of palliative care and 30 mg of octreotide for symptomatic treatment, with no active drug treatment.

Subsequent treatments were allowed in the post-progression phase and were applied as a fixed cost in the first cycle after disease progression. The frequency of subsequent treatment use was assumed to be zero in the base-case analysis; scenario analyses considered applying the same costs of subsequent treatments as in the pNETs and GI and lung models from Novartis. 33 This choice of base case reflected the fact that (1) the A618111181 trial of sunitinib did not collect information on subsequent treatments, which led Novartis to apply the same costs of subsequent treatments to both arms, and (2) the way that Novartis implemented subsequent treatment costs in its models is unreliable (see Chapter 5, Critical appraisal of the company submissions). In the GI and lung and GI (midgut)-only analyses, we adopted the same costs and implementation of subsequent treatments as in the GI and lung model of Novartis, which used detailed information on differences between trial arms in the frequency of treatment use post progression in RADIANT-4. 35 As we had no information for 177Lu-DOTATATE, we assumed that it had the same subsequent treatment costs as applied to everolimus.

The costs of disease monitoring in our model were obtained from the pNETs and GI and lung models from Novartis. 33 Based on the opinion of our clinical experts, we adopted a smaller number of visits for the GI and lung evaluation than were used by Novartis in its model of the same location (see Costs of medical management and disease monitoring).

Utility values for the Peninsula Technology Assessment Group model

Of the eight independent sources of data identified through the systematic search of utility studies, only a limited proportion of evidence was suitable to populate the PenTAG cost-effectiveness model. This was mainly because of the differences in the treatments being evaluated and the discrepancies between the definitions of HRQoL outcome measures in the studies and the model’s health states.

The utility values used in the PenTAG cost-effectiveness model and their sources are presented by tumour location, in accordance with the NICE scope. No utility values for HRQoL outcomes measured in pNET patients under treatment with everolimus were found (Table 21). The only available estimates for everolimus in pNETs were those reported by the preference elicitation study of Swinburn et al. ,131 who asked approximately 100 members of the general public to assess descriptors (vignettes) of pre-progression and post-progression health states of patients with GI NETs and pNETs, using the time trade-off method. This study also elicited utility decrements resulting from the occurrence of some of the most common AEs associated with treatment therapies (see Appendix 10, Pancreatic studies, for details; AEs considered include neutropenia, hypertension, palmar–plantar erythrodysaesthesia syndrome, leukopenia, diarrhoea, stomatitis, thrombocytopenia, anaemia, hyperglycaemia, fatigue, infections, pneumonitis and nausea). In the base-case analysis submitted to NICE, Novartis33 used the utility value reported by this source for the pre-progression health state, adjusted for the disutility associated with the incidence of the common types of grade 3 or 4 AEs in RADIANT-3. 34 As these values do not meet the NICE reference case37 we considered them only in scenario analyses.

In the absence of HRQoL outcomes measured in pNET patients treated with everolimus, the base-case value was assumed to be the same as for sunitinib (see Table 21 and its description). This assumption was adopted after calculating the net difference in disutility from grade 3/4 AEs between everolimus and sunitinib according to the disutility values reported by Swinburn et al. ,131 which was equal to 0.002 quality-adjusted months. Giving the uncertainty associated with other parameters in the model, including the limited quality of AE data available for economic evaluation purposes, we considered such a difference insignificant.

Estimates of the utility of pNET patients undergoing treatment with sunitinib in both the pre-progression and the post-progression health states (see Table 21) were obtained by mapping individual patient data onto responses to the EORTC QLQ-C30 from A618111181 provided by the manufacturer (Pfizer, data request through NICE) using the algorithm of McKenzie and van der Pol. 132 Using these data the utility values that the company used in the model-based cost-effectiveness evidence submitted to NICE were validated. Although the methods used by the company were not properly described in the submission itself,32 the values used and the methods had been reported in the study by Mucino Ortega et al. 121 The validation exercise therefore sought to replicate the utility values submitted to NICE by following the methods described by Mucino Ortega et al. 121 This consisted of fitting a linear mixed-model equation to the EORTC QLQ-C30 data mapped to the EuroQol-5 Dimensions (EQ-5D) using the algorithm developed by McKenzie and van der Pol132 to estimate the effect of random group allocation (sunitinib vs. placebo) on EQ-5D utilities, adjusting for baseline EQ-5D score and treatment cycle (information available from the authors). Confidential information has been removed. Nonetheless, the estimate obtained employing the model is an approximation, as the estimate of utility in progressive disease was based only on data for the end-of-treatment follow-up time point relating to the placebo arm. This approximation relies on the fact that > 90% of patients in the placebo arm had progressed by the end of the study.

Confidential information has been removed. This difference is the result of the incorrect use of the data in the analysis conducted by the manufacturer of sunitinib (Pfizer) in its submission to the Scottish Medicines Consortium (SMC)133 and used by Novartis in its sensitivity analysis of the economic evaluation submitted to NICE. 33 The company incorrectly used baseline utility as the utility for the pre-progression health state, thus omitting the effects of treatment on patients’ utility during stable disease. In contrast, our replication of the utility values in Mucino Ortega et al. ’s121 study led us to the following estimated linear mixed-model equation for stable disease:

where EQ-5Dm is the mapped utility score from the EORTC QOLQ-C30 in A6181111,81 using the algorithm from McKenzie and van der Pol. 132 To estimate utilities for the two trial arms in the stable disease state, this model was fitted to data excluding the last follow-up, that is, the end-of-treatment follow-up observations, when some patients may have experienced disease progression, resulting in their withdrawal from treatment.

For GI and lung NETs patients under treatment with everolimus (Table 22), we used unpublished treatment-specific utility values, which were presented by Novartis as part of the evidence submitted to NICE and used by the company in sensitivity analyses of its economic evaluation. These data were preferred to published estimates from pooled analysis74 (i.e. utilities by health state regardless of treatment arm) used by the company in its base-case economic model submission, as they incorporate the impact of treatment-specific AEs and comorbidities on HRQoL. The treatment-specific utilities in the company’s submission were based on unpublished individual patient data from RADIANT-435 that were not available to us for review. To validate such estimates, we mapped mean FACT-G scores in RADIANT-4,35 reported by Singh et al. 74 in a poster also reporting pooled utilities by health state, using a linearised version of the algorithm from Longworth et al. 134 (see Appendix 15). The values that we obtained were approximately equal to those produced by the company from individual patient data in the pooled analysis with the original, non-linear mapping algorithm (see Appendix 22).

In the absence of data specific to lung or to GI-only patients for everolimus plus BSC and BSC only, we assumed the same utility values for these subgroups as for the overall RADIANT-435 population.

No data on HRQoL for patients treated with 177Lu-DOTATATE were identified through the systematic review. For this reason, we had to rely on unpublished data submitted by AAA. 31 The utility values in the PenTAG base-case model are based on a Dutch single-arm, uncontrolled study that has not been published at the time of writing. Utility values for pre-progression GI NET patients (see Table 22) were measured in the ERASMUS study, a single-centre, non-controlled, Phase I/II open-label study, conducted in 810 Dutch patients with different somatostatin receptor-positive tumour types. These data were used to estimate the utility of GI NET patients in the pre- and post-progression health states in the base-case cost-effectiveness model. Empirical data on HRQoL associated with 177Lu-DOTATATE in GI NET patients obtained from the Guy’s and St Thomas’ (UK) hospital registry were used by the company to estimate the utility of pre-progression patients. As no justification for this inconsistency in the choice of sources between the two health states was provided, the evidence from the ERASMUS study was preferred.

The utilities adopted for the de novo model by the AG, presented in Tables 21 and 22, were further adjusted for the effect of ageing in the model using the following linear equation, which was estimated by the AG from EQ-5D data in the Health Surgery for England 2012135 following the approach of Ara and Brazier:136

This adjustment was applied to all utilities irrespective of health state or treatment arm.

Summary

There is a lack of published evidence on HRQoL, especially for progressive disease in pNETs. In addition, for one of the comparators evaluated in this location, everolimus, there is no evidence available on the HRQoL outcomes in actual patients. In contrast, the present review benefited from access to individual patient data on HRQoL outcomes in patients from one of the main trials in the assessment, A6181111,81 provided by Pfizer through a NICE request.

The AG was able to validate the utilities derived by Novartis for the progressive disease and stable disease states in the GI location from RADIANT-435 data, without having access to the individual patient data but only aggregate HRQoL domain scores, by a linear approximation to the best-fitting algorithm from Longworth et al. ,134 used by Novartis to map individual FACT-G scores onto the EQ-5D. Linearising the best-fitting non-linear algorithm using first-order approximations enabled the successful validation of published mapped utilities.

In the absence of data specific to lung or to GI-only patients for everolimus plus BSC and BSC only, we assumed the same utility values for these subgroups as for the overall RADIANT-435 population.

The analyses of individual patient data highlighted the importance of requesting such data from the trial sponsors. This allowed the identification of fundamental errors in the interpretation of the data contained in the submissions to the three bodies responsible for making resource allocation decisions for England, Wales and Scotland.

Cost parameters and assumptions

The unit costs of treatments and resources were sourced according to the NICE Guide to the Methods of Technology Appraisal 2013. 37 Only costs that relate to the included interventions for the treatment of NETs, and to resources under the control of the NHS and PSS, were included. Value-added tax was excluded. Costs that were common and equal in all treatment strategies over the time horizon of the analysis were excluded. The cost-effectiveness results reflect the present value of costs and benefits accruing over the time horizon of the analysis.

We modelled the following costs, which were inflated to the cost year 2016, and used an annual discount rate of 3.5%:

• drug acquisition

• drug administration

• medical management and disease monitoring

• SAE management

• end-of-life care.

Costs of drug acquisition

Costs of drug administration

There is significant variation across the comparator treatments in the resource requirements for their administration. Everolimus and sunitinib are ingested orally as a tablet and capsule, respectively, and are usually self-administered, whereas 177Lu-DOTATATE is administered in the secondary care setting by intravenous infusion over 20–30 minutes. 139,140 The cost of hospital pharmacy dispensing, applied at each outpatient clinic visit, was included in our costing for the oral preparations. This required 12 minutes of hospital pharmacist time, equating to £14.40.

In contrast, the administration of 177Lu-DOTATATE is resource intensive. As a radiolabelled and intravenously delivered drug, it requires specialist oversight and a hospital setting. AAA31 concurs in its data submission (section 2, p. 25) by stating that its expectation of routine treatment is delivery ‘in a nuclear medicine department within a secondary care hospital as an outpatient appointment’. However, we are guided by expert clinical opinion (consultants in nuclear medicine) that current standard practice is to admit patients overnight. We understand that selected patients at a single specialist centre in England are managed as day cases and this approach may be expanded in the future.

Table 28 presents the cost of drug administration for 177Lu-DOTATATE. The estimated quantity of resource required is the average elicited from two NHS consultants in nuclear medicine. Unit costs were obtained from standard sources. 141,142

The costs associated with administering supportive treatments are presented in Table 29. Unit costs were obtained from standard sources. 141,142 Supportive treatments in Tables 25 and 114 but not listed here did not attract an administration cost; the cost of dispensing was not included for any supportive treatment.

Costs of medical management and disease monitoring

Medical management and disease monitoring resource estimates include the following categories of resource:

• hospitalisation: general and emergency

• outpatient clinic consultation

• procedures and tests

• other supportive procedures.

In the absence of published disease-specific detailed estimates of NHS resourcing, we relied on a source of unpublished evidence supplied by Novartis to tell us which resources are used and what the expected rates of utilisation are. In an industry-sponsored survey, nine physicians from seven UK centres were asked in 2016 to confirm the nature of NETs resourcing (type and rate) from a previous resource use survey of 32 clinicians in England in 2011. The validation process was framed in the context of personal practice in the previous year, across various disease stages and primary tumour locations.

The physicians were asked to list the various types of resource that a patient with NETs requires during the course of the disease and estimate the number of times that a patient would see each physician each month. The resource use for the given NETs patient population was then calculated as a weighted average of the annual number treated by each clinician relative to the total number treated annually across all clinicians. These estimates were then weighted according to the respective proportions of patients with pNETs in RADIANT-334 and GI NETs and lung NETs in RADIANT-4,35 to determine resource use for the overall trial populations.

For the BSC strategy in the pNETs evaluation, which was not modelled by Novartis, the resource utilisation of patients with stable disease was assumed to be the same as that presented for patients on active treatment. For patients with pNETs who progressed on BSC, resource utilisation was assumed to be equal to that for those who progressed on active treatment.

In some instances we modified the raw survey findings for our modelling:

• The frequency of resource use for people who progressed following active treatment was adjusted downwards according to the proportion of people who resided ‘in observation’ in clinical trials (32.7% following progression on active treatment, 33.3% following BSC), which was effectively BSC.

• The frequencies of consultations and procedures and tests for people with stable pNETs receiving BSC were reduced to below the estimates for active treatment. The reductions were proportionate to the differences observed between active treatment and BSC in GI and lung and GI (midgut) NETs (approximately 4 : 1 for consultations and 2 : 1 for tests/procedures).

• The frequencies of consultation between people with GI and lung NETs/GI (midgut) NETs and the medical oncologist were adjusted downwards to the average of estimates from our expert clinicians. Utilisation survey estimates gathered by Novartis appeared high in absolute terms but also compared with frequencies in people with pNETs.

We used standard sources for the unit costing of disease management and monitoring. 141 These unit costs are presented in Table 30.

Rates of hospital resource use are presented in Table 31 for the pNETs evaluation and Table 32 for the GI and lung NETs and GI (midgut) NETs evaluations.

Costs of adverse event management

Adverse events experienced by patients on treatment attract additional health-care resources. To approximate the cost of managing those treatment-related AEs that could influence the cost-effectiveness of included treatments, we included only grade 3 and 4 AEs (SAEs) occurring in ≥ 2% of patients in either arm of the trials (National Cancer Institute Common Terminology Criteria for Adverse Events, version 4.03111).

In the evaluation of pNETs this included SAEs reported in RADIANT-334 and A6181111. 81 In the evaluation of GI and lung NETs, as well as GI (midgut) NETs, this included SAEs reported in RADIANT-435 and NETTER-1. 102

In pNETs, an ITC was conducted to match the trial populations of A618111181 and NETTER-1102 to the trial populations of the respective RADIANT trials (see Table 135). For each active treatment in the three evaluations, an OR was applied to the weighted average rate to give a relative rate by strategy for each event type (see Table 115). In GI and lung NETs (see Table 116), and GI (midgut) NETs (see Table 117), the unadjusted proportions of patients experiencing a SAE as reported in RADIANT-435 and NETTER-1102 were used.

Based on the assumption that no patient would report more than one SAE of any specific type during their time on treatment, we applied the costs of SAE management only to the initial Markov cycle in the progression-free health state.

Cost of end-of-life care

On the basis that the average cost of health resource use in the final weeks of the life of a cancer patient is a reasonable surrogate for patients with a NET, we used an estimate from the literature for cancer patients in England and Wales (£4346.19). This includes elective and non-elective inpatient admissions, outpatient appointments, accident and emergency visits and district nurse and general practitioner (GP) visits from the point at which a strong opioid is first used. 143