Selective internal radiation therapies for unresectable early-, intermediate- or advanced-stage hepatocellular carcinoma: systematic review, network meta-analysis and economic evaluation

The SARAH and SIRveNIB randomised controlled trials

Two large RCTs compared SIR-Spheres with sorafenib in patients who were not suitable for curative treatments: the SARAH trial was conducted in France19,20 and the SIRveNIB trial was conducted in the Asia-Pacific region. 21 Both trials were considered to have a low overall risk of bias (see Appendix 5). Further details of these trials are presented in Table 5.

As shown in Table 5, there were methodological differences between the SARAH and the SIRveNIB trials. In the SIRveNIB trial, patients could receive only one SIRT delivery, whereas in the SARAH trial patients could receive more than one delivery of SIRT; 69 out of 184 (37.5%) patients who received SIRT received more than one delivery to either the ipsilateral or the contralateral lobe.

The SARAH trial was conducted in France and the SIRveNIB trial was conducted in the Asia-Pacific region. This has implications for the generalisability of the SIRveNIB trial results to the UK population. HCC in European patients is more likely to be caused by alcohol or hepatitis C, whereas in Asia it is more likely to be caused by hepatitis B. The natural history of these diseases is different. Treatment options are also different, as hepatitis B-related liver disease is often less advanced than in alcohol-related or hepatitis C-related disease; therefore, patients may have had more treatment prior to receiving systemic therapy.

The Sirtex submission stated that patient selection in the SARAH trial did not reflect UK clinical practice, as the trial included patients with a poor survival prognosis who would be considered for only systemic therapy or BSC [e.g. because of a high tumour burden, main PVT or impaired liver function (Child–Pugh class B)]. Therefore, this has implications for the generalisability of the SARAH trial results to the UK population who would be eligible for SIRT in clinical practice.

In both trials, patients were assessed for suitability of SIRT after randomisation. In the SARAH trial, 53 out of 237 (22.4%) patients allocated to SIR-Spheres did not receive SIRT, 26 of whom were treated with sorafenib. In the SIRveNIB trial, 52 out of 182 (28.6%) patients allocated to SIR-Spheres did not receive SIRT, three of whom were treated with sorafenib (where reported; subsequent treatments were not reported for 31/52 patients). Results were presented for both the intention-to-treat (ITT) and the per-protocol populations; patients who did not receive their allocated treatment were excluded from the per-protocol analysis (those who received sorafenib instead of SIRT were not included in the sorafenib arm in the per-protocol analysis).

The SARAH and SIRveNIB trial publications reported baseline characteristics for both the ITT and the per-protocol populations. 19,21 The SIR-Spheres and sorafenib groups were generally similar at baseline in the ITT populations (see Table 5). However, in the per-protocol population, patients in the sorafenib arm appeared to have slightly worse disease characteristics than those in the SIR-Spheres arm in the SARAH trial (BCLC stage C: 69.4% vs. 65.5%; Child–Pugh class B7: 14.6% vs. 11.5%; median tumour burden: 20% vs. 12.5%, respectively) and in the SIRveNIB trial (BCLC stage C: 45.1% vs. 38.5%; PVT: 29.6% vs. 23.1%; tumour size > 50% of liver: 21.6% vs. 17.7%, respectively).

The SARAH randomised controlled trial subgroup analysis (low tumour burden/low albumin–bilirubin grade)

The Sirtex company submission selected a subgroup of patients from the SARAH trial with ≤ 25% tumour burden and albumin–bilirubin (ALBI) 1 for their base-case analysis in the economic model; the company stated that these patients are considered the most appropriate candidates for SIR-Spheres in clinical practice, as they are the most likely to benefit from SIRT. This is not a clinically recognised subgroup and was based on a post hoc analysis; therefore, these results should be prospectively validated before being considered relevant for clinical practice.

This subgroup included 37 (16%) patients in the SIRT group and 48 (22%) patients in the sorafenib group; 92% of those allocated to SIRT received treatment after work-up. Baseline characteristics were relatively well balanced between treatment groups, although more patients in the SIRT arm had BCLC stage B disease, single tumours and received previous TACE (these patients generally have a better prognosis than patients who are diagnosed at a later stage and are not eligible for TACE) than in the sorafenib arm. More patients in the sorafenib arm had an ECOG performance status of 0 and unilobar liver involvement. Table 6 presents the baseline characteristics and results for the full ITT population and the low tumour burden/low ALBI grade subgroup of the SARAH trial.

As shown in Table 6, median OS and PFS appeared to be better in the SIR-Spheres arm than in the sorafenib arm in the post hoc subgroup analysis, although the difference between treatment groups was not statistically significant. The proportion of patients who went on to have potentially curative therapy was higher in the SIR-Spheres arm than in the sorafenib arm, although numbers were very low (five and one patients, respectively). Tumour response rate, HRQoL and AEs were not reported separately for the low tumour burden/low ALBI grade subgroup.

Prespecified and post hoc subgroup analysis results were presented in the SARAH trial publication for OS. 19 Tumour burden was included as a post hoc subgroup. However, neither the ALBI grade nor the combination of low tumour burden and low ALBI grade was presented.

The SIRveNIB trial did not report subgroup analysis results for the subgroup of low tumour burden/low ALBI grade patients. However, ALBI grade was included in the OS subgroup analysis. Results favoured SIR-Spheres in the subgroup of ALBI 1 patients (HR 0.89, 95% CI 0.6 to 1.4; p = 0.58), whereas results favoured sorafenib for the subgroup of patients with ALBI 2/3 (HR 1.24, 95% CI 0.9 to 1.7; p = 0.14).

Other randomised controlled trials of SIR-Spheres

The SIRTACE is a small RCT rated as being at a high risk of bias that compared SIR-Spheres (n = 13) with TACE (n = 15) in patients with unresectable HCC without portal vein occlusion. 22 A higher proportion of patients in the SIRT group had BCLC stage A disease (38.5% vs. 26.7%) and Child–Pugh liver function class A (92.3% vs. 86.7%) than in the TACE group. The average number of tumour nodules was higher in the TACE group (5.0 vs. 3.5). Therefore, patients in the SIR-Spheres treatment arm had a better prognosis than those in the TACE arm.

At 6 months, 69.2% of SIRT patients and 86.7% of TACE patients were still alive. At 12 months, 46.2% of SIRT patients and 66.7% of TACE patients were still alive. PFS, disease control rate and the proportion of patients who went on to have potentially curative therapy were similar between treatment groups. The proportion of patients with a partial response was higher in the SIRT group than in the TACE group (30.8% vs. 13.3%), although patient numbers were very small.

There were no statistically significant differences between treatment groups in HRQoL by week 12, despite Functional Assessment of Cancer Therapy Hepatobiliary-Pancreatic Symptom Index (FACT-Hep) scores being lower in the SIRT group at baseline (indicating lower quality of life). However, 10 out of 28 patients had missing baseline data and were excluded from HRQoL analyses. The proportion of patients reporting TRAEs was higher in the TACE group than in the SIRT group (33.3% vs. 23.1%), although the proportion of patients reporting at least one AE was higher in the SIRT group (92.3% vs. 66.7%), as was the number of patients with grade ≥ 3 AEs (three vs. two patients) and serious AEs requiring hospitalisation (seven vs. five patients).

A small RCT by Pitton et al. ,23 with some concerns regarding bias, compared SIR-Spheres (n = 12) with DEB-TACE (n = 12) in patients with unresectable intermediate (BCLC stage B) HCC with preserved liver function (Child–Pugh class A–B7). Treatment groups appeared reasonably similar at baseline, although more patients in the SIRT group had received prior local ablation (four vs. one) and more patients in the DEB-TACE group had received prior resection (five vs. three). Median OS and PFS were longer in the DEB-TACE arm than in the SIR-Spheres arm (788 days vs. 592 days and 216 days vs. 180 days, respectively), although the difference between groups was not statistically significant. Median TTP was 371 days in the SIRT arm and 336 days in the DEB-TACE arm. AEs were not reported.

The SORAMIC RCT compared SIR-Spheres followed by sorafenib with sorafenib alone in patients with unresectable intermediate or advanced (BCLC stage B or C) HCC with preserved liver function (Child–Pugh class ≤ B7) and ECOG performance status of < 2, who were poor candidates for TACE. Only safety and tolerability data for the first 40 patients have been published to date, rated as being at a high risk of bias. 24 More patients in the sorafenib-alone group had PVT (35% vs. 15%) and BCLC stage C disease (70% vs. 60%), indicating poorer prognosis in this group. There were 196 treatment-emergent AEs reported in the SIRT plus sorafenib arm and 222 events in the sorafenib-alone arm, of which 21.9% and 21.2%, respectively, were considered to be grade ≥ 3. The most common grade 3 or 4 AEs (hypertension, hand–foot skin reaction and diarrhoea) were reported in a similar number of patients in both treatment arms. Grade 3 or 4 fatigue appeared more common in patients receiving SIRT plus sorafenib (20% vs. 10%). Grade 3 or 4 infection and anorexia appeared more common in patients receiving sorafenib alone (20% vs. 5% and 0% vs. 10%, respectively). Grade 3 or 4 laboratory-related events were more common in patients receiving sorafenib alone (elevated gamma-glutamyltransferase level 45% vs. 30%, elevated aspartate aminotransferase level 15% vs. 0% and elevated alanine aminotransferase level 10% vs. 0%). One patient experienced a grade 3 gastric ulcer that was probably (but not proven to be) related to SIRT microspheres deposition.

Further details of each of these trials are presented in Appendix 6.

Ongoing studies

There are three ongoing studies of SIR-Spheres including patients with HCC: the Austrian Cardiovascular and Interventional Radiological Society of Europe Registry for SIR-Spheres Therapy (CIRT),70 the RESIN tumour registry in the USA71 and the RESIN tumour registry in Taiwan. 72 The CIRT study was completed in January 2020, the RESIN tumour registry study in the USA is due to be completed in August 2022 and the RESIN tumour registry study in Taiwan was due to be completed in December 2019.

There is also an ongoing individual patient data prospective meta-analysis of patients from the SIRveNIB and SARAH trials: VESPRO. 73

Ongoing studies

There is one ongoing RCT of TheraSphere in patients with HCC: STOP-HCC, which has an estimated study completion date of February 2020; final results are not anticipated before at least December 2020. 74

The BTG submission presents 12 additional ongoing or planned studies of TheraSphere.

Ongoing studies

There are three ongoing studies of QuiremSpheres including patients with HCC: HEPAR Primary,75 HORA EST HCC76 and Hope166. 77 All three studies are currently recruiting patients.

Overview

Two studies by Rognoni et al. 90,97 reported on the cost-effectiveness of SIRT in HCC from an Italian heath service perspective. Both studies used the same basic model design and inputs, but investigated different treatment strategies. One study97 compared SIRT with sorafenib in two HCC subpopulations: intermediate (BCLC B) and intermediate-advanced (BCLC C) disease. The other study90 compared SIRT followed by TACE and possibly sorafenib with SIRT followed by sorafenib in patients with intermediate disease (BCLC B).

Both studies presented a probabilistic Markov model consisting of up to five health states: stable disease, progression, posttransplant, death from disease and death from other causes. The post-transplant health state was used only for the comparison of SIRT with sorafenib in patients with intermediate disease. Transition probabilities were drawn from three Italian oncology centres, which were compared using propensity score matching. HRQoL measures were not reported in this cohort; utilities were, therefore, derived from cost-effectiveness analysis registries. Utilities were assumed to be the same across the patient populations. Italy-specific costs were used in the model, and were derived primarily from official local tariffs and reference costs.

For intermediate-stage patients, the estimated incremental cost-effectiveness ratio (ICER) for SIRT compared with sorafenib was €3302 per quality-adjusted life-year (QALY) gained. In advanced-stage patients, SIRT was found to dominate sorafenib. These results appear to be driven primarily by the relatively low costs of the SIRT procedure relative to the acquisition costs of sorafenib, combined with significant clinical benefits of SIRT resulting in additional life-years gained (LYG). In the comparison of SIRT followed by TACE and possibly sorafenib, with SIRT followed by sorafenib, SIRT-TACE-sorafenib was found to dominate SIRT-sorafenib.

Commentary

The two studies appear to be comprehensive and well implemented, accounting for all major sources of costs and benefits, including long-term benefits in patients receiving liver transplant. However, the fitting and selection of parametric functions to survival data were poorly described and explored. Variability in cost-effectiveness estimates was explored using a one-way sensitivity analysis, showing that the results were robust to a wide range of assumptions.

However, the two studies suffered from a number of potential limitations. Foremost among these is the use of non-randomised data to produce estimates of relative effectiveness. Although propensity scoring was used to adjust for baseline imbalances, this process may have affected the results. The comparison between SIRT and sorafenib in the BCLC C subgroup is of particular concern, as a significant survival benefit was predicted for patients receiving SIRT. This is inconsistent with the results of the SARAH19 and SIRveNIB21 trials reported in Chapter 3, Efficacy and safety of SIR-Spheres, which show no such benefit. The HRQoL values used were generally not reflective of the population under consideration, and matched poorly with those used in previous NICE technology appraisals (TAs) in this indication. The study was also limited in its capacity to inform the present appraisal as the costs and resource use evidence reflected an Italian health-care setting, and the choice of comparators does not represent current UK practice.

Overview

The study by Rostambeigi et al. 91,92 (also presented as a conference abstract) sought to assess the cost-effectiveness of SIRT versus conventional TACE in three subgroups (BCLC A, B and C) of patients with HCC from a US Medicare perspective.

The model presented was a patient simulation that followed 750 patients (split evenly between BCLC A, B and C) through a treatment pathway comprising treatment with either SIRT or TACE. The simulation was repeated for each treatment type and patient subgroup over a time horizon of 3 or 5 years. The model structure adopted is not clearly reported, but appeared to allow for disease recurrence, mortality and liver transplant.

Probabilities for each outcome were drawn from the literature for each patient subgroup according to BCLC stage. Exponential curves were used to estimate survival based on reported survival rates, with a 10% increase in mortality for 1 month following recurrence of HCC and re-treatment. Transplant rates of 29%, 16% and 5% were applied for patients in BCLC stages A, B and C, respectively, although is it unclear how this affected model outcomes. The model assumed disease ‘recurrence’ rates of 40%, 60% and 80% every 10 months for SIRT patients, whereas TACE patients had a recurrence rate of 60%, and could receive 4 to 10 procedures. An assumed probability of 0.5 was used for SIRT re-treatment at the beginning of every 10-month treatment interval, and patients were assumed to receive a maximum of two or three SIRT treatments depending on the scenario. Costs applied in the model were obtained from Medicare reimbursement costs; HRQoL was not considered.

The ICERs presented were estimated using an unconventional approach, calculated by dividing the incremental mean cost per month of survival (i.e. total costs divided by OS in months) by the overall incremental survival in months. The authors did not account for dominance in their calculations, presenting a number of negative ICERs without sufficient interpretation of their different meanings. ICERs in which SIRT was less costly and less effective, less costly and more effective, and more costly but less effective than TACE were presented without further distinction.

In the main analysis in which each procedure could be repeated every 10 months for up to 5 years, the AG calculated SIRT to increase mean survival by 3.80 months in BCLC C patients at a reduced cost. In the scenario in which procedures are repeated every 6 months for up to 3 years, SIRT was more effective (2.90 months incremental survival), with reduced costs compared with TACE in BCLC C patients. In all other patient groups and treatment regimens, SIRT was dominated by TACE.

Commentary

The limited reporting of the model structure and assumptions adopted prevents a detailed critique or discussion of the appropriateness of the model to estimate the relative costs and benefits of SIRT and sorafenib. A number of key structural assumptions appear to have been made arbitrarily, and poor reporting of model inputs limits the generalisability of this study to other settings. As the resource use and costs are specific to the USA, they are unlikely to be relevant to an NHS setting. The choice of comparators and outcome measures (e.g. LYG) further limits comparison with UK practice.

Treatment costs of transarterial chemoembolisation and drug-eluting bead transarterial chemoembolisation

Sirtex provided three alternative scenarios for the cost of TACE and DEB-TACE. In one scenario, these costs were based on those estimated by Fateen et al. ,105 a single-centre retrospective database study from the UK. This study collected cost data for 101 procedures in 43 patients between 2006 and 2012 at a centre in Nottingham, UK. In this study, 25% of patients received DEB-TACE and the remaining 75% of patients received TACE. Costs reported in Fateen et al. 105 were for the 2012 cost year: these were inflated to 2018 costs. 106

A second scenario used unit costs from National Schedule of Reference Costs 2017–2018107 for hospitalisation, applied to resource use as estimated in the Fateen et al. 105 study. The mean cost per day of hospitalisation was estimated as £1757 (from Elective Inpatient, Percutaneous, Chemoembolisation or Radioembolisation, of Lesion of Liver, YR57Z), and was assumed to include the cost of delivering TACE.

A third scenario incorporated the results of the resource use survey commissioned by Sirtex, which were used to estimate the number of TACE and DEB-TACE procedures received by each patient, and the proportion of patients receiving DEB-TACE and TACE. The resource use survey was completed by five medical professionals from UK hospitals, comprising two oncologists, one hepatologist and two specialist nurses. This scenario was presented to reflect that resource use might have changed since the time that the Fateen et al. 105 study was undertaken. The survey estimated that a greater proportion of CTT patients receive DEB-TACE in the survey than in the earlier-conducted Fateen et al. 105 study (63% vs. 25%), and that, on average, there are fewer procedures undertaken for a given TACE patient (2.5 vs. 3.03) but a greater number of DEB-TACE procedures (2.83 vs. 1.43).

The costs of providing CTT, estimated as a weighted average of DEB-TACE and TACE costs, ranged from £8792.59 in the scenario based on the Fateen et al. 105 study (scenario 1), to £13,702.37 in the scenario incorporating the results of the resource use survey for the number of TACE and DEB-TACE procedures (scenario 3). A full breakdown of costs is provided in Appendix 15, Table 48.

Treatment costs of selective internal radiation therapy

Procedure costs relating to the administration of SIR-Spheres were assumed to comprise the device costs, the cost of work-up and the SIRT administration procedure (see Appendix 15, Table 49, for a detailed breakdown).

The acquisition cost for a single administration of SIR-Spheres and TheraSphere was assumed to be £8000.

Sirtex provided a range of scenarios to explore work-up and procedure costs, using alternative sources and assumptions to provide a range of plausible costs. Work-up costs were based on the number of work-ups and the total length of hospital stay for a work-up. SIRT procedure costs were based on the number of procedures and the total length of inpatient stay. If the hospital stay was < 1 day, the cost of an outpatient visit was instead applied.

Adverse event costs

The unit costs applied in the CTT-eligible model are reproduced in Appendix 15, Table 50. Sirtex derived the unit costs for treating each event from previous NICE TAs, and AE rates were obtained from Salem et al. ,25 a Phase II RCT that compared TheraSphere with TACE in a population of early-stage HCC patients with intent to transplant. Rates of AEs for SIR-Spheres were assumed to be equivalent to those for TheraSphere. This study estimated a higher burden of AEs in CTT patients, in particular neutropenia and elevated aspartate aminotransferase. Consequently, a higher cost was applied in the model (£346 for CTT vs. £109 for TheraSphere).

Cost-minimisation analysis

The AG considered the presentation of a CMA for this population to be inappropriate and potentially misleading. Such an analysis is appropriate only if there is compelling and unambiguous evidence for equivalent efficacy between interventions. When a CMA is considered by NICE in other appraisals, it is typically accompanied by an extensive and conclusive assessment of equivalence between treatment arms. 109–111 Clinical equivalence is a dynamic concept and any demonstration of clinical equivalence should be sustained over time. Therefore, it is important to assess whether or not the two therapies are equivalent not just in response rate, but in terms of if PFS and OS are also similar.

Results of the AG systematic review found no high-quality evidence in this population. As discussed in Chapter 3, Clinical effectiveness results, the RCTs directly comparing SIR-Spheres with TACE and DEB-TACE were very small and of poor quality, and appeared to favour the chemoembolisation procedure over SIRT in terms of survival outcomes. Although one RCT comparing TheraSphere with TACE reported longer TTP, a higher proportion of patients undergoing transplant and a small but non-significant OS benefit in the TheraSphere arm, this study enrolled a small number of patients and was rated as having a high risk of bias. 22

Therefore, although the AG acknowledges the cited limitation in the effectiveness evidence for this population, and agrees that the development of a cost–utility model is inappropriate, the AG does not consider the identified evidence sufficient to make the strong assumption of equivalence between CTT and SIRT. Furthermore, a focus on treatment costs excludes possible important outcomes regarding people who are downstaged after treatment and become eligible to receive curative therapy, or who receive subsequent therapy after progression of disease.

Cost of treatment with conventional transarterial therapy

The cost analysis of CTT highlighted significant uncertainties in the number of CTT treatments that are typically given, and the impact on the total costs. The applicability of the available sources was limited, and included the only single UK centre collecting data between 2006 and 2012,105 and a survey of five UK-based clinicians. These two sources were used to provide a range of the number of treatments that CTT patients might receive in practice. For TACE, the estimated range was narrow and estimated at between 2.5 and 3.03 treatments. A much wider range was, however, estimated for DEB-TACE (1.43 to 2.83). To consider the plausibility of the presented estimates, the AG searched for alternative estimates of the number of TACE and DEB-TACE procedures. The AG identified two alternative sources of representative data: a UK-based multicentre trial of DEB-TACE enrolling patients between 2010 and 2015 found that a mean of 2.18 DEB-TACE treatments were given,22 and clinicians at a centre in the UK with experience in delivering TACE reported that patients (up to 2010) received a mean of 2.56 treatments with TACE (Dr Jai Patel, Leeds Teaching Hospitals NHS Trust, 2019, personal communication). These estimates both fall within the ranges presented by Sirtex.

Number of selective internal radiation therapy procedures

Sirtex explored the cost impact from using a range of sources to estimate the number of procedures with SIR-Spheres and with TheraSphere. Patients receiving treatment with SIRT typically receive multiple procedures on the basis of their tumour burden (i.e. bilobar involvement requiring sequential treatment visits), with patients not typically re-treated with SIRT on disease progression. Therefore, the number of procedures required would not be expected to differ between treatment arms, and the range of total treatment costs for SIR-Spheres and TheraSphere estimated by this analysis might be expected to be more similar.

Overall survival

Overall survival for patients downstaged and in receipt of palliative care was modelled separately, with the proportion of patients downstaged based on observed values in the low tumour burden/ALBI 1 subgroup of the SARAH trial. 19

Overall survival for patients who are not downstaged to curative therapies in the economic model was based on observed survival in the SARAH trial,19 using data on the low tumour burden/ALBI 1 subgroup of patients, including 37 SIRT patients and 48 sorafenib patients.

Before fitting parametric functions to the available KM data, diagnostic plots were used to assess the plausibility of assumption of proportional hazards. The plots revealed some evidence to suggest that the proportional hazards assumption may not hold, as in some instances the lines were not parallel and indeed crossed in some cases. 102 The Schoenfeld residuals, however, suggest no significant deviation from the proportion hazards assumption. Given this uncertainty, Sirtex opted to fit separate parametric functions to the KM data.

The following parametric survival models were fitted to the observed KM data: Weibull, log-normal, log-logistic, exponential and gamma functions. Assessment of the most appropriate parametric extrapolation was made with reference to statistical goodness of fit, visual fit to the observed data and assumptions made in previous TAs. 11–13 Assessment of statistical fit (see Sirtex company submission,102 appendix F) revealed a similar statistical fit for the majority of curves, with the exponential curve observed to have the highest statistical fit. In assessing visual fit, Sirtex noted that the generalised gamma, Weibull and Gompertz curves crossed, which is not seen in the KM curves until the last few patients, whereas the log-normal and log-logistic curves did not cross. Sirtex further noted that in previous TAs of sorafenib (TA47419) and lenvatinib (TA55112), the log-logistic and log-normal curves were considered the most appropriate, and in the analysis of the SARAH19 ITT population the log-normal distribution fitted best, in terms of both goodness-of-fit statistical criteria and visual inspection. On these grounds, Sirtex therefore selected the log-normal function for its base-case analysis. Assessment of uncertainty in curve selection was also partially explored in two scenario analyses considering the log-logistic and Weibull distributions.

Overall survival outcomes for patients downstaged to curative therapy were not drawn from the SARAH trial,19 as OS data were censored on receipt of curative therapy. Survival outcomes for these patients were, therefore, based on a US cohort study,112 which reported the outcomes for patients who did and did not receive curative therapy. The survival HR for downstaged patients was 0.29 (95% CI 0.18 to 0.47). To model survival in the downstaged patients, this HR was applied to the treatment-specific survival curves for SIR-Spheres and sorafenib patients. Importantly, because this HR was applied to the individual survival curves for SIR-Spheres and sorafenib, the model implies differential OS following receipt of curative therapies depending on the initial treatment received.

Progression-free survival

Progression-free survival was defined as the time from the closest date of radiological examination before the first administration of the study treatment to disease progression (per investigator assessment), or death from any cause. Because progression events were observed across patients who were and were not downstaged to receive curative therapy, a common PFS curve was assumed for all patients irrespective of whether or not they received subsequent curative therapy. Sirtex’s base-case analysis drew PFS data from the low tumour burden/ALBI 1 subgroup of the SARAH trial. 19

Assessment of the proportional hazards suggested a degree of uncertainty in whether or not this assumption holds. Assessment of statistical fit based on Akaike information criterion (AIC) and Bayesian information criterion (BIC) of the jointly fitted data found that the (assuming proportional hazards) log-logistic and log-normal, as well as the independently fitted (no proportional hazards) log-normal, distributions had the best statistical fit. Aligning with assumptions made for OS, Sirtex’s base-case analysis used independently fitted log-normal distributions. Uncertainty in curve selection was partially explored in a scenario analysis in which the log-logistic and Weibull distribution were used.

Health-related quality of life

The primary source of utility data used by Sirtex was the SARAH trial,19 which measured HRQoL using the EORTC QLQ-C30 questionnaire. There were a significant number of missing responses over the course of the study, ranging from 19% at baseline to 56.8% at 18 months, with an overall rate of missing data of 38.5%. To calculate health state utilities from this data set, the mapping algorithm by Longworth et al. 113 was used to generate EQ-5D scores adjusted to reflect UK population weights. Sirtex did not consider the SARAH trial19 to show evidence of an independent treatment effect on utility, and there was no significant difference between the HRQoL of those treated with SIR-Spheres and those treated with sorafenib. The company submission,102 however, also notes a statistically significant difference in reported global health scores between treatment arms, and applies treatment-specific utility values based on the subgroup of patients with a tumour burden of ≤ 25% and ALBI 1. The values used in the base-case model are reported in Appendix 15, Table 51.

Selective internal radiation therapy procedure costs

Procedure costs relating to the administration of SIR-Spheres were assumed to comprise the device costs and cost of the work-up and treatment procedures. All patients in the SIRT arm of the model were assumed to undergo at least one work-up procedure, with 5% of patients also assumed to undergo a second work-up based on clinical opinion. To account for the fact that not all patients will go on to receive SIRT (e.g. owing to excess shunting), only a proportion of patients were assumed to receive SIRT. Sirtex’s base case used the low tumour burden/ALBI 1 subgroup of the SARAH trial19 to derive this figure. The model also permitted SIRT patients to be re-treated with SIRT. Sirtex did not consider the average number of SIRT treatments in the SARAH trial19 to represent likely UK practice, as the SARAH trial19 mandated separate administrations where bilobar disease was present. Sirtex instead used data from the CIRT114 (Belgium, France, Germany, Italy, Spain and Switzerland) as well as the ENRY study showing that patients with bilobar disease typically receive a single administration of SIRT with both lobes treated simultaneously. 69 The number of SIRT administrations was, therefore, based broadly on the CIRT, with 1.20 treatments assumed per patient. Uncertainty in the number of SIRT administrations was also explored in scenario analyses based on the SARAH trial,19 the SIRveNIB trial,21 the ENRY study69 and The Christie NHS Foundation Trust data (personal communication).

Costs relating to the work-up and SIRT procedures were based on National Schedule of Reference Costs 2017–2018,107 with the cost of SIR-Spheres assumed to be £8000 per administration. Appendix 15, Table 52, summarises the assumptions and costs of the SIRT procedure.

Drug acquisition costs: systemic therapies

Drug acquisition costs for sorafenib and lenvatinib were taken from the British National Formulary (BNF). 115 Dosing of sorafenib was based on the SARAH trial,19 assuming that 24% of patients received an 800-mg dose and 76% received a 600-mg dose. In scenarios in which lenvatinib was included as a comparator, dosing was based on TA551,12 with 65% assumed to receive an 8-mg dose and 35% assumed to receive a 12-mg dose. 12 Duration of sorafenib therapy was based on the time-to-discontinuation curve from the SARAH trial,19 which was extrapolated using a log-normal function. Duration of lenvatinib therapy was estimated by applying a HR to the sorafenib time-to-discontinuation curve taken from TA551. 12

Subsequent treatments

Modelled subsequent treatments without curative intent were based on expert elicitation, as the subsequent treatments received in the SARAH trial19 were not considered reflective of NHS practice. Drug costs were taken from the electronic market information tool (eMIT)116 and the BNF. 115

For patients downstaged to receive curative therapies, the modelled therapies were based on those received in the ITT population of the SARAH trial,19 consisting of resection, liver transplantation and tumour ablation. The proportion receiving each type of therapy is summarised in Appendix 15, Table 53. Costs of resection were based on NICE TA474,11 and costs of ablation and liver transplantation were based on National Schedule of Reference Costs 2017–2018. 107

Health state costs

Resource use estimates were based on a survey of clinical experts, and included medical staff contacts [e.g. general practitioner (GP) appointments], diagnostic procedures, inpatient care and Personal Social Services (PSS) contacts. Unit costs were derived from National Schedule of Reference Costs 2017–2018. 107 Total costs by health state are reported in Appendix 15, Table 54.

Adverse event costs

The costs of grade 3 or 4 TRAEs experienced by ≥ 5% of the population were modelled, with rates drawn from the SARAH19 and REFLECT81 trials. Costs for each AE were sourced from previous TAs and inflated to the 2018 cost year as appropriate. See Appendix 15, Table 55, for a summary of included AE costs.

Model results

The headline results presented in the Sirtex company submission102 are based on the deterministic version of the model. Uncertainty surrounding model parameters was explored using deterministic sensitivity analysis (DSA) and PSA. The probabilistic results were estimated from 1000 Monte Carlo samples. Uncertainty was represented using tornado diagrams, cost-effectiveness planes and cost-effectiveness acceptability curves (CEACs).

Table 20 presents the base-case estimates of cost-effectiveness using the list price for sorafenib. Based on the probabilistic version of the company’s model, SIR-Spheres are expected to generate an additional 0.682 QALYs at an incremental cost of –£1979 compared with sorafenib; SIR-Spheres were, therefore, estimated to be dominant, producing greater health benefits at lower overall cost. The deterministic version of the model produces similar results, with SIR-Spheres estimated to dominate sorafenib.

Figure 9 presents the results of the company’s DSA. The most influential parameters (of those assessed by the company) relate to predicted OS (SIR-Spheres and sorafenib) and the proportion of patients downstaged to receive curative therapy. Additional scenario analyses presented by the company showed that the estimated ICER was generally robust to a range of alternative assumptions, including alternative extrapolations of survival data. However, this analysis also showed that estimated ICERs increased very significantly when the source of effectiveness estimates was changed from the low tumour burden/ALBI 1 subgroup to the ITT or per-protocol population from the SARAH trial,19 which yielded ICERs of £58,763 and £680,276 per QALY gained, respectively.

FIGURE 9.

Sirtex DSA: tornado diagram. Reproduced with permission from Sirtex Medical Ltd. 102

Reproduced with permission from Sirtex Medical Ltd. 102

Relevance of modelled population

The company’s health economic analysis is limited to a subpopulation of patients with a tumour burden of ≤ 25% and with preserved liver function (ALBI 1). The company cited clinical opinion and published literature in its justification for focusing on this group, stating that the ITT and per-protocol population recruited to the SARAH trial19 was unreflective of that eligible in the UK, while also highlighting that the trial included patients with high tumour volume, PVT and poor liver function. The company also outlined that this subpopulation increased the probability of receiving SIRT and the probability of going on to access curative therapy, citing figures from the SARAH trial. 19

Consultation with the AG’s clinical experts confirmed that this subgroup could be identified prospectively and treated with SIRT. However, they also noted that ALBIs are not routinely used to assess liver function in UK practice, and that this definition did not represent a widely accepted clinically distinct subgroup of patients.

The AG is further concerned that the selection of this subgroup is based on a post hoc analysis of a relatively small subgroup of the SARAH trial,19 representing < 20% of the total trial population. Comparison of the results for this subgroup on key outcomes, such as PFS and OS, revealed no statistically significant differences between this group and the remaining population. Furthermore, the randomisation procedure for the SARAH trial19 did not stratify by these baseline characteristics, increasing the risk of baseline imbalances. This can be observed in the sample size of this group between treatment arms, with 37 patients in the SIRT arm and 48 patients in the sorafenib arm. A further consequence of using this subgroup is that potentially relevant data from the SIRveNIB trial21 cannot be used, as data on this subgroup were not available to the company. This is important for two reasons: (1) it reduces the available sample size with consequences for precision and (2) it does not allow for a confirmatory analysis of the PFS and OS benefits observed in this subgroup.

The AG is, therefore, concerned that the purported treatment effects in this subgroup are potentially an artefact of imbalances in characteristics between treatment arms. Available data do not allow further analysis to establish the validity of the observed PFS and OS gains in this subgroup.

Model structure and clinical plausibility of downstaging

The company’s model allows a proportion of patients to move on to receive curative therapy. This is a significant driver of the model results, as 66% of incremental QALYs are generated by patients who received curative therapies.

The SARAH trial19 was used to support the downstaging paradigm used in the model, in which a small number of patients went on to receive curative therapy. The plausibility of downstaging at such high rates in UK practice is unclear. The AG was advised that downstaging of patients with advanced HCC to transplant and other curative options is rare in UK clinical practice, with very few if any of these patients receiving curative therapies. It is also notable that the SIRveNIB trial,21 which recruited a similar population, makes no mention of any patients going on to receive curative therapy. Similarly, none of the previous TAs that assessed systemic cancer treatments for advanced HCC modelled the possibility of curative therapies. The AG is, therefore, concerned that the very sizable benefits resulting from curative therapy would not be realised in practice, and that the rarity of downstaging means that any resulting incremental benefits are subject to very considerable uncertainty.

Modelling of overall survival

The company fit independent parametric survival functions to the observed data from the SARAH trial. 19 This method makes fewer assumptions than a treatment–covariate-based approach, and is in line with NICE DSU guidance on survival analysis. 117 However, the AG does not accept the company’s rationale for selecting the log-normal curve, which was based primarily on visual fit and its use in previous HCC appraisals. The AG notes that the log-normal is the most optimistic of all the fitted parametric curves, and has among the worst statistical fit. The log-normal also has a much longer tail, and, in the AG’s view, fits poorly to the tail of the observed data for the SIR-Spheres arm of the SARAH trial. 19 Clinical advice to the AG indicated a preference for the Weibull function, which predicts substantially shorter survival gains and also has better statistical fit.

In addition to the above, the AG is concerned that the parametric functions were fitted to the observed data that had not been censored to exclude those patients downstaged to receive curative therapy. In the economic model, the outcomes for these patients are modelling independently and, therefore, using the uncensored data means that the OS benefits experienced by these patients are double-counted. The impact of this double-counting is significant, and leads to a substantial overestimation of survival gain. For example, based on a log-normal extrapolation (used in the Sirtex base case) and using the uncensored data, estimated OS gain on SIR-Spheres is 8.27 months. Using the log-normal function on the same data censored for downstaging results in a much reduced predicted OS gain of 1.55 months.

Further to the above issues regarding the plausibility of downstaging, the AG has concerns around the methods used to model the OS benefits associated with curative therapy. Postcurative OS is modelled by using the HR from the Kanwal et al. 112 cohort study to the OS curve for each treatment. This HR is assumed to reflect the improvement in survival outcomes post curative therapy. The application of this HR is treatment specific (i.e. is applied to the SIR-Spheres OS curve for SIR-Spheres patients and to the sorafenib OS curve for sorafenib patients). This implies that OS postcurative therapy will differ depending on the initial treatment received, and thus favours SIR-Spheres. Expert advice received by the AG, however, considers this implausible and that outcomes will be the same post curative therapy regardless of previous therapy received.

Furthermore, the application of a HR to the log-normal curve is inappropriate, as the log-normal function is an accelerated failure time model and does not make assumptions about proportional hazard assumptions. Consequently, survival times are considerably over estimated. The AG also questions the appropriateness of the HR of 0.29 used by the company, noting that this figure was not based on the primary analysis presented in the cited study, but on a scenario analysis in which classification of patients was based on both BCLC stage and ECOG performance status.

Modelling of progression-free survival

The company’s approach to modelling PFS was similar to that of modelling OS, with independent parametric survival functions fitted to the observed data.

The AG is satisfied that the company’s approach of using independent curves was appropriate given the presented evidence to support the non-proportionality of hazards. The AG, however, questions the appropriateness of fitting parametric functions to PFS data at all, given that the available KM data are all but complete; no patients remain at risk in the sorafenib arm and only one remained in the SIRT arm. The company could, therefore, have used the observed data directly, avoiding any uncertainty in the choice of parametric function.

The AG is also concerned that the modelled data were not censored for downstaging events and, therefore, double-count patients who were downstaged to receive curative treatment. As with OS, this results in PFS gains being overestimated, although to a lesser degree than OS. Mean PFS gain assuming a log-normal function was 3.7 months using the uncensored data and 2.35 months using the censored data.

Concerns regarding costs of selective internal radiation therapy

It is assumed in the Sirtex model that patients with bilobar tumours receive SIRT in both liver lobes during the same treatment session. This is in contrast with how patients were treated in the SARAH trial,19 which mandated that patients receive separate treatments with a delay between the first and the second administration. Sequential treatment is implemented to mitigate the risk of REILD, which is more likely to occur if both lobes are treated simultaneously. The company put forward evidence from the European CIRT, and suggested that (confidential information has been removed).

The impact of this assumption is to reduce the costs of providing SIR-Spheres, as sequential treatment involves additional administration and acquisition costs. However, clinical advisors to the AG disagree with the assertion that simultaneous treatment would be implemented in the UK, and contend that in UK practice it is likely that sequential treatment would be used as per the SARAH trial. 19 Furthermore, the AG notes that, although the company adjusts costs to account for the use of simultaneous treatment, no corresponding adjustment is made to health outcomes to account for the increased risks associated with simultaneous treatment.

Failed work-up procedures

In the Sirtex model, a proportion of patients are assumed to fail the work-up procedure and are thus ineligible to receive SIR-Spheres. The proportion of patients receiving work-up who do not go on to receive SIRT was drawn from the low tumour burden/ALBI 1 subgroup of the SARAH trial,19 which was substantially lower than for the population as a whole (8.1% vs. 18.6%). The AG is concerned about the appropriateness of this figure, given the post hoc nature of the analysis. The primary reason patients become ineligible for SIRT following work-up is a high rate of shunting of radioactive material to the lungs. Although this may be plausibly linked to tumour volume and liver status, any such association has not been demonstrated, and it is not clear that the proportion of patients who experience excessive lung shunt will vary substantially between patient groups.

Furthermore, the company’s model assumes that patients who fail work-up will move to the sorafenib arm of the model. The AG considers this inappropriate as only 62% of patients in the SARAH trial19 who failed work-up subsequently received sorafenib. The outcomes of patients in the SARAH trial19 who received work-up but no SIRT were inferior to those who successfully received SIR-Spheres or were randomised to the sorafenib arm. Assuming that patients who fail work-up receive sorafenib outcomes is therefore likely to overestimate the PFS and OS for those allocated to receive SIR-Spheres.

Subsequent therapy costs

The company noted in its submission that the subsequent treatments received by patients in the SARAH trial19 included a number of therapies (e.g. capecitabine and doxorubicin) not used in UK practice. The treatments received following primary therapy in the model were, therefore, based on a survey of 12 clinicians instead.

The AG considers the proportions of patients receiving subsequent therapies in the model to be subject to substantial uncertainty, and notes that these differ substantially from those reported in the SARAH trial. 19 The proportion of patients assumed to receive sorafenib following SIR-Spheres is higher than that observed in SARAH,19 as is the proportion of patients receiving further treatments post sorafenib. The AG also notes that post-sorafenib treatment is based on the ITT population of the SARAH trial19 and, therefore, does not reflect the modelled low tumour burden/ALBI 1 subgroup. Given that the low tumour burden/ALBI 1 subgroup represents a particularly healthy population, it may be anticipated that a much higher proportion of these patients would go on to receive subsequent systemic therapies. As no figures on subsequent therapy in the low tumour burden/ALBI 1 subgroup are reported, this cannot be verified.

Duration of subsequent sorafenib and lenvatinib therapy was drawn from the REFLECT trial,12 and subsequent regorafenib was based on the RESORCE trial. 101 The approach taken to define time on treatment was inconsistent, as median values were used for sorafenib and lenvatinib, whereas a mean value was used for regorafenib. The AG considers mean values more appropriate than the medians used by the company, as the aim of the model is to calculate the mean costs of subsequent therapy. The AG is also concerned that the REFLECT trial12 considers the use of sorafenib and lenvatinib in a first-line setting, particularly as this implies that patients receiving sorafenib as a subsequent therapy will receive treatment for much longer than those who received it as a first-line therapy. The AG, therefore, considers that these values are likely to overestimate time on treatment, and that it may be better to base duration of subsequent therapy on the RESORCE trial,101 which considers systemic therapy use in a second-line setting.

Omission of palliative care costs

The Evidence Review Group notes that the company model does not include end-of-life costs to account for palliation at the end of life. However, the impact of this omission is small, as fewer than 1% of patients remain alive at the end of the modelled time horizon, meaning that nearly all modelled patients incur this cost.

Downstaging outcomes

In this model, it was assumed that the impact of treatment with SIRT compared with CTT was limited to differences in the likelihood of patients being downstaged and becoming eligible for curative therapy.

Non-mortality outcomes for the ‘watch and wait’ health state were estimated from a single-centre, non-randomised comparison of TACE and TheraSphere patients. 118 The study was undertaken in a population of unresectable HCC patients who did not meet the Milan criteria9 at presentation, specifically including patients who were of T3 United Network for Organ Sharing (UNOS) status. This is defined as patients with either a single nodule of > 5.0 cm or two or three nodules, at least one of which is > 3.0 cm in size,119 and downstaging was defined as a decrease in the maximal tumour dimension to 3.0 cm.

The probability of remaining in the watch and wait health state for all therapies was estimated by the company using the median time to downstaging in the TheraSphere arm of the Lewandowski et al. 118 study. The company assumed that the median time to downstaging represented the median time to ‘prognosis’ (i.e. either to downstaging or to pharmaceutical management). The median time to downstaging in the study for TheraSphere patients was 3.1 months; the median time to downstaging in the TACE arm of the study had not been reached. The company converted the median time of 3.1 months to a per-cycle probability of leaving the watch and wait health state of 18.6%, resulting in a per-cycle probability of remaining in this health state of 81.4%.

Of the proportion who leave the watch and wait health state in each cycle, the company used the probability of downstaging from the Lewandowski et al. 118 study to estimate the transition of patients to the pre-transplant state. The remaining living patients entered the pharmacological management health state. The study reported a probability of downstaging from TheraSphere treatment of 58% (25/43), compared with 31% (11/35) downstaged from TACE.

The efficacy of SIR-Spheres and QuiremSpheres was assumed to be equal to that of TheraSphere, and the efficacy of DEB-TACE and TAE was assumed to be equal to that of TACE.

Owing to a lack of data specific to this outcome, the probability of death in each model cycle for the ‘watch and wait’ health state was assumed to be equivalent to that of patients on the wait list, which was estimated from a cohort of NHS patients awaiting liver transplant (see Transplant wait list outcomes). The mortality rate was assumed to be equal between all treatment arms. The greater predicted benefits of SIRT in this model are, therefore, entirely attributable to a greater proportion of patients being successfully downstaged.

Appendix 15, Tables 56 and 57, summarise the transition probability values and mortality rates, respectively, used in the model.

Transplant wait list outcomes

The probability of successfully receiving a transplant once on the wait list was calculated by the company using the median wait time of 130 days for a liver transplant in the UK. 120 This data set is based on a cohort of 2706 NHS patients who were registered for a liver transplant between April 2013 and March 2016, and is not specific to an indication of HCC. This was converted to a per-cycle probability of 13.9%. The probability of transplantation was not conditional on initial treatment.

Patients could transition from the pre-transplant state to pharmacological management, in the case that a patient becomes ineligible for transplant while on the wait list. The probability of this happening was informed by clinical advice to the company, with 16 cases of patients leaving the wait because of disease progression for every 103 transplants (National Audit for Liver Transplant, incomplete source provided by the company).

Mortality in the pre-transplant wait list health state was estimated from a figure quoted in an NHS service specification for Liver Transplantation Service in Adults,121 in which ‘up to 18% of patients die while on the liver transplant waiting list’ (reproduced with permission; contains public sector information licensed under the Open Government Licence v3.0), and converted to a per-cycle mortality rate using the median time to transplant of 130 days.

Pharmacological management outcomes

Patients entering the pharmacological management health state are assumed to remain there until death. The mortality rate applied was based on the median OS of BSC patients reported in the NICE sorafenib submission (34.4 weeks). 11 Per-cycle mortality was estimated assuming that OS followed an exponential distribution; the applied per-cycle mortality rate was 7.7%. This rate was applied to patients in this health state regardless of their initial treatment.

Post-transplant outcomes

Mortality in the three cycles (12 weeks) following transplant was estimated using data from a study of early-stage HCC patients, Bellavance et al. ,122 which reported a 30-day mortality probability of 1.5%.

The post-transplant mortality rate beyond these three cycles was assumed to be lower, and was estimated from NHS 5-year survival rates following transplantation120 of liver patients who had a transplant between 2010 and 2012, which was estimated at 81%. These data reflect a general liver transplant population and are not specific to those who have HCC. Furthermore, for the patients in the population who did have HCC, they are also not specific to patients who had been downstaged after having previously been ineligible for transplant before active treatment for HCC. The company justified the assumption that the mortality rates for a downstaged population can be assumed to be equivalent to a population who were not originally downstaged, on the basis of a systematic review by Gordon-Weeks et al. 123

Adverse events

For TheraSphere and SIR-Spheres, data on grade 3 and 4 TRAEs were sourced from a systematic review of AEs. 79 Event rates for QuiremSpheres were assumed to be the same as for SIR-Spheres. Rates of TRAEs for TACE and DEB-TACE were sourced from a RCT of DEB-TACE versus TACE in HCC. 58 The company’s model included severe TRAEs that occurred in > 5% of patients in at least one arm.

Total TRAE utility decrements and treatment costs were applied in the first model cycle. The estimates of utility decrements were based on the assumption that grade 3 and 4 AEs were associated with a utility decrement of 0.012, which was multiplied by AE rates reported for each event. The total TRAE disutility for TheraSphere, SIR-Spheres, QuiremSpheres and TACE was estimated as –0.002, with –0.009 for TAE and 0.000 for DEB-TACE. Total TRAE costs ranged from £5.59 for DEB-TACE, to £111.33 for SIR-Spheres, and £384.15 for sorafenib. Further details of TRAE rates and associated costs are provided in Appendix 15, Tables 58 and 61, respectively.

Health-related quality of life

BTG drew on a variety of external sources for the utility values in its economic model (see Appendix 15, Table 59). Utility values for all health states with the exception of the post-transplant tunnel states were the same as the pre-progression values used in the TA551124 submission for lenvatinib (equal to 0.75), which were estimated from EQ-5D data collected from patients in the REFLECT trial. 81 The utility applied to the ‘pharmacological management’ state is taken to be an average of the pre-progression and post-progression health state values, as BTG states that this population comprises patients in both progression states equally. Post-transplant utilities were derived from a study by Lim et al. ,125 which used an average of literature-derived utilities equal to 0.69. A scenario analysis was carried out using significantly lower pre- and post-liver transplant utilities from Ratcliffe et al. ;126 however, these values were taken from a primarily non-HCC population.

Utilities were adjusted according to age and gender norms reported in Kind et al. ;127 however, this adjustment was applied incorrectly, which resulted in patients experiencing a much lower HRQoL than reported in the cited sources. When this was highlighted to the company, it stated that this was intentional, and considered the use of lower utility values appropriate and consistent with methods reported in Kind et al. 127

Costs of selective internal radiation therapy treatment

Procedure costs relating to the administration of SIRT were assumed to comprise microsphere (SIRT) acquisition costs, the cost of the work-up and procedure costs relating to the administration of SIRT. The mean number of SIRT treatments per patient was informed by an elicitation exercise undertaken by BTG. Each patient was estimated as having an average of 1.2 SIRT treatments, with one work-up per patient. Only patients who are eligible for SIRT enter the model, and so the costs of work-ups that did not result in treatment with SIRT were not included.

The work-up procedure costs were based on a microcosting from The Christie NHS Foundation Trust, Manchester, and were estimated as being £467.91. These costs included the time of the personnel involved with the work-up (a technician, clinical scientist and radiologist) and a macroaggregated albumin (MAA) body SPECT. The AG requested additional details of this microcosting; however, little further granularity was provided. In addition, BTG identified further relevant cost items in the work-up procedure, which increased the cost to £860.32 per work-up. The company assumed that the resources required for the work-up associated with TheraSphere, SIR-Spheres and QuiremSpheres would be the same.

Costs relating to the administration of the SIRT work-up and the SIRT procedure were based on National Schedule of Reference Costs 2017–2018,107 and the cost of each SIRT was assumed to be £8000 per procedure. Further details are provided in Appendix 15, where Table 60 summarises the assumptions and costs of the SIRT work-up procedure and Table 62 summarises the associated unit costs.

Treatment costs of conventional transarterial therapy

Each patient in the TACE and TAE arms was assumed to have three initial treatments in their respective arms, and patients in the DEB-TACE arm had an average of 1.5 initial treatments. The unit cost and the frequency of their use was informed by clinician input.

The cost of administration involved in each CTT was assumed to be captured in the Healthcare Resource Group (HRG) code for the embolisation procedure (£2790, National Schedule of Reference Costs 2017–2018,107 HRG code YR57Z).

Second-line treatment

After patients move into the pharmacological management health state, they were assumed to receive sorafenib (33% of patients) or BSC (67% of patients). Patients remain in this state until death. The unit cost of sorafenib was obtained from the BNF,115 with the total per-cycle cost estimated assuming a posology of 400 mg twice daily. It was assumed that sorafenib would not be associated with administration costs and that patients would orally self-administer this treatment. It was unclear whether or not the costs of treating AEs associated with sorafenib treatment were captured within the model. Costs associated with BSC were assumed to be captured within the health state resource use.

Health state resource use

Owing to an absence of evidence from published literature for resource use for the CTT-eligible health states, expert opinion was sought from The Christie NHS Foundation Trust (see Appendix 15, Table 63, for a summary of health state costs). These consisted of the following:

• physician visits (oncologist, hepatologist, Macmillan nurse, gastroenterologist, radiologist, clinical nurse specialist and palliative care physician)

• laboratory tests [alpha-fetoprotein test, liver function test, international normalised ratio (INR), complete blood count, biochemistry and endoscopy]

• radiological tests [computerised tomography (CT) scan, MRI scan and ultrasound scan]

• hospitalisation

• hospital follow-ups (specialist, GP and nurse)

• transplant aftercare (immunosuppressants).

Unit costs for each of these items, plus the cost of a transplant procedure, were obtained from national sources. 106,107

The AG requested additional details of how these resource use estimates were obtained. BTG clarified that resource use estimates were provided by a single clinical expert whose role is consultant interventional radiologist at a centre in the UK that uses SIRT. Opinion was elicited via an unstructured telephone conversation, and the estimates were given verbally and were entered directly into the model; no transcripts of this conversation were collected. Therefore, the AG cannot verify the estimation of the resource use inputs.

Additional one-off costs were applied at the point of progression, relating to laboratory and radiological tests (estimated as £95.32 in total) and were obtained from TA555. 13

Palliative care costs

The company’s model also included a cost of £8191 to account for costs of palliation at the end of life, which was applied on death. This was derived from a joint Nuffield Trust and Marie Curie report into end-of-life cancer care and inflated to 2017/18 prices. 128

Base-case results

Results of the base-case analysis are summarised in Table 22. In the company’s main analysis, TheraSphere, SIR-Spheres and QuiremSpheres were associated with virtually identical numbers of QALYs, owing to the assumption of equal efficacy between interventions. They were all estimated to have similar total costs, with TheraSphere estimated to have marginally lower costs owing to lower rates of AEs requiring treatment.

Similarly, for TACE, DEB-TACE and TAE, marginal differences were observed owing to assumed differences in AE rates and unit costs of treatment.

DEB-TACE was estimated as being the strategy with the lowest costs owing to the fewer procedures required, and was used as the reference treatment in the incremental analysis. This resulted in an ICER of £24,647 for each of the SIRT versus DEB-TACE, and TACE and TAE being dominated versus DEB-TACE.

The probabilistic version of the model produced similar results, with the ICER relative to DEB-TACE being £25,052 per QALY.

Probabilistic results

Uncertainty surrounding model parameters was explored using scenario analyses and PSA; the executable model also included a number of DSAs that were not presented in the company submission or appendices. The company’s probabilistic results were estimated from 1000 Monte Carlo samples and were presented using CEACs and cost-effectiveness acceptability frontiers (CEAFs) only, with no ICERs from the probabilistic model presented in the company submission.

Figure 11 presents the results of the company’s PSA. Up to a threshold of approximately £25,000 per QALY, the company model estimated the treatment with the highest likelihood of being cost-effective to be DEB-TACE. After this point, the probability of being cost-effective was highest for the three SIRTs, which had similar probabilities of cost-effectiveness.

FIGURE 11.

Cost-effectiveness acceptability curve (CTT-eligible population). Reproduced with permission from BTG Ltd. 103

Reproduced with permission from BTG Ltd.

Scenario analyses

Table 23 presents the results of the company’s scenario analysis. The most influential parameters, of those assessed by the company, relate to the proportion of patients who transition to resection, and the proportion of patients who were downstaged after treatment with TheraSphere. Although the amount by which the proportion of patients was varied was arbitrary, and the ICER does not specifically represent a potential upper bound, this analysis showed that the model was most sensitive to this parameter.

Downstaging and role of transplant in the UK

The company assumed that patients who are successfully downstaged become eligible for transplantation, and that no patients receive any other kind of curative therapy including resection or ablation. This was justified on the basis that few patients are expected to receive these other therapies. The company provided two sources in support of this assumption. In these studies, of the patients who received radical curative therapy after downstaging, the proportion who received resection ranged from approximately 5.9%129 to 10%. 118

Clinical advice received by the AG also suggested that at least a proportion of these patients would go on to receive resection rather than a transplant. This AG therefore considers the assumption that all patients will go on to receive a transplant to be unreasonable and likely to favour SIRT, as outcomes following resection have been demonstrated to be associated with poorer outcomes (recurrence and survival) than those following transplantation. 122 The relevance of downstaging to transplantation in UK practice is also unclear. Eligibility for transplantation in the UK has historically been defined by the Milan criteria,130 and only recently has a service evaluation been introduced in which eligibility criteria have been expanded to permit downstaged patients to receive a transplant. 131,132 Furthermore, at the time of writing, this study has recruited only a small number of patients, and does not represent established national practice.

Modelling of pharmacological management

The progression status of patients in the pharmacological management health state was estimated as a 50 : 50 average of patients in the pre-progressed and post-progression states. This split is arbitrary and unlikely to accurately reflect the actual proportion of patients in each health state. A visual comparison of the PFS and OS extrapolation plots for sorafenib and BSC in the SHARP study69 appears to show that a greater proportion of time is spent in the post-progressed health state. A more reasonable estimate of the ratio of patients in each group is likely to be 33 : 67. Furthermore, given that the PFS and OS plots for SHARP are available, time in state could have been explicitly modelled, avoiding the need for such an assumption. The implications of this assumption are important and may lead to overly pessimistic estimates for patients in this health state, as this split is used to estimate utility and cost of active treatment. Based on the 50 : 50 split assumed, this will tend to overestimate total QALYs as too many patients are assumed to be in the pre-progressed state, as well as overestimating costs associated with time on sorafenib, where treatment duration is linked to progression.

Exclusion of patients who received selective internal radiation therapy work-up procedure but not treatment with selective internal radiation therapy

An important omission from the economic analysis is the costs and outcomes associated with patients who receive work-up associated with SIRT but who subsequently do not receive SIRT. These costs should be included in the economic analysis because work-up costs will be incurred by the NHS if SIRT was to be implemented in practice. Furthermore, patients who fail the work-up procedure are likely to be different from those who go on to receive treatment, as demonstrated in the SARAH trial,19 in which patients who failed work-up had significantly poorer outcomes than those who went on to receive SIRT. Excluding these patients from the analysis therefore underestimates total costs in the SIRT treatment arms and is likely to overestimate treatment benefits.

Modelling of comparator treatments

The company assumed equivalent efficacy between the SIRT treatments due to the paucity of comparative data, which the AG considered reasonable given the lack of data, and similarities in the treatment modalities. However, the BTG company submission103 states that it considers this assumption to be conservative, and that it might be expected that TheraSphere would provide superior outcomes. The AG notes that no plausible clinical argument or clinical evidence was provided in support of this statement.

Downstaging outcomes

The key benefit of SIRT in this analysis was through the increased proportion of patients who achieved downstaging after treatment, which indirectly led to increased numbers of patients receiving curative therapy. The probability of downstaging was estimated using data from a study of TheraSphere and TACE patients. 29 The AG had concerns relating to the robustness and generalisability of this study. The study was retrospective and single centre, with non-randomised cohort arms, which could have left it open to confounding bias. Furthermore, the study retrospectively identified patients who were most likely to be downstaged to curative therapies and therefore the modelled population is not representative of the broad CTT-eligible population in the scope of the analysis, and predicts higher rates of downstaging than would otherwise be observed for this broader population.

There are also issues regarding the generalisability of the downstaging criteria applied in the Lewandowski et al. 118 study, which were based on tumour dimensions only. However, UK criteria, used in the UK service evaluation of downstaging, also take into account alpha-fetoprotein level. 131 This may mean that there are differences between these patients and those considered eligible for transplant in the NHS.

To estimate the transition of patients to the pre-transplant wait list, the observed probability of downstaging from the Lewandowski et al. 118 study was applied to the proportion of patients who remained in the ‘watch and wait’ health state, rather than being applied directly in the model. As a result, this method underestimated the proportion of patients who were downstaged: for TheraSphere, the model predicted that 48% patients were downstaged, compared with 58% reported by Lewandowski et al. ,118 and for TACE, the modelled versus observed proportion who were downstaged was 26% vs. 35%.

The company assumed that the mortality rate of patients in the ‘watch and wait’ health state was equivalent to that of the pre-transplant mortality rate, citing a lack of data to model this specific outcome. However, the Lewandowski et al. 118 study reported mortality rates that were censored to curative therapies, and it was unclear why these were not leveraged in the model. The same mortality rate was applied to both treatment arms, thereby assuming that the only impact of treatment on mortality is through the bridging of patients to transplant. Furthermore, the data used to estimate pre-transplant mortality was from a cohort of patients,121 of whom only a proportion had HCC. The Lewandowski et al. 118 study also reported progression outcomes, which again were not used in the economic analysis.

The use of different sources for downstaging, progression and mortality outcomes also means that the evidence was derived from very different study populations, which led to a lack of internal consistency, and made it more difficult to validate the predictions of the model.

Transplant wait list outcomes

The data source used to estimate the time spent on the transplant wait list was estimated for a cohort of patients not specific to HCC. Patients on the transplant wait list are prioritised by their Model for End-Stage Liver Disease (MELD) score;133 however, the presence of HCC adds ‘exception points’ to MELD, meaning that the wait list time is generally shorter for HCC patients. The AG obtained data from a report on the 1-year outcomes following the introduction of the National Liver Offering Scheme, which was implemented on 20 March 2018. 134 The median waiting time under the old offering scheme may not accurately reflect how long patients may wait under the new offering scheme. The median waiting time to transplant for HCC patients who received a transplant between 20 March 2018 and 19 March 2019 was 49.5 days, which is substantially lower than the value for the overall cohort.

The company provided an incomplete reference on the source of the data used to estimate the transition to pharmacological management, and so it was not possible to comment on the suitability of this source. In an interim report on a service evaluation of transplantation following downstaging of HCC patients in the UK,132 of 27 patients enrolled in the programme to date, only one was removed from the wait list owing to the deterioration of their condition. This provides a much lower estimate of dropout than that estimated by the company, although the AG acknowledges that it is based on a smaller subset of patients.

The AG questions whether or not it is appropriate to apply the same transition probabilities and mortality rate to patients regardless of their initial treatment; however, the AG is not aware of any directly applicable evidence for a differential rate. There are many factors that determine the rate at which patients receive transplant; some of these will not be treatment dependent, including the availability of donor grafts, and some are dependent on treatment. Previous studies of SIRT and CTT with intent to downstage have demonstrated differential outcomes of transplantation and progression between treatment arms; although these are based on very small patient numbers, there does appear to be a small benefit in favour of SIRT. 25,47 Although TheraSphere and TACE were given as downstaging rather than bridging therapies in the Lewandowski et al. 118 study and so are not directly applicable to outcomes for patients on the transplant wait list, OS censored to curative therapies was also significantly different between arms in favour of SIRT, particularly after 2 and 3 years. Similarly, the rate at which patients receive curative therapy following downstaging is also likely to differ between arms, as evidenced in the Lewandowski et al. 118 study. As such, the AG considers it unlikely that outcomes would be equivalent across different treatment modalities, although it is not possible to estimate directly without estimates of survival conditional on downstaging.

Pharmacological management

Outcomes for patients in the pharmacological management health state were based on the BSC arm of the SHARP trial;69 the company stated that this was to avoid applying any benefit associated with a particular HCC treatment in the model, as patients modelled to receive pharmacological management would be given different treatments. This is not representative of patients in this health state, as a proportion of these patients would receive further active therapy, assumed by the company to be sorafenib. Because patients receiving sorafenib experience better outcomes than patients on BSC (as demonstrated by a HR of 0.69 for OS in SHARP69), this approach underestimates survival for patients in this health state. A more accurate approach would be to calculate outcomes separately for sorafenib and BSC and then weight according to the proportion of patients in the health state over time.

Furthermore, the SHARP trial69 is unrepresentative of the patients who would receive BSC in this population for a number of reasons. Approximately 50% of patients in SHARP had extrahepatic spread, and would thus be contraindicated for SIRT treatment. A subgroup analysis of SHARP patients demonstrated that the sorafenib treatment effect was higher in patients with no extrahepatic spread (HR of 0.55 compared with 0.69 in the ITT population). Data from REFLECT,12 which compared lenvatinib with sorafenib, also demonstrated that the prognosis for patients with extrahepatic spread is worse than for those without: in the ITT population, the median OS was 12.3 months, compared with 18.0 months in a population with no extrahepatic spread. In addition, the SHARP trial enrolled only patients who had not received previous treatment with systemic therapy, so BSC patients in SHARP do not represent the patients in the pharmacological management health state who previously received TACE or SIRT. The AG was advised that patients who present with HCC and are eligible for sorafenib are typically associated with a more rapidly progressing form of the disease and will have a higher mortality rate.

As a result, the cost-effectiveness analysis is biased in favour of SIRT through the selection of unrepresentative comparator data. The use of these data from SHARP69 underestimates survival in the pharmacological management health state, thereby further inflating the relative treatment effect of SIRT, as fewer patients on SIRT than on other therapies enter this health state.

Post-transplant outcomes

The AG has concerns about the applicability of the sources used to estimate mortality following liver transplantation, and considers it uncertain whether or not the assumed treatment pathway is reflective of clinical practice.

The data set used to estimate long-term mortality after transplant is not specific to patients with HCC. Patients with HCC are at risk of tumour recurrence, which is linked to increased mortality. 122 This can be illustrated by a comparison of survival in the general liver transplant population and in a HCC population. The AG obtained a HCC-specific data set of survival outcomes for liver transplant recipients in the UK since 1994. 134 In this data set, patients with HCC (restricted to those aged > 60 years as a proxy for intermediate-HCC patients) had a 5-year survival of 71%. This was lower than for those in the general liver transplant data set, whose 5-year survival was estimated as 81%. Therefore, benefits estimated by the company model are likely to be overestimated.

By excluding tumour recurrences, the treatment pathway is also misrepresented by the model. Both the Bellavance et al. 122 and Lewandowski et al. 118 studies report on recurrences that occur after transplantation: approximately 20% of patients in the Lewandowski et al. 118 study and 14% of patients in the Bellavance et al. 122 study experienced recurrence after transplantation, with a 1-year relapse-free survival rate of between 73% and 89%. In addition, the AG found that, in its analysis of the HCC-specific transplant data set, > 10% of transplant recipients in the UK in this population experienced a recurrence in the first 5 years post transplant. The patients who experience a recurrence are at an elevated risk of death,122 and these patients often experience a reduced quality of life and additional treatment-related costs. 135 By excluding recurrence after transplant, the model overestimates the QALYs and underestimates costs generated for transplant recipients, which biases the results in favour of the SIRT arm owing to a higher proportion of patients being downstaged.

Health-related quality of life

The total number of QALYs generated by the model is likely to be underestimated, owing to the source chosen and an error in how age-related disutility was applied.

Health state utility values were estimated from a range of sources, but were primarily based on the NICE appraisal of lenvatinib (TA551),12 which enrolled patients with advanced HCC, of whom approximately 60% had extrahepatic spread. This population therefore had more advanced disease and does not reflect the model population of intermediate-HCC patients. Therefore, the utilities drawn from TA55112 are likely to underestimate the quality of life for a CTT-eligible population, and disadvantages any treatment arm associated with increased life-years.

The AG considers the company to have incorrectly implemented age-related disutilities in the model, although the company contends that the application was appropriate. This ‘error’ has an impact on all health states, and results in patients experiencing much lower utilities than those observed in the cited sources. In the company’s model, the decrement associated with ageing is estimated by estimating an absolute utility decrement for each health state relative to full health (i.e. 1 minus the reported health state utility) and then subtracting this decrement from the age- and gender-adjusted population norm from Kind et al. 127 For example, as patients enter the model at the age of 65 years, the age-adjusted utility started at 0.78, and the literature-derived absolute utility for ‘watch and wait’ patients was 0.75. 12 This meant that the age-adjusted utility for patients in the ‘watch and wait’ health state was 0.53 (0.78 – 0.25). The application of age-adjusted utilities in this way is inappropriate and ignores the fact that each health state utility is derived from an age-appropriate source, and thus already accounts for any age-related decline in HRQoL. Furthermore, this method is inconsistent with previous TAs136–138 in which age-related disutilities have been applied, where age-related decrements are applied as a multiplier to health state utilities rather than as an absolute decrement.

Resource use estimates

Resource use was estimated in the model based on feedback from a single clinician at a centre in the UK that uses SIRT. As the company could not provide details of the questionnaire or transcript of the interview, it has not been possible to verify how these data were estimated. Therefore, there are a number of uncertainties regarding which treatment costs are included, such as AEs relating to subsequent therapy (sorafenib) or to transplant, or whether or not any bridging therapy was provided for patients on the transplant wait list.

The company’s clinical expert advised that TACE and TAE patients had around three initial treatments in their respective arms, whereas patients in the DEB-TACE arm had an average of 1.5 initial treatments. As described in Chapter 5, Sirtex submission: conventional transarterial therapy-eligible analysis, there is apparent variation in the number of treatments that patients receive in practice, with values identified between 1.43 and 2.83 per patient for DEB-TACE and between 2.5 and 3.03 for TACE patients. The uncertainty in these numbers was not explored by the company. By implementing a single embolisation cost for each CTT procedure, the company also did not explore any differences in the length of hospital stay between the different CTT treatments.

A proportion of patients in the pharmacological management health state receive sorafenib. This was estimated using data obtained from a survey of clinicians; as there were limited details provided on how the proportion was estimated, the underlying assumptions could not be validated. It appears that the cost of sorafenib was applied for the time that patients were in the pre-progression health state; however, this would overestimate the cost of treatment, because mean time on treatment with sorafenib is less than mean TTP. 11 The analysis also excludes patients who receive lenvatinib instead of sorafenib, and the proportion of patients who progress on sorafenib and receive subsequent treatment with regorafenib; clinical advisors to the AG suggest that this would be approximately 20% of patients.

The company assumed that the work-up procedure for each SIRT would be associated with the same resource use. This underestimates the costs for QuiremSpheres, as the use of QuiremScout is required and is associated with an additional procurement cost.

Overall survival

Overall survival for patients receiving TheraSphere was based on a single-arm Phase II trial of 52 patients with intermediate and advanced HCC. 139

The following standard parametric survival models were fitted to the observed data: Weibull, log-normal, log-logistic, exponential and gamma functions. Assessment of the most appropriate parametric extrapolation was made with reference to statistical goodness of fit and clinical plausibility of survival estimates. The log-logistic and log-normal curves were eliminated on this basis, as they predicted that a small proportion of patients would not die within the time horizon of the model. The Weibull function was selected for the base-case analysis; no other extrapolations were explored in scenario analysis.

Estimation of OS for comparator therapies was based on a NMA of studies identified in the presented clinical effectiveness review. The NMA drew evidence from RCTs as well as non-comparative studies. The primary NMA reported better survival for TheraSphere than for sorafenib [HR (confidential information has been removed), 95% CrI (confidential information has been removed)], although this was not statistically significant.

Progression-free survival

Modelling of TTP for TheraSphere was implemented by fitting standard parametric functions to reported KM data from the same Phase II study used to model OS. 139 TTP was defined from first SIRT to first progression at any site. TTP, therefore, excluded mortality events, as the model only permits death following progression. As with OS, standard parametric curves were fitted to available KM data and the exponential function was selected as the most appropriate survival model based on the clinical plausibility of predicted outcomes. No other parametric functions were explored in the presented scenario analyses.

Owing to inconsistent reporting of TTP in the studies identified in the systematic review, a NMA for TTP was not feasible. Time-to-progression outcomes for comparator therapies were, therefore, based on a naive comparison, generated via median TTP and PFS data from relevant sources, which were converted to survival curves by assuming that TTP followed an exponential function. Median TTP for SIR-Spheres was based on a retrospective cohort study of patients who received SIR-Spheres,45 with TTP assumed to be the same for QuiremSpheres owing to a lack of appropriate data. Median TTP for sorafenib was based on a weighted average of values reported in TA474,11 TA55112 and a retrospective cohort study. 45 Lenvatinib TTP was sourced from TA551,12 and median TTP for regorafenib was sourced from TA555. 13 Note that all values sourced from TAs were based on PFS rather than TTP.

Health-related quality of life

The primary source of utility data used by BTG was TA551,12 which drew evidence from the REFLECT trial81 comparing lenvatinib with sorafenib, which collected EuroQol-5 Dimensions, three-level version (EQ-5D-3L), values from participants. The values used assume no differences in HRQoL between treatment arms, but do not attempt to account for differences in HRQoL as a result of AEs. This was carried out by applying a one-off utility decrement in the first cycle of the model, which was estimated by applying a 0.012 decrement per grade 3 or 4 event. Note that the BTG company submission103 erroneously reports that a 0.014 decrement was applied in the model and miscalculates the decrement to be applied in the executable model.

In addition to the above, adjustments were also made to the health state utilities to account for the impact of ageing. This was undertaken by applying a decrement to every model cycle. The decrement applied was estimated by subtracting 1 from the age- and gender-adjusted population norm. Note that the BTG company submission103 erroneously reports the decrements applied as 0.26 for the progression health state and 0.32 for the progressive disease health state, when the model applies a common decrement to both health states, which changes over time to reflect the increased age of the cohort. General population utility norms were sourced from Kind et al. 127 Utility values applied in the base-case analysis along with utility decrements are reported in Appendix 15, Table 64.

Selective internal radiation therapy procedure costs

See the review of the CTT-eligible population model (Chapter 5, Evidence used to inform the company’s model) for details of SIRT procedure costs.

Drug acquisition costs: systemic therapies

Drug acquisition costs for sorafenib, lenvatinib and regorafenib were taken from the BNF. 115 Respective dosing was 800 mg, 12 mg and 160 mg per day. Dosing was based on recommended doses for HCC patients, described in their respective European Medicines Agency summary of product characteristics (SmPC). Duration of systemic therapy was based on progression, with patients assumed to continue systemic therapy until either progressive disease or death. Appendix 15, Table 65, summarises the drug acquisition costs applied in the model.

Subsequent treatments

A proportion of the patients receiving SIRT were assumed to receive sorafenib therapy following SIRT, with patients assumed to receive sorafenib after cycle 1 until disease progression or death. In the base-case analysis, the proportion of patients assumed to receive sorafenib was 33%, based on ‘data on file’. Patients not receiving concomitant sorafenib were assumed to receive BSC. No subsequent therapies were modelled following disease progression in either model arm (SIRT or systemic therapy).

Health state costs

Resource use estimates were based on a survey of clinical experts conducted to inform resource use in TA189,140 TA47411 and TA551. 12 These included physician visits, laboratory and radiological tests, and hospital stays. Unit costs were derived from TA189,140 and updated using National Schedule of Reference Costs 2017–2018. 107

In addition to the above, a one-off cost was applied on treatment progression based on the costs applied in TA551. 124 This comprised additional laboratory and radiological tests (see Appendix 15, Table 67).

Total costs by health state are reported in Appendix 15, Table 66, along with a summary of one-off progression costs.

Adverse event costs

Unit costs associated with AEs were drawn from National Schedule of Reference Costs 2017–2018107 and are summarised in Appendix 15, Table 68. No information or justification was presented with regard to how the specific costs used were selected.

Palliative care costs

The company’s model includes a cost of £8191 to account for costs of palliation at the end of life. This was derived from a joint Nuffield Trust and Marie Curie report128 into end-of-life cancer care and inflated to 2017/18 prices. This cost was applied on a patient’s death and was applied for all modelled interventions.

Model results

The headline results presented in the BTG company submission103 are based on the deterministic version of the model. Uncertainty surrounding model parameters was explored using scenario analysis and a PSA. The executable model also included a number of DSAs that were not presented in the company submission or appendices. The company’s probabilistic results were estimated from 1000 Monte Carlo samples and were presented using CEACs and CEAFs only, with no ICERs from the probabilistic model in the company submission.

Table 25 presents the company’s base-case estimates of cost-effectiveness using the corrected version of the model at the list price for sorafenib, lenvatinib and regorafenib. Based on the probabilistic version of the company’s model, regorafenib was estimated to be the most cost-effective therapy. The results of the fully incremental analysis suggested that SIR-Spheres, QuiremSpheres and lenvatinib were dominated by one or more therapies, whereas sorafenib was extendedly dominated by TheraSphere. The estimated ICER for TheraSphere compared with regorafenib was £69,070 per QALY and estimated that TheraSphere generates an additional 0.185 QALYs at an additional cost of £12,778. The deterministic version of the model produces similar results, with an ICER relative to regorafenib of £66,624 per QALY.

Figure 13 presents the results of the DSA generated by the AG. The most influential parameters (of those assessed by the company) relate to the OS HR for regorafenib and the proportion of patients assumed to go on to receive post-SIRT sorafenib. Additional scenario analysis presented by the company showed that the estimated ICER was influenced significantly by assumptions made about post-SIRT treatment. In the presented scenario analysis, in which no concomitant sorafenib was assumed, TheraSphere was estimated to be the most cost-effective intervention, with a deterministic ICER of £5870 per QALY.

FIGURE 13.

BTG DSA: tornado diagram. Reproduced with permission from BTG Ltd. 103

Reproduced with permission from BTG Ltd.

The AG questioned the face validity of the utility values applied, and were concerned that the company had made a calculation error with respect to the calculation of the utility decrements. After clarification from the company, BTG confirmed that the utility decrements applied in the model were as intended by the company. See below for further critique of the utility values applied.

Inappropriate inclusion of regorafenib as a comparator

The base-case analysis presented in the BTG economic analysis includes three systemic therapies: sorafenib, lenvatinib and regorafenib. The AG is of the view that regorafenib should not have been included as a comparator, as it is used only as a second-line therapy following sorafenib. This is stated in the SmPC for regorafenib and NICE’s recommendation13 for regorafenib, which restricts use to patients who have been previously treated with sorafenib. 141 The AG considers it entirely reasonable to model subsequent regorafenib use following sorafenib, but it should not have been directly compared with SIRT and the other systemic therapies.

Work-up without selective internal radiation therapy procedure

An important omission from the BTG economic analysis is the costs associated with patients who received work-up but did not continue on to the SIRT procedure. In the SARAH19 and SIRveNIB21 trials, 18.6% and 28.6% of patients, respectively, received work-up but did not continue on to receive SIRT. The AG considers the cost of patients who do not proceed to SIRT treatment important, as they constitute part of the incremental costs of implementing SIRT in the NHS. The AG further notes that many of these patients will receive other active therapies instead of SIRT, and it is therefore appropriate to model the associated costs and outcomes. For example, in the SARAH trial,19 62% of patients for whom work-up failed went on to receive sorafenib. The AG, therefore, considers that the costs associated with the administration of these alternatives should also be included in the economic analysis. The AG also notes that the clinical effectiveness data used to populate the model were based on the ITT population and, therefore, the clinical outcomes of these patients for whom work-up failed are implicitly included. This is inconsistent with BTG’s stated position that only patients receiving therapy were considered.

Network meta-analysis and estimation of relative overall survival benefits

BTG conducted a NMA to compare TheraSphere with sorafenib for the treatment of unresectable HCC patients. Seven studies formed the primary network: two RCTs, one prospective study and four retrospective studies. There are differences in the studies included in the NMAs conducted by BTG and the AG. The BTG network included only studies conducted outside Asia, owing to known differences in both aetiology and treatment patterns in Asian populations. The AG identified additional studies that the company did not include or identify in its systematic literature review. 40 Unlike the AG, the company did not split the NMA into different populations of patients with differing stages of HCC disease. Therefore, the baseline BCLC stage, Child–Pugh status and the proportion of patients with PVT differed across studies. However, the population in the primary network was mostly advanced-stage HCC patients.

The validity of results from the NMA relies on the quality of the studies that make up the evidence base. However, there are considerable concerns regarding the quality of the prospective and retrospective studies. The prospective observational study by Woodall et al. ,37 comparing TheraSphere with BSC, which was excluded from the AG’s NMA, presented significant baseline imbalances and evidence of selection bias, as patients who failed to meet the pre-treatment TheraSphere requirements formed the ‘no-treatment’ arm. In addition, the retrospective studies39,44,45 were all associated with a high risk of bias as there are significant baseline imbalances, unclear reporting of blinding and missing outcome data, and were excluded from the AG’s primary NMA for these reasons.

Although the NMA reports better survival for TheraSphere than for sorafenib, this appears to be on the basis of the inclusion of a particular retrospective study, Biederman et al. ,39 which reports a very strong treatment effect on OS with TheraSphere compared with SIR-Spheres (HR 0.40, 95% CrI 0.20 to 0.78). As discussed in Chapter 3, Risk of bias, the four retrospective studies (including Biederman et al. 39) and the prospective observational study are poor quality and were rated as being at a high risk of bias, which reduces the reliability of the NMA results.

Limited exploration of uncertainty surrounding survival functions

The BTG company submission103 does not include any consideration of the uncertainty surrounding the range of potentially plausible survival functions for OS. Although a number of parametric functions were fitted to the available data for OS, the impact of alternative functions was not explored in the company’s presented scenario analyses. Furthermore, there is no functionality in the presented executable model to implement alternative survival functions.

Omission of downstaging

The AG notes that the BTG economic model did not consider the possibility that patients may be downstaged to receive curative therapy. As stated in relation to the Sirtex CTT-ineligible model, the relevance of downstaging in an advanced-HCC population is unclear, with the AG’s clinical experts suggesting that this would be a very rare occurrence in UK practice. However, downstaging was observed in a small number of patients in the SARAH trial19 and, therefore, the potential benefits of downstaging represent an important uncertainty. Therefore, although the AG recognises that the inclusion of downstaging in the company’s base case may be inappropriate, this uncertainty should have been explored in scenario analysis.

Modelling of progression-free survival

The BTG company submission states that it was not possible to obtain estimates of relative PFS from the NMA and, therefore, PFS was based on a naive comparison of reported estimates from studies identified as relevant by the company. The AG considers there to be a number of significant weaknesses in the company’s approach, and that the selected median PFS for TheraSphere lacks face validity. Although the AG acknowledges that a NMA could not be run for PFS outcomes, based on the studies included in the company’s network, the AG does not agree that a relevant network could not have been constructed (see Chapter 4). Importantly, as reported in Chapters 3 and 4, there are randomised comparisons of SIRT (SIR-Spheres) and systemic therapies (sorafenib) on which estimates of median PFS could have been based. The AG would consider such an approach preferable to the company’s naive comparison, which used populations poorly matched with the modelled population. The AG further notes that this randomised evidence was ignored in favour of studies used in the relevant NICE appraisals, which focused on populations including a significant proportion of patients with extrahepatic spread, and, with respect to regorafenib, had already failed previous sorafenib therapy.

Further to the above, the AG also questions the plausibility of the modelled median PFS for TheraSphere. The modelled value of 11 months is 3.5 times longer than the value used for SIR-Spheres (3 months) and longer than the median OS reported in the SARAH trial19 for both SIR-Spheres and sorafenib. Given the broad clinical similarity between TheraSphere and SIR-Spheres, and the lack of high-quality comparative evidence, the AG considers that it is unreasonable to assume such a large disparity in PFS.

Dosing and time on systemic therapy

Dosing of systemic therapies in the BTG economic analysis was based on the relevant SmPC, with a dose of 800 mg, 12 mg and 160 mg assumed for sorafenib, lenvatinib and regorafenib, respectively. These figures are likely to overestimate the dose received for all three drugs, as dose reductions and interruptions are common in patients receiving systemic therapy, and were observed in all relevant trial data. For example, the mean dose of sorafenib received in the SARAH trial19 was 648 mg, not 800 mg. The company’s model also does not account for the fact that the dosing of lenvatinib is weight dependent, with patients < 60 kg receiving 8 mg daily; 13% of patients in the Western subgroup of the REFLECT trial81 weighed < 60 kg.

Time on systemic treatment in the BTG economic analysis is assumed to align with PFS. This is consistent with the SmPC for both sorafenib and lenvatinib, both of which indicate that therapy should continue for as long as clinical benefit is observed, or until toxicity becomes unacceptable. However, sorafenib, lenvatinib and regorafenib are all associated with significant tolerability issues, which means that many patients discontinue therapy prior to disease progression. This is seen in the pivotal trials, in which time on systemic therapy is always less than PFS. For example, median time on sorafenib in the SARAH trial19 was 2.8 months, whereas median PFS was 3.7 months. Using PFS as an indicator of treatment discontinuation, therefore, may produce overestimates of time on treatment and consequently total drug acquisition costs for sorafenib, lenvatinib and regorafenib.

Subsequent therapy costs

The BTG economic analysis assumes that a proportion of patients receiving SIRT treatment (TheraSphere, SIR-Spheres or QuiremSpheres) move on to receive subsequent systemic therapy immediately following initial SIRT. These patients are assumed to continue therapy until disease progression. The AG considers the modelling of subsequent therapy in this way to be inconsistent with likely NHS practice and the supporting trial evidence, and that initiation of systemic therapy following SIRT would typically happen following disease progression. The AG acknowledges that, in the SARAH trial,19 a proportion (11/52, 21%) of patients did receive subsequent systemic therapy prior to progression. However, there is no evidence to suggest that this was initiated immediately following SIRT; indeed, the SARAH19 and SIRveNIB21 trial protocols stipulated that further therapy should not commence until disease progression.

A further issue relating to the company’s modelling of subsequent therapy is the assumption that patients receiving first-line sorafenib therapy will not receive further active therapy following progression. This is inconsistent with clinical practice, in which a proportion of patients will receive second-line regorafenib as per NICE’s recommendations. 13 It is also not consistent with the modelled trial evidence, as a proportion of patients in the SARAH and SIRveNIB trials went on to receive subsequent therapy following discontinuation of sorafenib.

Application of age-adjusted utilities

Similar to the BTG economic analysis in the CTT-eligible population, the estimation of age-related disutility was implemented incorrectly, resulting in health state utilities being applied that are inconsistent with values used in previous TAs, as well as values reported in the SARAH trial. 19 For further details of this error, see Chapter 5, Evidence used to inform the company’s model.

Further to the above, the AG considers age adjustment unnecessary in an advanced population in which the majority of patients die within 5 years; the application of age-adjusted utilities is unnecessary and not in keeping with norms for this type of model.

Calculation errors

A small number of calculation errors were identified and corrected as part of the AG’s assessment of the BTG economic analysis. These errors related to:

• the estimation of the comparator TTP, which used an incorrectly estimated HR

• the calculation of per-cycle mortality and progression, which were estimated using a monthly cycle, whereas the rest of the model used a 4-week cycle.

These errors have only a marginal effect on the reported ICER, increasing the deterministic ICER from £64,693 to £66,624 per QALY.

Overall survival

The analysis of OS for the base-case analysis was based on time-to-event data from the SARAH trial supplied by Sirtex, and KM curves from the SIRveNIB trial. 21 Pooled KM curves for the base-case population are presented in Appendix 16, Figures 26 and 27. Survival estimates can be found in Appendix 16, Table 71.

Standard parametric survival functions were fitted to the survival data available for each of the considered populations, and log-cumulative hazard plots were generated to assess any changes in hazards over time (see Appendix 16, Figure 28). Plots of each of the fitted parametric models with the observed KM OS curves are presented in Figures 15 (SIR-Spheres) and 16 (sorafenib). Model fit statistics are summarised in Appendix 16, Table 72, which showed that the generalised gamma model had the best fit, with the log-normal and log-logistic curves also having similar statistical fit, thereby providing little justification to discriminate between these models on this basis of fit statistics. The generalised gamma, log-normal and log-logistic models are, however, all accelerated failure time models and, as such, a HR cannot be applied to estimate outcomes for lenvatinib patients, and would likewise not permit scenarios in which differential outcomes are assumed for TheraSphere, which would similarly require the application of a HR. To accommodate the use of HRs, the AG base-case analysis, therefore, selected the Weibull function, which has the best statistical fit from the remaining curves, and was considered the most clinically plausible. The AG considered this reasonable given the limited data to accommodate accelerated failure time functions and the small variation in predicted incremental survival across all six functions, but acknowledges this as a limitation of the presented base-case analysis. Scenario analysis is, therefore, presented, in which the generalised gamma, log-normal and log-logistic functions are used to model OS. In these scenarios, equivalence is assumed between sorafenib and lenvatinib.

FIGURE 15.

Extrapolation of OS: SIR-Spheres.

FIGURE 16.

Extrapolation of OS: sorafenib.

For scenarios run on the SARAH trial19 subpopulations described previously, the Weibull function was retained to model OS outcomes. Fit statistics for the SARAH trial whole population, low tumour burden/ALBI 1 subgroup and no-MVI subgroup are reported in Appendix 16, Table 74. Plots of each of the fitted parametric models with the observed KM OS curves are presented in Appendix 16, Figures 30 and 31 (SIR-Spheres) and Figures 32 and 33 (sorafenib). In all three scenarios, the Weibull function had a good statistical and visual fit to the observed data.

Progression-free survival

The analysis of PFS for the base-case analysis was based on supplied time-to-event data from the SARAH trial19 and KM curves from the SIRveNIB trial. 21

Similar to the approach previously described for OS, standard parametric survival functions were fitted to the survival data available for each of the considered populations (Figures 17 and 18), and log-cumulative hazard plots generated to consider the change in hazards over time (see Appendix 16, Figure 29). Plots of each of the fitted parametric models with the observed KM OS curves are presented in Appendix 16, Figures 34 and 35 (SIR-Spheres) and Figures 36 and 37 (sorafenib). Similar to OS, model fit statistics for the generalised gamma, log-normal and log-logistic functions were superior to other functions (see Appendix 16, Table 73). These functions were, however, rejected to accommodate the application of a HR for lenvatinib and the implementation of scenarios assuming differential effectiveness for TheraSphere. The Weibull function was, therefore, selected in the AG base-case analysis as this had the best statistical and visual fit to the observed data and was considered clinically plausible.

FIGURE 17.

Extrapolation of PFS: SIR-Spheres.

FIGURE 18.

Extrapolation of PFS: sorafenib.

Overall survival for patients downstaged to curative therapy

The base-case analysis does not allow for downstaging to curative therapies, owing to uncertainties over whether or not this is realistic in a population of patients with advanced disease. A number of scenarios are presented in which downstaging is allowed for. The proportion of patients downstaged is based on the values reported in the SARAH trial19 and varied depending on the efficacy subgroup used (see Appendix 16, Table 69). Outcomes for patients downstaged to curative therapy were based on a US prospective cohort study,112 which recruited 267 patients with HCC, including 191 with intermediate and advanced disease. This study compared outcomes for patients who had received palliative care with those who received potentially curative therapies (liver transplantation, surgical resection or tumour ablation). Using Cox multivariate proportional hazards, the HR for OS with potentially curative treatments versus non-curative treatment was 0.29 (95% CI 0.18 to 0.47). This HR was applied to the pooled sorafenib ITT arms of the SARAH and SIRveNIB trials in all scenarios. This was carried out to prevent the outcomes of downstaged patients varying depending on the patient population selected or by treatment arm; advice from clinical advisors to the AG suggested that outcomes post-curative therapy would be similar regardless of patient characteristics or treatment received to achieve downstaging. The sorafenib ITT arm was used as this was considered to best match care received in the analysed patient cohort, and is most representative of the current standard of care in UK practice.

Age-related disutilities

Age-adjusted UK population norms from Szende et al. 146 were applied to the utility values included in the model. Age-related decrements were estimated in the form of a multiplier, with decrements applied relative to the populations on entering the model. This allows for the trial-derived utilities applied in the model to account for age-related decline in HRQoL as the population ages over time.

Selective internal radiation therapy health state utilities

The health state utilities associated with SIRT in the CTT-ineligible model were based on the per-protocol subgroup of the SARAH trial as calculated by Sirtex in its evidence submission (see Chapter 5, Evidence used to inform the company’s model, for details). EORTC QLQ-C30 summary scores were mapped to EQ-5D using the algorithm developed by Longworth et al. ,113 and utilities were calculated based on UK general population weights.

The per-protocol utilities were considered to better reflect the HRQoL associated with SIRT than those derived from the ITT population, as 22.4% of patients randomised to SIRT did not receive SIRT in the SARAH trial. These patients may have received other systemic therapies or BSC, or were otherwise too unwell to receive SIRT; thus, the ITT utility values may not have represented those of a SIRT-treated population. There were no further utility decrements applied to these utilities as these are likely to have been captured in the SARAH trial results. The health state utilities applied in the model are presented in Table 28.

Systemic therapy health state utilities

Health state utilities applied to modelled patients receiving the systemic therapies sorafenib and lenvatinib were taken from the per-protocol subgroup of sorafenib patients in the SARAH trial. 19 The difference in utility between SIRT and sorafenib in this subgroup was 0.011, which the AG considered to account sufficiently for the ostensibly greater burden of AEs associated with these drugs. Utilities applied to patients who received work-up but ultimately did not receive SIRT were weighted by the proportion on systemic therapy versus BSC (61.9% and 38.1%, respectively). This assumes that patients not on systemic therapy had a utility equivalent to those on SIRT, which may overestimate the HRQoL of BSC patients, as a proportion were likely to have been too unwell to receive systemic therapy.

Post-transplant health state utilities

The AG scenarios 6 and 10 include the possibility for downstaging; therefore, post-transplant utilities were considered for use in the model. Pre-transplant health state utilities are assumed to be equal to those experienced in pre-progression for SIRT, systemic therapies and BSC. Post-transplant health state utilities are assumed to be equal to those experienced on SIRT, regardless of which treatment a patient received before downstaging to transplant. However, it is likely that patients who received a transplant may have a better HRQoL than the per-protocol population of the SARAH trial.

Despite multiple studies showing that recipients of liver transplant enjoy increased HRQoL post transplant in comparison with pre transplant,113,147–149 a lack of generalisability between these studies and the population included in the model renders the absolute utility values reported in the literature too uncertain for inclusion. Studies also show that HRQoL remains lower for liver transplant recipients than for healthy patient controls. 150–152 However, as with the pre- and post-transplant utilities, there is insufficient evidence to suggest that these studies are generalisable to the modelled population. Given the lack of evidence to definitively suggest that utility values in the post-transplant HCC population are lower than in the general population, the AG believes that the utility values observed in the general population represent the upper bound of the utility expected in the post-transplant population.

Work-up costs and number of procedures

Patients allocated to receive SIRT must first undergo a work-up procedure to assess their suitability for treatment with SIRT, and to plan the procedure through angiographic evaluation and occlusion of any vessels that could carry microspheres away from the liver to the gut. Although work-up is a one-off procedure, those patients who required a second SIRT procedure owing to an unsuccessful or incomplete first procedure are likely to need a second work-up.

In the SARAH trial,19 17 of the 184 patients who received SIRT required re-treatment owing to an unsuccessful or incomplete first procedure (nine received a second work-up but were not re-treated). Therefore, patients who received any of the SIRTs incurred the cost of 1.09 work-up procedures to account for re-treatment. As the model independently considered the costs and outcomes for patients who underwent work-up but ultimately did not receive SIRT, these individuals were assumed to receive 1.0 work-up procedures. The AG’s base case assumed that 18.6% of patients who underwent work-up did not go on to receive SIRT in line with the SARAH trial19 data. However, in recognition of the uncertainty around this value, a number of alternative scenarios are presented in Sensitivity analyses results.

Work-up costs used in the AG base case were based on the values BTG elicited from The Christie NHS Foundation Trust (see Appendix 15, Table 60). The largest expenditures were staff costs and SPECT/CT. The total cost of a single work-up procedure for SIR-Spheres and TheraSphere used in the AG model was £860.32, and the work-up cost of £5178.32 for QuiremSpheres comprised the list price of QuiremScout and the BTG-elicited value excluding the £74 cost of the technetium-99m MAA agent. This does not include the PAS discount available for QuiremScout.

Selective internal radiation therapy treatment costs and number of procedures

Patients in the AG model received an average of 1.21 SIRT procedures. This is based on the assumption that patients requiring bilobar treatment will require two separate SIRT procedures, separated by a few weeks (as per the SARAH protocol153), and that patients will be re-treated owing to an incomplete or unsuccessful first treatment. The clinical advisors to the AG stated that it would be very unlikely that both lobes would be treated in the same treatment session in UK practice owing to an increased risk of REILD. SIRT patients in the SARAH study19 had 1.28 separate SIRT treatments on average [222 treatments, 173 patients (one or two treatments only)]. This broadly reflects the results of the Sirtex resource use survey (1.2 treatments per patient). This value excludes the 11 patients who had three separate SIRT treatments, and includes only one procedure for the nine patients who received a second treatment owing to disease progression, as it was unclear whether or not this would be permitted in UK practice.

The acquisition cost of a single SIRT treatment was taken from each company submission: SIR-Spheres, £8000; TheraSphere, £8000; and QuiremSpheres, £9896.

The cost of the SIRT procedure applied in the AG model was taken from National Schedule of Reference Costs 2017–2018107 (YR57Z). The average cost of ‘Percutaneous, Chemoembolisation, or Radioembolisation, of Lesion of Liver’ was £2790. This cost was incurred for each separate SIRT administration for patients receiving TheraSphere and QuiremSpheres in the AG model. The Sirtex company submission102 stated that SIR-Spheres administration procedures use intermittent contrast-medium injection to assess the distribution of the microspheres under radiography over the course of approximately 1 hour. The AG, therefore, included an additional cost of £209 for the SIR-Spheres administration procedure (RD32Z – Contrast Fluoroscopy Procedures with duration of more than 40 minutes), for a total of £2999.

Costs of systemic therapies

The pack costs for sorafenib (£3576.56), lenvatinib (£1437.00) and regorafenib (£3744.00) were taken from the BNF. 115 The confidential PAS discounts available for sorafenib, lenvatinib and regorafenib are not included in this report. For results of the AG’s economic analysis that include these discounts, please refer to Appendix 17.

The daily dose of sorafenib used in the AG base case was based on the SARAH trial19 (648.5 mg), and the mean time on treatment was calculated by applying an exponential function to the median time on treatment reported in the SARAH trial19 (exponential mean 122.95 days).

The base-case daily dose of lenvatinib was 10.2 mg per day, based on the Western subgroup of the REFLECT trial81 for lenvatinib. This value was considered by the technology appraisal committee in TA55112 to better represent the average weight-based dose used in UK practice. The AG considered the time on treatment reported in the REFLECT trial81 for lenvatinib to be excessively long compared with SARAH,19 and reflective of differences in the baseline characteristics of the populations recruited to these trials. To avoid inflating the relative cost of lenvatinib, the AG applied the reported HR of PFS between lenvatinib and sorafenib in REFLECT to the SARAH time on treatment to produce an estimate of 124.07 days on treatment.

Wastage was accounted for in the AG model using the simple assumption that if a new pack was started then in the case of treatment discontinuation, the remainder could not be used to treat other patients. However, this may be a conservative assumption, as it was reported in TA55513 that many centres have measures in place to reduce wastage of expensive cancer treatments, such as issuing only a 1-month supply of tablets at a time (approximately one pack of sorafenib). However, as it generally cannot be predetermined when therapy will be discontinued owing to AEs, death or non-compliance, it can be reasonably assumed that some wastage will occur.

Cost of subsequent treatment

The interventions used following first-line treatment in the SARAH trial19 were not representative of current UK practice; however, as the efficacy data used in the model are derived from these patients, the trial values are most appropriate. Therefore, the proportion of patients who received subsequent systemic therapy (98% sorafenib) following SIRT in the SARAH trial19 (28.8%) was used to estimate the size of this population in the AG model. The AG was advised that current NICE recommendations mean that lenvatinib is rarely used in practice, as this would preclude second-line use of regorafenib. Therefore, 95% of patients continuing to subsequent systemic therapies following SIRT treatment are assumed to receive sorafenib, and 5% are assumed to receive lenvatinib.

As a number of chemotherapeutic/systemic agents administered to patients following sorafenib in the SARAH trial19 have now been displaced in practice by regorafenib, or are otherwise no longer in use, the AG model assumes that the proportion of those who received systemic therapies after sorafenib in the trial (12.04%) would receive regorafenib in UK practice. A small proportion (3.47%; i.e. 12.04% of 28.8%) of SIRT patients also receive regorafenib following second-line sorafenib treatment. Duration of therapy and dose intensity of each of the three systemic agents modelled is assumed to be the same as first-line treatment, whereas regorafenib is assumed to have the same time on treatment as sorafenib (122.95 days), with a mean daily dose of 160 mg (RESORCE trial). 101

Scenario 1: efficacy data from SARAH only

The first scenario analysis explores the effect of using only data from the European SARAH trial19 to inform efficacy estimates for SIRT and sorafenib, on the basis that these might better represent the patient population and clinical practice in the UK. Deterministic and probabilistic results are presented in Table 32. The probabilistic results are based on 5000 model iterations. As with the AG base case, each SIRT is associated with almost the same number of life-years and QALYs; however, this scenario predicts lower OS (and thus life-years/QALYs) than in the base case, which makes SIR-Spheres marginally cheaper than lenvatinib.

Scenario 2: low tumour burden/albumin–bilirubin 1 subgroup (SARAH)

This scenario explores the use of the company’s preferred post hoc grouping of patients from the SARAH trial19 as the source of efficacy data for SIRT and sorafenib. Further changes from the AG base case are the use of the higher low tumour burden/ALBI 1 subgroup utilities from the SARAH trial,19 and the significantly lower proportion of patients who receive work-up but not SIRT (8.1% vs. 18.6%). Note that although Sirtex used a proportion of 2.9% for work-up failures in this population, it was unclear how this figure was reached. Increasing the number of work-up failures, however, increases the cost-effectiveness of SIRT.

This scenario predicts the cost-effectiveness of an optimised decision in which only patients who have a tumour burden of ≤ 25% and a preserved liver function would be eligible to receive SIRT. As there is no equivalent evidence available for lenvatinib, this scenario assumes that the HR between sorafenib and lenvatinib remains the same as in the base-case population.

Table 33 shows that although the systemic therapies were less costly than SIRT in this scenario, SIR-Spheres generated an additional 0.139 QALYs compared with lenvatinib and 0.117 compared with sorafenib in the probabilistic model. This resulted in fully incremental ICERs of £20,926 per QALY gained for TheraSphere compared with lenvatinib, and £119,562 for SIR-Spheres compared with TheraSphere. However, the two technologies were distinguished only by the additional fluoroscopy cost associated with the SIR-Spheres procedure, resulting in very similar NMB at a £30,000 threshold. This is notably the only scenario in which TheraSphere and SIR-Spheres have a positive incremental NMB versus lenvatinib at a WTP threshold of £30,000 (excluding scenario 4). This is illustrated by the CEAC in Figure 20, which shows lenvatinib to have the highest likelihood of being cost-effective up to a WTP threshold of approximately £27,000, at which point it is surpassed by TheraSphere and by SIR-Spheres at a WTP threshold of ≥ £32,000.

FIGURE 20.

Cost-effectiveness acceptability curve for AG scenario 2: low tumour burden/ALBI 1 subgroup.

Results including the confidential PAS discounts for sorafenib, lenvatinib, regorafenib and QuiremSpheres can be found in Appendix 17.

Scenario 3: no macroscopic vascular invasion (SARAH)

This scenario limits the patient population to only those who had no MVI, referred to elsewhere as PVI, at baseline. These patients may be expected to benefit more from SIRT owing to a more favourable positioning and spread of their tumour, and were thus defined as a subgroup of interest in NICE’s scope. As there is no equivalent evidence for lenvatinib, this scenario assumes that the HR between sorafenib and lenvatinib remains the same as in the base-case population.

The probabilistic analysis in Table 34 found all three SIRTs to be dominated by lenvatinib, with a significantly lower NMB than either systemic therapy. Notably, the gap in QALYs produced by SIRT versus sorafenib widened in this analysis versus the base case, implying a reduced benefit of SIRT in this population.

Scenario 4: TheraSphere hazard ratio from the Biederman et al. and Van Der Gucht et al. network meta-analysis scenario

The results presented in Table 35 use the HR derived from the AG’s NMA scenario, which included the low-quality retrospective studies by Biederman et al. 39 and Van Der Gucht et al. 40 The patient population in Biederman et al. 39 was particularly mismatched with the others included in this analysis, as it included only patients with MVI, which appeared to have a substantial impact on the treatment effect associated with TheraSphere.

A HR of 0.46 versus SIR-Spheres was applied for both OS and PFS outcomes for TheraSphere. Based on the probabilistic analysis (5000 iterations), TheraSphere is expected to generate an additional 0.507 QALYs compared with lenvatinib, at an additional cost of £4068, producing an ICER of £8017 per QALY gained, and a NMB of £11,413. TheraSphere was associated with higher costs than SIR-Spheres owing to the increased disease management costs associated with lower mortality, but it also produced an additional 0.566 QALYs, yielding an ICER of £6060 per QALY gained.

Further scenario analyses

Table 36 presents a number of other scenarios on the AG base case that explore the impact of alternative assumptions, including sources of utilities, downstaging to curative therapy, resource use and survival models.

Scenarios 6 and 10 include the possibility for downstaging; in these scenarios, the distribution of the three liver-targeted treatments was derived from the SARAH trial. 19 Patients who received TACE or radiation therapy were excluded as these would not be permitted options in this population in UK practice. Liver transplant was undergone by 1.09% of SIRT patients and 0.46% of sorafenib patients; 1.63% of SIRT patients and 0% of sorafenib patients underwent liver resection, and 3.26% of SIRT patients and 0.92% of sorafenib patients received ablation therapy.

Only the deterministic results are produced for these analyses.

Table 37 presents the results of the base-case and selected scenario analyses in terms of their effect on the NMB ranking of the five technologies at list price. This shows lenvatinib to be consistently ranked first in terms of incremental NMB, except in those scenarios that use more favourable assumptions in favour of SIRT. As SIRT produces QALYs above the WTP threshold, increasing the proportion of patients who fail work-up (scenario 17) and do not go on to receive SIRT increases its cost-effectiveness, as overall costs are reduced and the more cost-effective QALYs produced on BSC and sorafenib are up-weighted.