Introduction
An economic evaluation was conducted to assess the cost-effectiveness of whole breast RT with 26 Gy delivered in 5 fractions over 1 week compared to the international standard 3-week schedule of 40 Gy in 15 fractions. The target population was UK adults who have undergone breast-conserving surgery or mastectomy for early breast cancer (stage I/II/IIIa).
A decision analytic model was constructed to estimate the costs and benefits of each option over a lifetime time horizon. The analysis followed the reference case outlined by the NICE in which health effects and costs which fall on the NHS were considered.68 The economic analysis excludes the trial arm schedule of 27 Gy delivered in 5 fractions over 1 week as this was associated with higher normal tissue effect risk compared to 40 Gy and 26 Gy, and is not considered a relevant comparator for future clinical practice.
Methods
Lifetime costs and health outcomes were estimated for each treatment using a Markov model. A discount rate of 3.5% per year was used for costs and health effects. Costs were set to a base year of 2019, adjusting for inflation.69 Quality-adjusted life-years (QALYs) were used to capture health impacts.
R Statistical programming language and was used to build the model and coding principle from the DARTH modelling group were followed.70,71
Model structure and patient population
The target population was UK adults who have undergone breast-conserving surgery or mastectomy for early breast cancer (stage I/II/IIIa). The patient population for the model matched the inclusion criteria for FAST-Forward as described in Chapter 2. This included individuals with tumour grades 1–3, ER positivity and negativity, HER-2 positivity and negativity and those with or without regional lymph node metastasis.
A Markov model was built to estimate the total QALYs and costs for each treatment alternative. The structure of the economic model is illustrated in Figure 16. Circular arrows indicate the possibility of staying in the same state for multiple cycles, and linear arrows represent the possible transitions between states. The model structure was developed in discussion with clinical experts. A single health state was used to capture locoregional relapse, which includes local relapse and regional relapse as these relapses have similar prognostic and cost implications. The Markov model cycle length was 1 year as outcomes occur over a long period. A half-cycle correction was applied.
FIGURE 16 Markov model structure of early breast cancer. (Figure not reproduced here; see the original article.)
Patients entered the model following local tumour excision and received RT in either 15 fractions or 5 fractions. The RT schedules differed in terms of resource use and the likelihood of inducing acute skin reactions. Following RT, patients began in the ‘alive and disease-free’ health state and over time were at risk of locoregional relapse, distant (metastatic) relapse or death. Radiotherapy modality determined how patients move through states over time.
Radiotherapy treatment period
This section concerns costs of delivering RT and the costs of managing acute side effects. Patients were assumed to receive treatment as specified in the FAST-Forward protocol.2 Radiotherapy costs were applied at model entry and are calculated as shown in Table 19. As RT delivery resource activity was not recorded in the trials, resource use was informed by expert opinion. A proportion of patients are instructed to follow breath hold techniques while receiving RT in order to protect the heart. This extends the duration of the RT appointment, and we estimated the consequent increase in cost.
| Input | Value | Source |
|---|---|---|
| Radiotherapy delivery | ||
| Radiotherapy planning cost | £315 | 72 |
| Cost of delivering one fraction of RT | £124 | 72 |
| Percentage receiving breath hold with whole breast RT | 25% | Expert opinion |
| Increase in fraction costs associated with breath hold | 30% | Expert opinion,73,74 |
| Treatment | Formula: (unit cost) × (units per patient) × (proportion of patients) | Cost |
| Whole breast 15 fractions (WB15F) | ||
| Planning cost | £315.00 | |
| RT without breath hold | £124 × 15 × 75% | £1395.00 |
| RT with breath hold | (£124 × 1.3) × 15 × 25% | +£604.50 |
| Total | £2314.50 | |
| Whole breast 5 fractions (WB5F) | ||
| Planning cost | £315.00 | |
| RT without breath hold | £124 × 5 × 75% | £465.00 |
| RT with breath hold | (£124 × 1.3) × 5 × 25% | +£201.50 |
| Total | £981.50 | |
We included costs relating to treating acute adverse skin reactions during treatment. These were assumed to be one-off costs which depend on the severity of AE as measured by worst RTOG score observed during treatment.3 Preference-based measures of health-related quality of life (HRQoL) were collected at baseline and at 3 months, but not during receipt of RT when acute skin reactions would be present. Literature review did not identify any studies that linked HRQoL measures to RTOG scores. Quality of life was assumed to be the same across 5 and 15 fractions during the treatment period due to an absence of preference-based quality of life data for this period. This assumption was viewed to be conservative towards the 26 Gy schedule given the numerically lower number of events compared to 40 Gy.
Post radiotherapy period
Time to locoregional relapse, distant relapse and some maintenance costs and HRQoL parameters were estimated from FAST-Forward individual patient data. Stata 16 was used to carry out all individual patient data analysis.75 The remaining model parameters were informed by searching the wider literature. All model inputs for this period are presented in Table 20.
| Input | Mean | Standard error | Distribution | Source |
|---|---|---|---|---|
| Population characteristics | ||||
| Average age at beginning of model | ||||
| Subgroup 1 | 63 years | - | - | FF |
| Subgroup 2 | 60 years | - | - | FF |
| Acute side effects during treatment | ||||
| Costs of treating adverse skin reactions for states: RTOG0, RTOG1, RTOG2a, RTOG2b, RTOG3 | £0, £0, £132, £132, £136 | - | - | Expert opinion90 |
| Probability of RTOG states: RTOG0, RTOG1, RTOG2a, RTOG2b, RTOG3 | ||||
| 15 fractions RT | 0%, 32%, 27.5%, 27.5%, 14% | - | Dirichlet | 3 |
| 5 fractions RT | 6%, 62%, 13.5%, 13.5%, 6% | - | Dirichlet | 3 |
| Transition rate between states | ||||
| Alive disease free to locoregional relapse: 15 fractions | ||||
| Subgroup 1 | 0.0038 | 0.0014 | Exponential | FF |
| Subgroup 2 | 0.0073 | 0.0014 | Exponential | FF |
| Alive disease free to distant relapse: 15 fractions | ||||
| Subgroup 1 | 0.0024 | 0.0007 | Exponential | FF |
| Subgroup 2 | 0.0128 | 0.0018 | Exponential | FF |
| Alive and disease free to death | Variable – dependent on age | - | - | 78 |
| Locoregional relapse to distant relapse | 0.0515 | 0.0045 | Exponential | 80 |
| Locoregional relapse to death | Variable – dependent on age | - | - | 78 |
| Distant relapse to death | ||||
| Subgroup 1 | 0.2196 | 0.5102 | Exponential | 79 |
| Subgroup 2 | 0.21 | 0.5102 | ||
| Relative treatment effect estimates | ||||
| Alive disease free to locoregional relapse: 5 fractions vs. 15 fractions | 0.66 | 0.167 | Log-normal | 23 |
| Healthcare costs | ||||
| Alive and disease free (annual) | ||||
| Subgroup 1 | £1216 | £82 | Gamma | FF |
| Subgroup 2 | £1412 | £68 | Gamma | FF |
| Additional costs of 1st year alive and disease free | £402 | £64 | Gamma | FF |
| 1st year of locoregional relapse | ||||
| Treatment costs | £4241 | ±20% | Gamma | 82–84,86 |
| Supportive care costs | £2995 | ±20% | Gamma | 50,86 |
| Beyond 1st year locoregional relapse (annual) | £2139 | ±20% | Gamma | 50,86 |
| Distant relapse (annual) | £13,426 | ±20% | Gamma | 85,86 |
| Health-related quality of life | ||||
| Alive and disease free | ||||
| Subgroup 1 | 0.7880 | 0.0130 | Gamma | FF |
| Subgroup 2 | 0.7730 | 0.0060 | Gamma | FF |
| Locoregional relapse | ||||
| Subgroup 1 | 0.7880 | 0.0130 | Gamma | FF |
| Subgroup 2 | 0.7730 | 0.0060 | Gamma | FF |
| Decrement for distant relapse relative to alive and disease free | 0.3030 | 0.1550 | Gamma | 88 |
| Quality of life decrement with age | Variable – dependent on age | - | - | 89 |
Transition probabilities
Transition probabilities estimated from FAST-Forward
The rate of locoregional relapse and distant relapse outcomes were estimated using separate models. This is the approach used in the clinical analyses in which separate Cox models are fit to locoregional and distant relapse (see Chapter 5). To choose an appropriate statistical model, the algorithm outlined in DSU guidance (2013) was used.76 It was found that an exponential survival model fit the data best for both locoregional relapse and distant relapse. Data from patients receiving 40 Gy and 26 Gy arm was used to fit the model (n = 2723) with a dummy to indicate the treatment schedule.
In the trial 45% would be considered at lower than average risk of local relapse and eligible for partial breast RT according to Royal College of Radiologists guideline.77 Our economic model estimates outcomes for the combined FAST-Forward population but allows for a separate analysis of low and high-risk populations.12,77 A subgroup dummy was included in the survival analysis to reflect differences in baseline risk of local relapse based on patient characteristics. The average ages in subgroups 1 and 2 were 63 and 60, respectively.
Definition of low- and high-risk subgroups
Subgroup 1: Lower risk of local relapse (eligible for partial breast RT under RCR guidelines). All of the following criteria: 50+ years old, and tumour grade 1–2, and tumour size ≤ 3 cm, and ER positive, and HER-2 negative, and with no regional lymph node metastasis (N0).
Subgroup 2: Higher risk of local relapse (not eligible for partial breast RT). Any of the following criteria: < 50 years old, or tumour grade 3, or tumour size > 3 cm or ER negative, or HER-2 positive or with regional lymph node metastasis (N1 or greater).
In the dataset, some individuals were recorded as having a distant relapse and locoregional relapse at the same time point. As a sensitivity analysis, we estimated the time to first event (either distant or locoregional), censoring individuals that experience more than one type of relapse (each participant only contributes their 1st event). For this we used a log-normal survival model which also fit the data well.
Estimated parameters for all survival analyses are reported in Appendix 2, Tables 44–51.
Transition probabilities for the remaining model states
The assumption was made that individuals who have not experienced a distant relapse have the same rate of all-cause mortality as the age- and sex-matched general population.78 Patients who developed distant relapse were at increased risk of death. This risk was based on a French study of metastatic breast cancer. In the base case, mortality with distant relapse was adjusted for age and based on the hormone receptor-positive and HER-2 negative molecular subtype. This was chosen as it is the most common subtype in FF. The impact of using alternative molecular subtypes was investigated in sensitivity analyses.79 The assumption was made that increased mortality from breast cancer occurs by first experiencing distant relapse. Therefore, the UK rate of breast cancer mortality was removed from all-cause mortality to avoid double counting.12
The rate of transition from locoregional to distant relapse was taken from a Dutch study of breast cancer.80 It was assumed that RT modality did not affect the rate of transition from locoregional to distant relapse or the mortality risk with distant relapse.
Treatment effects
To model the rates of locoregional relapse for 5 fractions relative to 15 fractions, we applied the HR for locoregional relapse estimated in FF to the baseline rate of locoregional relapse estimated in the survival model. For both 5 fractions and 15 fractions, a common rate of transition from alive and disease free to distant relapse was assumed. This assumption was based on the clinical rationale that RT is a local treatment and therefore its causal impact on developing distant relapse (if any) can only occur through reducing locoregional relapses. The impact of this assumption was explored in a sensitivity analysis. For the base case, treatment effects were assumed to persist over time.
Costs
Costs for the alive and disease-free state estimated from FAST-Forward
The resource use questionnaire collected in FAST-Forward covered activities related and unrelated to breast cancer such as general practitioner costs, nursing costs and hospitalisations. To construct per-patient costs, unit costs were applied to resource use, see Appendix 2, Table 52.
A panel of cost data was constructed for each data collection time period. Individuals only contributed data if there were classified as being in the alive and disease-free state. Summary statistics are shown in Appendix 2, Table 53.
A GEE model was fit to the 3-month cost data.75 This model had an exchangeable within-individual correlation structure. A gamma family distribution was chosen to reflect the non-negativity and skewed distribution of cost data. A log link was chosen given the typical structure of cost data. The GEE model utilises all available observations of the dependent variable. This is equivalent to a complete case analysis (CCA). CCA was chosen as the degree of missing data in the study was low, < 25% at any time point. This implies a missing completely at random assumption (MCAR) that is missingness is independent of both observed and unobserved values.81 When fitting the models, baseline costs were discarded as it is the post-treatment costs that are to be estimated.
Plotting the data revealed a large drop between costs in the first 6 months after treatment and the costs from 6 months onward. This discontinuity was modelled by defining a binary covariate indicating 6 months pre- and post-treatment. A dummy for high versus low risk was also included in the analysis. Age and treatment were found not to be statistically significant and so were not included in the model.
The following values were estimated after scaling costs up from average costs over 3 months to 12 months:
- Yearly costs in subgroup 1: low risk = £1216, with SE = £82.
- Yearly costs in subgroup 2: high risk = £1412, with SE = £68.
- Additional costs associated with first year after treatment = £402 with SE £64.
Costs for the remaining model states
These costs were sourced from the wider literature as there were insufficient observations to estimate them from FAST-Forward. Locoregional relapse was associated with one-off mastectomy costs in addition to supportive care costs in the first year of the event.50,82–84 Following this first year supportive care was assumed to consist of one GP visit and one mammogram per year.50 A UK study was used to estimate supportive care and treatment costs for distant relapse.85
To make them consistent with the inclusion of related and unrelated costs in the FAST-Forward resource use questionnaire, costs unrelated to breast cancer were added to the breast cancer costs for locoregional and distant relapse health states.86
Health-related quality of life
Health-related quality of life for the alive and disease-free state estimated from FAST-Forward
Quality of life (QoL) was captured in FAST-Forward by the EuroQol EQ-5D 5 level questionnaire. NICE’s current advice is to map 5L scores to 3L using the method described by van Hout et al.87
Data was collected in the PRO sub-study (n = 1179) at baseline (before randomisation or RT), 3, 6 months post treatment and 1, 2 and 5 years post randomisation.
A panel of QoL data was constructed for each data collection time period. Individuals only contributed data if there were classified as being in the alive and disease-free state. Summary statistics are shown in Appendix 2, Table 54.
A generalised liner model (GLM) based on the wave of data closest to treatment was used for simplicity as time was found to have no statistical impact on results. A gamma distribution for disutility was used as this fits the range utility data. An identity link was chosen.
A complete case analysis (CCA) was chosen as the degree of missing data in the study was low at < 25% at any time point. This implies a missing completely at random assumption (MCAR) that is missingness is independent of both observed and unobserved values (Faria et al.).81
Using the full sample (6143 observations on 1044 patients), we estimated a GLM model. This model estimated disutilities (1 – EQ-5D score). It included a dummy for the alive and disease-free state, for high versus low risk, a time dummy equal to one for all post-treatment time points, all possible interactions between the previous dummies and a dummy for the treatment arm. The average EQ-5D score for patients in the alive and disease-free state in subgroup 1 was 0.788 (SE = 0.013, p-value < 0.001), for subgroup 2 this was 0.773 (SE = 0.006, p-value < 0.001).
Health-related quality of life for the remaining model states
Health-related quality of life post locoregional relapse was assumed to be the same as for the alive and disease-free state. The decrement in HRQoL with distant relapse was taken from a previous RT decision model.88 A Health Survey for England study was used to model the decline in HRQoL with age.89
Uncertainty analyses
The distributions described in Table 20 were applied to each of the parameters to represent uncertainty in the model. A probabilistic sensitivity analysis (PSA) was used to reflect the impact of this joint uncertainty on outcomes. This involved taking 10,000 draws from the distributions and calculating model outcomes for each iteration. One-way sensitivity analysis was used to explore the sensitivity of results to one-at-a-time changes in individual inputs and assumptions, see Table 21.
| Scenario | FAST-Forward population | |
|---|---|---|
| Whole breast 15 fractions (%) | Whole breast 5 fractions (%) | |
| ICER (probability treatment ICER < Threshold) | ||
| Base case (£15,000/QALY) | Dominated (0.1) | £0 (99.9) |
| Threshold £20,000/QALY | Dominated (0.2) | £0 (99.8) |
| Threshold £30,000/QALY | Dominated (0.4) | £0 (99.6) |
| Distant relapse HR estimated in FAST-Forward | £0 (66.2) | £6271 (33.8) |
| All treatment effects maintained for 10 years | Dominated (0.1) | £0 (99.9) |
| All treatment effects maintained for 5 years | Dominated (0.1) | £0 (99.9) |
| Mortality rate following distant relapse based on HER-2+ population79 | Dominated (0.1) | £0 (99.9) |
| Mortality rate following distant relapse based on TNBC (HR− and HER-2−) population79 | Dominated (0.1) | £0 (99.9) |
| Distant relapse costs reduced to £8934 per year94 | Dominated (0.1) | £0 (99.9) |
| Distant relapse costs increased to £16,111 per year (20% increase) | Dominated (0.1) | £0 (99.9) |
| Disutility resulting from distant relapse reduced to 0.2695 | Dominated (0.1) | £0 (99.9) |
| Disutility resulting from distant relapse increased to 0.3636 (20% increase) | Dominated (0.1) | £0 (99.9) |
| Use of breath hold assumed to not increase cost to deliver one fraction of RT (£124) | Dominated (0.1) | £0 (99.9) |
| Use of breath hold assumed to double the cost to deliver one fraction of RT (£248) | Dominated (0.1) | £0 (99.9) |
| Rate of adverse skin reactions set equal across all treatment options | Dominated (0.1) | £0 (99.9) |
| Health-related quality of life weight during RT set to zero | Dominated (0.1) | £0 (99.9) |
| Log-normal survival model in which each participant only contributes their 1st event | Dominated (12) | £0 (88) |
Cost-effectiveness analysis
Discounted costs and QALYs were calculated for each treatment option taking account of outcomes over the patient lifetime. Incremental cost-effectiveness ratios (ICERs) were calculated with whole breast 15 fractions as the comparator treatment. The ICER was calculated as the mean additional costs divided by the mean additional health gained for 5 fractions compared to 15 fractions. The ICER was compared to a cost-effectiveness threshold of £15,000/QALY. This was chosen as it is the value used by the Department of Health in the UK to represent the marginal rate at which NHS activities generate QALYs.91,92 Sensitivity analysis used the threshold values commonly used in NICE decision-making: £20,000–30,000/QALY.68
Validation
Model validation was carried out using the TECH-VER checklist.93
Subgroup analyses
FAST-Forward contained a combination of individuals who were both high and low risk. The model allowed for these subgroups to be analysed separately or as a combined population. The primary analysis reports results for the combined population. Secondary analysis reports results for each population separately.
Results
For the base case analysis, mean costs and QALYs for 15 fractions were £31,640 and 11.08 QALYs, respectively. For 5 fractions these were £29,638 and 11.12 QALYs. Therefore, 5 fractions therapy was expected to dominate with expected cost savings of £2002 (95% CI £1245 to £2804) and higher expected QALYs: 0.04 (95% CI −0.01 to 0.09). Across simulations there was a 99.9% chance that 5 fractions either dominated WB15F or had an ICER below £15,000/QALY. The probabilistic results are shown in the incremental cost-effectiveness plane in Figure 17. The dashed line represents the cost-effectiveness threshold of £15,000 per QALY. The points below this line represent simulations in which 5 fractions is cost-effective relative to 15 fractions. This diamond indicates mean cost savings and expected gains in health.
FIGURE 17 Base case cost-effectiveness plane comparing 5 fractions to 15 fractions. (Figure not reproduced here; see the original article.)
Subgroups
For the low-risk population (subgroup 1), 5 fraction therapy was expected to dominate with expected cost savings of £1881 (95% CI £1252 to £2648) and higher expected QALYs: 0.03 (95% CI −0.01 to 0.07). Across simulations there was a 99.9% chance that 5 fractions either dominated WB15F or had an ICER below £15,000/QALY.
For the high-risk population (subgroup 2), 5 fractions therapy was expected to dominate with expected cost savings of £2102 (95% CI £1230 to £3093) and higher expected QALYs: 0.05 (95% CI −0.01 to 0.11). Across simulations there was a 99.9% chance that 5 fractions either dominated WB15F or had an ICER below £15,000/QALY.
Sensitivity analysis
Based on expected outcomes, 5 fractions dominated all other options except when using the distant relapse HR results estimated in the trial (see Table 21). In this one scenario, 5 fractions compared with 15 fractions was expected to be less expensive with incremental costs of −£908 (95% CI −£2689 to £975) but less effective, with −0.14 incremental QALYs (95% CI −0.43 to 0.12). This placed the expected ICER in the southwest quadrant. With a threshold of £15,000/QALY 5 fractions was expected to have 33.8% chance of being cost-effective in this scenario.
Discussion
Across a range of scenarios, 5 fraction RT were expected to be cost-effective at a threshold of £15,000/QALY. Five fractions was expected to provide more QALYs and have lower costs compared to 15 fractions. The expected cost savings were primarily due to reducing the number of fractions of RT. The improvement in QALYs was driven by the modest expected reduction in locoregional relapses. Figure 17 illustrates how 5 fractions is associated with lower costs and greater QALYs. The figure also illustrates that the spread of points for 5 fractions lies almost completely to the south east of the £15,000/QALY cost-effectiveness threshold. This indicates that the cost-effectiveness of 5 fractions was associated only with minimal parametric uncertainty.
Resource savings from reduced fractionation may enable the same number of patients to be treated with lower linear accelerator capacity therefore freeing capacity to treat breast and other cancers. The increased capacity could also be used to introduce the many improvements in the quality of therapy, such as cardiac-sparing breath hold if not fully implemented. Further, the benefits to patients would immediately be realised in terms of reduced burden of treatment and hospital footfall, and this may have added benefits such as reduced exposure to COVID-19.96
The sensitivity analyses reported in Table 21 illustrated that these overall conclusions were robust to alternative inputs and assumptions. Across all scenarios the 5 fractions regimens remained the least costly alternative. Using different molecular subtype populations to model mortality following distant relapse did not impact results.79 This was because in this scenario relative rates of distant relapse were assumed common across arms. In the scenario in which relative rates of distant relapse were calculated using the HR estimated in FAST-Forward, 5 fractions was associated with fewer expected QALYs. With a cost-effectiveness threshold of £15,000/QALY, 5 fractions was not expected to be the cost-effective alternative. The relatively lower QALYs associated with 5 fractions in these scenarios was due to the HR for distant relapse for 5 versus 15 fractions estimated in FAST-Forward. This estimate indicated a higher rate of distant relapse with 5 fractions, HR = 1.27 (0.90 to 1.80) which in absolute terms is equivalent to an increase of 1% (−0.4% to 2.9%). Though this was not statistically significant (p = 0.17) distant relapse had a large impact on both morbidity and mortality in the economic model and so drove differences in expected QALYs between treatment options.
The key assumption in the base case analysis is that there was a common pattern of transition from alive and disease free to distant relapse across treatment options. This assumption was based on the clinical argument that RT is a local treatment and therefore has local effects. One alternative approach to capture treatment effects in the model would be to estimate the rate of any relapse (distant or local or regional). This approach could be used to estimate baseline and relative effects across RT modalities. This alternative approach was not carried out here so that the economic analysis followed the analysis in the published trial.23 Additionally, estimating the difference in any relapse between arms would imply a common treatment effect on local, distant and regional relapses which may not be clinically plausible. Even if this approach was used, it would remain necessary to separate out the different relapse types as they have different costs and health consequences. A fully comprehensive approach may require a multistate modelling framework and thus a different approach to both data collection and modelling.97,98 This approach would need to take account of recognised issues with effect identification.99,100 Further research is required in this area.
Other limitations of the analysis are described here. The HRQoL impact of acute skin reactions was not captured due to lack of EQ-5D data during the treatment period, only treatment costs were included. Including HRQoL impacts into the analysis would likely improve the relative benefits of 5 fraction therapy as this was observed to have fewer severe acute adverse reactions than 15 fractions in Brunt et al.3
To our knowledge, no other studies have examined the cost-effectiveness of 5 fraction RT in a UK context. For the US system, Deshmukh et al. compared whole breast RT delivered over 5–7 weeks to RT delivered over 3–4 weeks and found that reducing fractionation (and therefore duration of RT) dominated higher fractionation.101 This is in line with our base case results.
Based on the results presented, 5 fraction RT offers potentially significant benefits to the UK health system compared to 15 fractions.