Gene expression profiling and expanded immunohistochemistry tests to guide the use of adjuvant chemotherapy in breast cancer management: a systematic review and cost-effectiveness analysis

Clinical

Economic

Two economic studies were identified. Both studies compared the use of OncotypeDX with Adjuvant! Online. These studies were non-UK studies and were not considered generalisable to the UK setting. The economic evaluation submitted by Genomic Health estimated the incremental cost for treatment guided using OncotypeDX to be £6232 per QALY gained compared with current clinical practice in the UK, although a number of limitations with regard to the analysis were highlighted.

A de novo economic model was built by the EAG and estimated the cost per QALY gained to be £29,502 compared with current clinical practice, assuming that the test was offered to all woman with ER+, LN−, HER2– early breast cancer, under our base-case assumptions. This analysis assumed OncotypeDX to be predictive of the benefit of chemotherapy, based on evidence from the Paik et al. study, although weaknesses relating to this study are highlighted. (Paik S, Tang G, Shak S, Kim C, Baker J, Kim W, et al. Gene expression and benefit of chemotherapy in women with node-negative, estrogen receptor-positive breast cancer. J Clin Oncol 2006;24:3726–34.) A subgroup analysis was performed and showed that the incremental cost-effectiveness ratio (ICER) for OncotypeDX compared with current clinical practice was reduced to £9774 per QALY gained if OncotypeDX was to be offered to women with a (NPI > 3.4) only. Compared with current clinical practice, OncotypeDX had a 12.4% (all women) and 91.6% (NPI > 3.4) probability of being considered cost-effective when using a threshold of £20,000 per QALY gained respectively, although the quality of the data in the model was considered relatively weak. Key areas of uncertainly relate to assumptions about the benefits of chemotherapy in terms of relative risk reduction by risk group, the risk of recurrence over time and the impact of the new test on decision-making. The ICER increased substantially and was greater than £20,000 per QALY gained for both analyses when assuming the same relative reduction in the risk of recurrence from chemotherapy for all patients irrespective of the OncotypeDX recurrence score classification, that is, assuming that the test is prognostic only.

Clinical

Economic

An analysis was carried out by the EAG to evaluate the use of MammaPrint in England and Wales but because of the limitations in the evidence available this was considered exploratory only and no base-case ICER was presented.

Clinical

The evidence base for PAM50 is still relatively immature. The current review identified two analytical validity studies (reported in abstract form only) comparing the PAM50 test with standard IHC measurements. Four studies evaluated the clinical validity of PAM50; two of these are as yet unpublished. No evidence on clinical utility was identified.

Economic

The EAG did not model treatment guided using PAM50 because of gaps in the evidence base.

Clinical

The current review identified three studies that provided data to support the use of the Mammostrat test as an independent prognostic tool for women with ER+, tamoxifen-treated breast cancer. Although the evidence base for the Mammostrat test is relatively immature, these studies included a large sample size, appeared to be of reasonable quality and provided data from a UK setting (one study). One study was identified for clinical utility but limitations were identified relating to this study.

Economic

The EAG conducted an exploratory analysis using the same model structure as for the OncotypeDX evaluation and unpublished data from a small sample from a non-UK population; however, because of the limitations in the evidence base, any conclusions drawn from this analysis are subject to significant uncertainty.

Clinical

No studies on analytical validity of the test were identified. The current review identified one study on the clinical validity of IHC4, which reports that the IHC4 score is a highly significant predictor of distant recurrence. This study was based on a large sample size and detailed the development of the test in one cohort and the external validation of the test in an independent cohort. The study also reported evidence comparing IHC4 with OncotypeDX. The review did not identify any published evidence on the clinical utility of IHC4 in terms of its impact on treatment decisions or its ability to predict chemotherapy benefit by risk group.

Economic

The EAG evaluated the cost-effectiveness of IHC4 in parallel with that of OncotypeDX as there was direct evidence between the two tests in a UK population from the same data source used to evaluate the cost-effectiveness of OncotypeDX. The IHC4 test was predicted to be dominant compared with current clinical practice in patients with ER+, LN−, HER2– early breast cancer, providing more QALYs at a lower cost. An incremental analysis was conducted comparing OncotypeDX, IHC4 and current clinical practice. When the treatment decision using OncotypeDX was compared with that using IHC4, the ICER for OncotypeDX increased to £64,111 per QALY gained if the tests were to be offered to all women and £31,125 if the tests were to be offered to women with a NPI > 3.4 only. IHC4 was predicted to remain dominant assuming the test to be prognostic only, that is, all women receiving chemotherapy derive the same relative benefit in terms of reduction in distant recurrences. However, because the evidence base for IHC4 is less developed than that for OncotypeDX, additional assumptions were required and the results are subject to greater uncertainty.

Clinical

Based on the limited available data identified for these tests, no firm conclusions can be drawn about their analytical validity, clinical validity (prognostic ability) and clinical utility. Further evidence on the prognostic and predictive ability of all of these tests is required.

Economic

No studies were identified in the systematic review of the economic literature. The EAG did not model treatment guided using these tests because of significant gaps in the evidence base.

OncotypeDX

Marchionni et al. 33 reported that OncotypeDX was furthest along the validation pathway, with strong retrospective evidence that it predicts distant metastasis and chemotherapy benefit to a clinically relevant extent over standard predictors in a well-defined clinical subgroup with clear treatment implications. A more detailed summary of the main results is provided in Appendix 4.

MammaPrint

The evidence reported by Marchionni et al. 33 for MammaPrint was more limited. A more detailed summary of the main results is provided in Appendix 5.

OncotypeDX

MammaPrint

OncotypeDX

• Analytical validity – there is limited evidence for the reproducibility of the tests in terms of reproducibility across different samples of the same block and across samples from different blocks. Centralisation was considered to be a current strength of OncotypeDX with regard to reproducibility.

• Clinical validity (prognostic ability of the tests) – there is fairly strong support for OncotypeDX over and above standard clinical predictors, but only in a well-defined population (ER+, LN−). Evidence is required to assess the stability of risk categories in other populations.

• Clinical utility – very few of the studies, particularly in isolation, provided compelling evidence of the test's clinical utility.

MammaPrint

• Analytical validity – there were limited data on variability and reproducibility, with a limited number of patients and a moderate number of replications.

• Clinical validity (prognostic ability of the tests) – evidence was based on retrospective data using clinically heterogeneous cohorts; evidence from RCTs is needed.

• Clinical utility – very limited evidence was available on clinical utility; robust evidence on the prediction of chemotherapy benefit is required.

Marchionni et al. 32 concluded (at the time of publication) that for both tests the relationship of predicted to observed risk in different populations needed further study, as did their incremental contribution, optimal implementation and relevance to patients on current therapies. Smartt34 concluded that the largest volume of evidence related to the OncotypeDX test.

Analytical validity

No new data examined analytical validity.

Clinical validity

Using 1231 tissue samples from the UK TransATAC (Arimidex, Tamoxifen, Alone or in Combination trial) trial, Dowsett et al. 79 assessed postmenopausal, hormone receptor-positive and majority LN− patients. They demonstrated that a 50-point increase in RS in all LN− patients (e.g. RS = 55 vs. RS = 5) was significantly associated with an increased risk of distant recurrence [hazard ratio (HR) 3.92, 95% CI 2.08 to 7.39; p < 0.001] when adjusted for the effects of tumour size, local grade (grade derived from case record forms), age and treatment. They also reported that, when local grade was replaced with central grade (assessed using the Elston and Ellis system) in multivariate analysis, adjusted RS was also significantly associated with risk of distant recurrence (HR 5.25, 95% CI 2.84 to 9.73; p < 0.001). RS was also significantly associated with TTDR in both node-negative (HR 5.25, 95% CI 2.84 to 9.73; p < 0.001) and node-positive patients (HR 3.47, 95% CI 1.64 to 7.38; p < 0.002). Correlation between RS-predicted distant recurrence and Adjuvant! Online-predicted recurrence was low but statistically significant by central grade (Spearman rank correlation = 0.23; p < 0.001) or local grade (Spearman rank correlation = 0.22; p < 0.001). Only approximately 5% of the variability in the estimates of recurrence using either of these scores was explained by the other. The authors concluded that the findings demonstrated that RS is an independent predictor of distant recurrence in LN− and LN+ hormone receptor-positive patients treated with anastrozole, adding value to estimates with standard clinicopathological features. As the patients were recruited as part of a large-scale trial this study benefits from a large sample size of UK-based patients and has a relatively long follow-up time (9 years).

Yorozuya et al. 88 reported a very small case–control study (10 cases, 30 control subjects) of ER+, LN− Japanese patients. The cases were those who had metastases after surgery; control subjects did not develop metastases. Significant differences were shown between the groups in terms of mean RS (cases: mean RS = 40.0, 95% CI 21.1 to 58.9; control subjects: mean RS = 17.8, 95% CI 13.8 to 21.9; p < 0.001). The study found significant differences between cases and control subjects in the proportions assigned to different risk categories [low: 3 (30%) cases vs. 19 (63%) control subjects; intermediate: 1 (10%) vs. 8 (27%); high: 6 (60%) vs. 3 (10%); p = 0.005]. Multivariate logistical regression analysis of age, ER score, PR score, RS, histological grade and lymphatic invasion compared with distant metastases showed that RS was not significant [RS ≥ 50 vs. RS < 50, p = 0.579, odds ratio (OR) 2.85, 95% CI 0.07 to 115.552] although the authors conclude that the OR indicates that it has value. The authors concluded that both histological grade and RS classification were effective in identifying women at risk of developing distant metastases after initial therapy for ER+, LN− stage I or IIA breast cancer. There are significant limitations in terms of the generalisation of the findings because of the extremely small sample size used in this study; furthermore, as the study was Japan based, generalisations to the UK setting are limited.

Toi et al. 87 examined the prognostic ability of OncotypeDX in 200 ER+, LN− Japanese patients. They showed that patients categorised as low risk had a significantly lower risk of distant recurrence than patients in the high-risk category (p < 0.001, log-rank test). No recurrences were identified in the intermediate recurrence group. Continuous RS was significantly associated with the risk of distant recurrence for a 50-point increase in RS (HR 6.20, 95% CI 2.27 to 17.0). In multivariate analyses the continuous RS maintains statistical significance when adjusting for age and clinical tumour size (HR 6.03, 95% CI 2.17 to 16.7). For risk of recurrence the HR was 3.38 (95% CI 1.32 to 8.69), for risk of recurrence or death the HR was 2.09 (95% CI 0.84 to 5.20) and for risk of death the HR was 2.67 (95% CI 0.93 to 7.62). The authors concluded that OncotypeDX has value in providing prognostic information in Asian populations with ER+, LN− breast cancer. This study had a small sample size and as it was conducted using Japanese patients generalisability to UK practice may be limited; however, the study does benefit from the fact that the tumour samples used were from patients who presented and were treated relatively recently (1992–8).

Cuzick et al. 84 reported data that aimed to assess how much of the information in the RS is contained in standard IHC markers (data from this report relating to the IHC4 test is presented in IHC4 test). Patients comprised a retrospective cohort from the TransATAC trial (multinational including the UK). The 1125 patients were mainly LN− and hormone receptor positive, and there were a total of 195 recurrences of which 145 were distant recurrences. In LN− women there were 101 recurrences of which 67 were distant recurrences. The authors reported the mean change in likelihood ratio chi-squared for the addition of GHI-RS (Genomic Health Recurrence Score) v to the classical score in the validation halves of 100 random splits of the data (higher values indicate more added prognostic information) for TTDR and time to recurrence (all recurrences). For TTDR the likelihood ratio chi-squared for all patients was 25.3 (95% CI 25.2 to 25.9) and for LN− patients was 20.9 (95% CI 20.7 to 21.6). For time to recurrence the LR-_x² for all patients was 25.6 (95% CI 25.2 to 25.9), and for LN− patients was 25.7 (95% CI 25.4 to 26.4). The authors report that the OncotypeDX RS adds prognostic information to traditional clinicopathological measures. This study has been rated as high quality and benefits from a large sample of patients.

Mamounas et al. 80 (and Tang et al. ,81 reported in the following section) undertook a retrospective analysis of ER+, LN− patients who had been recruited into two large US trials (NSABP B14 and B20). They showed a significant association between RS and the proportion of patients with locoregional recurrence at 10 years for 355 placebo-treated patients (NSABP B14), 895 tamoxifen-treated patients (NSABP B14 and B20) and 424 tamoxifen plus chemotherapy-treated patients (NSABP B20). Multivariate Cox regression analysis in the cohort of 895 tamoxifen-treated patients showed that RS was a significant predictor of locoregional recurrence (HR 2.16, 95% CI 1.26 to 3.68; p < 0.005). The authors concluded that a significant association exists between RS and risk for locoregional recurrence. This information has biologic consequences and potential clinical implications relative to locoregional therapy decisions for patients with LN− and ER+ breast cancer. These studies appeared to be of reasonable quality and, as the patients were recruited as part of two large-scale trials, the studies benefit from a large sample size with a long follow-up. However, across the two trials tumour samples were collected as long ago as 1982 until 1993. This means that there may be differences in diagnostic criteria applied and this may limit the generalisability of these findings to current practice.

Clinical utility

Analytical validity of OncotypeDX

In the earlier systematic reviews evidence exists on the technical and operational aspects of the test and on assay variability and reproducibility. Studies showed reasonable within-laboratory replicability.

Our findings indicated no new evidence.

Clinical validity (prognostic ability) of OncotypeDX

In earlier systematic reviews the evidence shows that RS was significantly correlated with DFS and OS. RS alone was shown to be a better predictor of distant recurrence at 10 years than traditional clinicopathological predictors. 42 Key gaps relate to the stability of risk categories in populations other then ER+, LN− patients.

Our findings indicate that further larger studies now exist which support the prognostic capability of OncotypeDX. In particular, a large UK study in 1231 postmenopausal women with hormone receptor-positive, LN− early breast cancer concluded that an increase in risk score was significantly associated with an increased risk of distant recurrence. 90 This study and the Mamounas et al. 80 study provide new evidence on the clinical validity of OncotypeDX, which employs cohorts of patients from large-scale RCTs and is rated as high quality. Furthermore, the evidence base has been extended to include the LN+ population83 and there are the beginnings of an evidence base for the validity of OncotypeDX in different populations such as in Japanese patients. 87,88

Clinical utility of OncotypeDX

In the earlier systematic reviews, evidence on clinical utility is limited. Paik et al. 49 demonstrated a significant benefit from the use of chemotherapy in the OncotypeDX high-risk group, although the review highlighted that the study may have been subject to bias as some patients in the validation data set were also in the training data set. Clinical experts indicated that more effective chemotherapy regimes are currently used in the UK. In total, > 44% of patients were aged < 50 years. The benefit of chemotherapy (reduction in distant recurrence) was greater in this population than in women aged > 50 years. The HR for the benefit of chemotherapy (reduction in distant recurrence) in women aged > 50 years compared with younger women was 2.02 (95% CI 0.75 to 5.47; p = 0.162).

Further supporting evidence was needed. Key gaps relate to the extent to which the test added to the management of patients and the proportion of patients who would benefit from the test. The role of the OncotypeDX test in guiding treatment of HER2-positive patients was unclear, as most of these patients were classified in the high-risk RS group in the initial trials. Prospective confirmation of the clinical utility of OncotypeDX was required.

Our findings indicate that there are no prospective studies reporting the impact of OncotypeDX on long-term outcomes such as OS. Four new studies76–78,82 presented further evidence on the impact of OncotypeDX on clinical decision-making. These indicate that the use of OncotypeDX leads to changes in decision-making for between 31.5% and 38% of patients. However, only one of these studies was UK based, and limitations in relation to study design were identified for this study. Specifically, these data were based on a small sample size (n = 106) and were derived from a conference poster,78 which was lacking the detail necessary to make judgements about the quality of the evidence. Two new studies (with three related citations81,83,86) provided evidence supporting the case that OncotypeDX predicts chemotherapy benefit. The Tang et al. 81,86 studies were based on ER+, LN− patients and Albain et al. 82 reported evidence for ER+, LN+ patients. These studies were based on trial data and the sample sizes were moderate in the case of Albain et al. 83 (n = 367) and large in the Tang et al. 81,86 analyses (n = 625–1319). These studies also had long follow-up times. Study quality was judged to be medium83,86 or high,81 although as Tang et al. 86 was a conference abstract we were unable to access the detail necessary to make adequate judgements about the quality of the evidence.

The first evidence relating to improvements in quality of life and reductions in patient anxiety as a result of using the test have been reported, although generalisations should be made with caution because of the small sample sizes employed. Further research in this area is required.

Key gaps in the evidence remain:

• Few of the studies were considered to be of high quality (n = 3). A number of studies in the current review were judged to provide medium-quality (although retrospective) evidence for OncotypeDX (n = 9). One of the most characteristic features of the studies was their heterogeneity. The studies varied considerably in their size, study design, patient populations and objectives. A large proportion of the OncotypeDX studies were small and retrospective. Many studies used old archived tumour samples and included the use of retrospective chart review to elicit treatment recommendations before and after OncotypeDX testing. There was a lack of standardised decision-making tools both within and between studies, and non-standardised methods of patient selection for OncotypeDX testing were used.

• Further direct evidence of the clinical utility of OncotypeDX is still required. This will be addressed by the ongoing TAILORx trial.

• The generalisability of the findings may be limited because of the small number of studies that were conducted in the UK setting and because a number of the studies were funded by the manufacturer, giving rise to possible conflicts of interest and publication bias.

Analytical validity

Our searches did not reveal any studies that examined analytical validity.

Clinical validity (prognostic ability)

Kok et al. 99 assessed whether or not analysing both MammaPrint score and hormone receptors provides superior prediction of outcome than hormone receptors alone in 272 Dutch patients. One group comprised LN+, ER+, tamoxifen-treated patients and a second group comprised LN−, ER+ patients who had received no adjuvant systemic treatment. Hormone receptors were evaluated using the St Gallen consensus recommendations102 (highly endocrine responsive: ER and PR ≥ 50%; incompletely endocrine responsive: ER and/or PR low or with either one absent). In patients treated with adjuvant tamoxifen (mainly LN+), both MammaPrint (adjusted for endocrine response categories, HR 2.78; 95% CI 1.30 to 5.94) and the endocrine response categories (adjusted for MammaPrint, HR 7.22; 95% CI 2.17 to 24.0) were associated with BCSS at 10 years. Also, in patients treated with tamoxifen for metastatic disease, combined analysis of MammaPrint and ER/PR revealed additional value (multivariate Cox regression; p = 0.013). In patients who did not receive tamoxifen, only MammaPrint was associated with outcome. The authors concluded that both methods provide independent information on outcome after tamoxifen for LN+ breast cancer. There are a number of limitations to this study: the second patient group comprised patients included in two previously reported evaluations, the overall sample size was small and tumour samples had been collected over a number of years (1982–97), which has implications for changes in diagnosis, treatment and standards of care. The study did benefit from an adequate follow-up time of 10 years.

Ishitobi et al. 96 examined risk classification using MammaPrint and disease outcome for 102 LN−, majority ER+, relatively young breast cancer patients in Japan. The results relating to clinical validity are presented here and the results relating to clinical utility are presented in the relevant section below. Among the 102 patients, 20 (20%) were classified as low risk and 82 (80%) as high risk. The authors reported that the probability of DMFS at 5 years was 100% for the low-risk group and 94% for the high-risk group. They did not report a HR. The NPV for time to distant metastasis was high (100%, 20/20), whereas the PPV was quite low (9.8%, 8/82). The NPV indicates the proportion of patients classified as low risk who were correctly classified using MammaPrint, whereas the small PPV indicates that many of the cases classified as high risk were incorrectly classified. The authors concluded that the 70-gene prognosis signature accurately identified Japanese breast cancer patients as being at low risk of developing recurrences, as 100% of the individuals in the low-risk group remained metastasis free for the duration of the observation period. The authors suggest that the low number of individuals in the low-risk group is consistent with previous findings on patient groups of ≤ 54 years. However, these low numbers make any generalisations of the findings limited. The young patient population may also explain why the probability of DMFS was also very high for the high-risk group, together with the fact that this was assessed at only 5 years, given that the majority of distant recurrences and deaths from breast cancer occur > 5 years after diagnosis. This study employed a very small sample size and, furthermore, as this study was performed in a Japanese population any generalisations to the UK population are significantly limited.

In a Netherlands-based study, Bueno-de-Mesquita et al. 93 assessed 123 LN−, majority ER+ patients who had been assigned MammaPrint risk categories. They reported risk classification and probability of disease outcome (time from surgery to distant metastasis and OS). OS probability was 97% (±2%) for good and 82% (±5%) for poor prognosis signature patients (p-value not reported) with an estimated HR of 3.4 (95% CI 1.2 to 9.6; p = 0.021). The probability of remaining free of distant metastasis (as first event) was 98% (±2%) for good and 78% (±6%) for poor prognosis signature patients (p-value not reported) with an estimated HR of 5.7 (95% CI 1.6 to 20; p = 0.007). In multivariate analysis, the authors demonstrated that MammaPrint was a strong independent prognostic factor, outperforming the clinicopathological risk indexes. They concluded that the 70-gene prognosis signature is also an independent prognostic factor in LN− breast cancer patients for women diagnosed in recent years. Again, as this study used a small sample size and the follow-up assessment was limited to 5 years, generalisations of the findings are limited. This study also reported reclassification findings, which are detailed in the relevant section below.

Mook et al. 75 examined 148 LN− and majority ER+, specifically postmenopausal patients. The study, conducted in the Netherlands, investigated disease outcome (DMFS and BCSS at 5 years), and prediction of early breast cancer-specific death (BCSD) using MammaPrint risk categories. The authors also assessed reclassification and these findings will be presented below. DMFS at 5 years was 93% in the low-risk group and 72% in the high-risk group (p = 0.07) with an associated HR of 4.6 (95% CI 1.8 to 12.0; p = 0.001). At 10 years it was 80% in the low-risk group and 67% in the high-risk group (HR not reported, p-value not reported). Over the entire follow-up period the HR was 1.8 (95% CI 0.9 to 3.5; p = 0.07). BCSS at 5 years was 99% in the low-risk group and 80% in the high-risk group (p = 0.036) with an associated HR of 19.1 (95% CI 2.5 to 148, p = 0.005). At 10 years it was 90% for the low-risk group and 69% for the high-risk group (HR not reported, p-value not reported). Over the entire follow-up period the HR was 2.0 (95% CI 1.0 to 4.0; p = 0.04). In terms of the prediction of early BCSD, the HR for BCSS at 5 years was 14.4 (95% CI 1.7 to 122.2; p = 0.01) and at 10 years was 4.4 (95% CI 1.4 to 13.6; p = 0.01). Subgroup analyses showed that the HR for BCSS in hormonal therapy-naive patients (untreated) at 5 years was 10.8 (95% CI 1.2 to 94.7; p = 0.03). The authors concluded that the MammaPrint signature can accurately select postmenopausal patients at low risk of breast cancer in terms of related death within 5 years of diagnosis, although not at 10 years, and can be of clinical use in selecting postmenopausal women for adjuvant chemotherapy. Again this study employed a very small sample size and was based on postmenopausal women, limiting the applicability of the findings.

Clinical utility

Analytical validity of MammaPrint

In the earlier systematic reviews limited data are available on variability and reproducibility, with a limited number of patients and a moderate number of replications. The only validation study using the MammaPrint assay (rather than the underlying 70-gene signature) showed that only about 80% of fresh-frozen specimens were analysable.

Our review identified no new evidence.

Clinical validity (prognostic ability) of MammaPrint

Earlier systematic reviews identified a range of studies providing evidence on the prognostic ability of the test in heterogeneous populations. The evidence relating to the clinical validity of MammaPrint was not always conclusive nor supportive of the prognostic value of the test. Four studies suggested that the test could predict prognosis, one study failed to verify the prognostic utility of the test and in another the methods and results were at variance with those of other studies.

The current review identified four additional studies that contain data on clinical validity. Of these, two were rated as high quality and two as moderate quality. These studies demonstrated that the MammaPrint score is a strong independent prognostic factor and may provide additional value to standard clinicopathological measures. The majority of the evidence suggests that the test is reliable at predicting outcome at 5 years. 74 However, the population in all of these studies was relatively small, ranging between 102 and 272 patients. One of the studies was conducted in a Japanese population, making generalisation to UK practice difficult. Follow-up was limited to only 5 years in two of the studies.

Clinical utility of MammaPrint

Earlier systematic reviews identified one study on clinical utility, which demonstrated that MammaPrint had an impact on clinical decision-making. The addition of MammaPrint to the standard Dutch clinical assessment of risk (modified by patient preference) in a cohort of 427 patients increased the number of patients receiving adjuvant systemic therapy by 20. However, the follow-up was not long enough to provide evidence of its effect on clinical end points such as DMFS or its utility in predicting treatment benefit.

The current review identified six studies that contained data on the clinical utility of MammaPrint. Of these, two were rated as high quality and four as moderate quality. Five of the six studies reported on how the MammaPrint test reclassifies patients into high- and low-risk groups compared with the risk assigned in current practice. None of these was based in a UK setting. These studies reported that there was a high level of discordance between MammaPrint and current practice, although the studies did not demonstrate how this would impact on treatment decisions. One study reported that the use of MammaPrint would result in altered treatment advice for 40% of patients, but this was based on the assumption that all patients classified as high risk would receive chemotherapy and no patients classified as low risk would receive chemotherapy rather than by providing evidence of actual changes in practice. A study on the benefits of chemotherapy by MammaPrint risk group was identified but omitted from the systematic review because it was based on a pooled analysis of six primary studies (which were included in the review in their own right). This study reports findings on chemotherapy benefit for MammaPrint high-and low-risk groups but the findings are not considered to be robust as the authors do not reanalyse the tumour samples and it is unclear how individual patient data were combined. All of the studies on clinical utility were based on small sample sizes.

Key gaps in the evidence base remain:

• Robust evidence of clinical utility is needed. It is not yet clear whether or not the use of the MammaPrint test will change the management of patients, for example reduce the number of patients unnecessarily treated with chemotherapy or improve patient outcomes through increases in DFS and OS. The ongoing Microarray In Node-negative Disease may Avoid ChemoTherapy (MINDACT) trial, which is summarised in Ongoing randomised trial: the Microarray In Node-negative Disease may Avoid ChemoTherapy trial (see Appendix 8 for further detail of the MINDACT trial), will provide this evidence.

• Only two studies were considered to be of high quality. The rest of the studies in the current review were judged to provide moderate-quality (although retrospective) evidence for MammaPrint. All of the included studies employed very small sample sizes. One of the most characteristic features of the studies was their heterogeneity, particularly with respect to patient populations. All but one92 of the MammaPrint studies were retrospective, and many used old archived tumour samples and non-standardised methods of patient selection. In addition to patient heterogeneity, there is likely to be significant heterogeneity in the chemotherapy treatment as patients were diagnosed with breast cancer over a period of > 20 years (1984–2006) and the standards of care have changed considerably.

• Further issues that may limit the extent to which we can generalise the findings include publication bias and the fact that no studies were conducted in the UK setting.

Analytical validity

No available evidence.

Clinical validity

Stork-Sloots et al. 114 compared BluePrint directly with the intrinsic subtyping using the original intrinsic gene set as developed by Perou et al. ,26 from which the PAM50 gene set has originated. Using 469 independent samples and two publicly available data sets (n = 215, n = 159), the authors reported 5-year survival estimates for the subtypes and for the groups further separated by high-and low-risk MammaPrint categories. They reported that samples with a ERBB2-like or basal-like gene profile showed equally poor 5-year survival rates of ∼65%. However, the ERBB2-like subset of MammaPrint low-risk patients (15%) showed an 89% (95% CI 71% to 100%) survival rate without trastuzumab treatment. When the luminal-like subtype was separated into high and low risk by MammaPrint the survival rate was 56% (95% CI 46% to 68%) for high-risk luminal-like samples and 94% (95% CI 90% to 99%) for low-risk luminal-like samples. The authors concluded that the developed multigene profile can classify breast tumours into luminal-, ERBB2- and basal-like subgroups. By combining this molecular subtyping with MammaPrint risk classification, specific groups of patients can be recognised that are at high risk of recurrence. The low-risk patients within the luminal- and ERBB2-like subclasses have a very low risk of recurrence. There are significant limitations in making any interpretations from this evidence as it is derived only from an abstract. It has been shown that there may be discrepancies between data made available in abstracts and the reporting of results in subsequently published full-length articles. 89 Because of incomplete reporting the methodological quality of studies cannot be confidently assessed by systematic reviewers.

Clinical utility

No available evidence.

Analytical validity

Ebbert et al. 118 reported an analytical validity study using 171 tumour samples from US patients. They reported that within-platform cross-validation of the clinical subtype predictor showed 91.6% concordance. There was 100% reproducibility in subtype predictions across 46 runs testing different subtypes. Subtype predictions across platforms showed 88.1% concordance. Dilution experiments, introducing ‘normal’ breast tissue RNA into breast cancer RNA, showed a systematic switch towards the ‘normal’ signature, with luminal A and luminal B subtypes being most susceptible. The authors concluded that the PAM50 Breast Cancer Intrinsic Classifier is highly reproducible within and across platforms and that the clinical test has utility in the management of ER+ and ER− invasive breast cancer of all stages. The quality of this report was judged as low. Furthermore, the study was based on a small number of tumour samples and full details of the patient characteristics were not provided. In addition, there are significant limitations in making any interpretations from this evidence as it is derived only from an abstract. It has been shown that there may be discrepancies between data made available in abstracts and the reporting of results in subsequently published full-length articles. 89 Because of incomplete reporting the methodological quality of studies cannot be confidently assessed by systematic reviewers.

Bernard et al. ,119 using a cohort of 793 LN+ tumour samples from the GEICAM 9906 randomised trial, reported agreement between reverse transcription-quantitative polymerase chain reaction (RT-qPCR) gene expression and IHC scoring for clinical markers. They showed that there was good agreement between (gene/protein) ESR1/ER, PGR/PR and ERBB2/HER2. The accuracy was significantly lower for MKI67/Ki-67, EGFR/EGFR and KRT5/CK5/6. Discrepancies between the Hercep test and chromogenic in situ hybridisation (CISH) procedure for 2+ and 3+ staining-intensity samples showed that RT-qPCR agreed better with the Herceptest [area under the curve (AUC): 0.95 vs. 0.93). The authors concluded that calling cut-points based on RT-qPCR expression across subtypes is reproducible across data sets and has good agreement with expression by IHC for clinically used biomarkers. The quality of this report was judged as low. Although the study benefits from a relatively large sample size, as these data were reported in abstract form only there are significant limitations in using these results to make generalisations.

Clinical validity

Parker et al. 115 investigated the distribution of intrinsic subtypes in comparison with clinical marker status and risk of relapse models for prognosis in a cohort of 950 Canadian and US majority LN−, ER+ breast cancer patients. They reported that the intrinsic subtypes showed prognostic significance (for recurrence-free survival, RFS) in untreated (no systemic therapy) patients (p < 0.0001) and remained significant in multivariable analyses that incorporated clinical covariates (ER status, histological grade, tumour size and node status) (p < 0.0001). The authors concluded that the intrinsic subtype and risk predictors based on the PAM50 gene set add significant prognostic and predictive value to pathological staging, histological grade and standard clinical molecular markers. The quality of this study was judged to be moderate and the sample size was relatively large.

Nielsen et al. 116 used a cohort of 786 LN+ or higher-risk LN−, ER+ Canadian patients with tumours collected between 1986 and 1992 to assess the numbers of patients assigned to each intrinsic subtype and risk of relapse score against IHC (ER, PR, HER2, Ki67). Adjuvant! Online was used to generate breast cancer recurrence-free survival and disease-specific survival estimates for each patient. They reported that the included patients were considered to be high risk, with overall 10-year RFS of 62% and disease-specific survival of 72%. Those assigned to luminal A tumours had significantly better outcomes (10-year RFS 74%; disease-specific survival 83%) than those assigned to luminal B, HER2-enriched and basal-like tumours. In Cox models incorporating standard prognostic variables, HRs for breast cancer disease-specific survival over the first 5 years of follow-up, relative to the most common luminal subtype, were 1.99 (95% CI 1.09 to 3.64) for the luminal B subtype, 3.65 (95% CI 1.64 to 8.16) for the HER2-enriched subtype and 17.71 (95% CI 1.71 to 183.33) for the basal-like subtype (p = 0.0018). The authors concluded that IHC approaches do work and provide significant prognostic information; however, PAM50 is superior to these in terms of adding significant additional information and in its capacity to identify a particularly low-risk group. This study was judged to be of moderate quality and incorporated a relatively large sample size.

Chia et al. 121 (AIC information has been removed).

Cheang et al. 117 (AIC information has been removed).

Clinical utility

In addition to the clinical validity data reported above, Chia et al. 121 (AIC information has been removed).

(AIC information has been removed.) Using the same data reported by Bernard et al. 119 for a cohort of 793 LN+ tumour samples, Martin et al. 120 (AIC information has been removed). The quality of this report was judged as low. Although the study benefits from a relatively large sample size, as these data were reported in abstract form only there are significant limitations in using these results to make generalisations.

Cheang et al. 117 (AIC information has been removed).

Analytical validity of PAM50

Two abstracts118,119 reported data on analytical validity, both rated as low quality. One118 employed a relatively small sample (n = 171) whereas the other119 was based on a much larger sample (n = 793). These studies provide a comparison of PAM50 against standard IHC measurements. There are significant limitations in making any interpretations from this evidence as it is derived only from abstracts. It has been shown that there may be discrepancies between data made available in abstracts and the reporting of results in subsequently published full-length articles. 89 Because of incomplete reporting the methodological quality of studies cannot be confidently assessed by systematic reviewers.

Clinical validity (prognostic ability) of PAM50

Four studies,115–117,121 two rated as high quality and two rated as low quality, were identified that contain data on clinical validity. Two of these are as yet unpublished. These studies demonstrate that the intrinsic subtypes are significantly associated with outcome, provide additional information to IHC approaches and standard clinicopathological measures and can identify a particularly low-risk group. They demonstrate that prognostic ability has been validated in external cohorts. However, the population in most of the studies was LN+, with the exception of that by Parker et al. ,115 who assessed LN− patients; therefore, the generalisability of these findings to LN−, ER+ patients is limited. Furthermore, no studies were UK based, limiting the generalisation of the findings to UK practice. Finally, as two of these studies were unpublished there are significant limitations in drawing conclusions from this evidence.

Clinical utility of PAM50

(AIC information has been removed).

Key gaps in the evidence base remain and are summarised below:

• The evidence base to date is still immature as the majority of the evidence presented here is in abstract or unpublished form only. Only two studies were considered to be of high quality (these presented clinical validity and clinical utility data), two were of moderate quality and two were of low quality, although it should be noted that only two studies were full peer-reviewed papers.

• Further evidence around analytical validity is required as the current evidence is based on abstracts only.

• Robust evidence of clinical utility is needed. It is not yet clear whether or not the use of the PAM50 test will change the management of patients, for example reduce the number of patients unnecessarily treated with chemotherapy or improve patient outcomes through increases in DFS and OS.

• The fact that no studies were conducted in a UK setting may limit the extent to which we can generalise the findings.

Analytical validity

No available evidence.

Clinical validity (prognostic ability)

Jerevall et al. 122 reported a retrospective analysis of (a subcohort of) a randomised prospective trial cohort. The 588 patients were all postmenopausal, LN− and ER+. The authors reported the development and testing of HOXB13:IL17BR plus MGI as a continuous index (BCI) in a training set (G1) and a test set (G2) of the same trial data. In the training set (G1) BCI classified 59% of patients as low risk. The rate of distant recurrence for the low-risk group was 7.1% (95% CI 0% to 3.5%) and the rate of death was 1.1% (95% CI 0% to 2.6%). In total, 22% were classified as intermediate risk, with a rate of distant recurrence of 17.8% (95% CI 7.6% to 26.8%) and a rate of death of 14.5% (95% CI 5.2% to 22.9%), and 18.4% were classified as high risk, with a rate of distant recurrence of 20.0% (95% CI 8.7% to 30.0%) and a rate of death of 14.7% (95% CI 4.7% to 23.6%). In the test set (G2) 53% of patients were classified as low risk (rate of distant recurrence 8.3%, 95% CI 4.7% to 14.4%; rate of death 5.1%, 95% CI 1.3% to 8.7%), 27% were intermediate risk (rate of distant recurrence 22.9%, 95% CI 14.5% to 35.2%; rate of death 19.8%, 95% CI 10.0% to 28.6%) and 20% were high risk (rate of distant recurrence 28.5%, 95% CI 17.9% to 43.6%; rate of death 28.8%, 95% CI 15.3% to 40.2%). The authors concluded that the BCI was a strong prognostic factor for distant recurrence and BCSD, independent of tumour size, grade, HER2 status and PR status (although tumour size did contribute prognostic value to distant recurrence). The prognostic utility of the BCI was also assessed compared with Adjuvant! Online for the test set (G2). Both BCI and Adjuvant Online were significant predictors of BCSD (BCI: HR 2.3, 95% CI 1.5 to 3.7, p < 0.001; Adjuvant Online: HR 1.4, 95% CI 1.0 to 1.8, p = 0.04) and distant recurrence (BCI: HR 2.0, 95% CI 1.3 to 3.1, p = 0.001; Adjuvant! Online: HR 1.4, 95% CI 1.0 to 1.8, p0.03). The authors concluded that the retrospective analysis of this randomised, prospective trial cohort validated the prognostic utility of HOXB13:IL17BR plus MGI and it was used to develop and test a continuous risk model that enables prediction of distant recurrence risk at the patient level. The study had a long mean follow-up time and a moderate sample size. The study used tumour samples that dated back to 1976, introducing differences in the diagnostic criteria applied to patients.

Clinical utility

No available evidence.

Supplementary evidence

The manufacturer submitted a description of the data gathered on the Randox assay up to August 2011. 123 A summary of this information is provided here.

The objective of the study was to improve the discrimination, to include basal and unclassified (triple-negative) types, subdivide luminal groups into A and B and assign samples to an ERBB2 group. The main indicator of correct typing is whether the samples are correctly typed as ER+ or ER−. A total of 150 archived tumour samples were collected and used on the Randox biochip array. Exclusion criteria included a lack of ER status information or one or both of the housekeeping genes failing to reach adequate expression levels, preventing normalisation of the remaining gene set on the chip. The sample set included information on the following: DFS, OS ER status, PR status and other standard clinical measurements. However, HER2 (ERBB2) status was not available for any patients; thus, hormonal status was either luminal positive or negative. The initial summary, using a patient cohort of 78 individuals, shows an overall agreement of 79% between hormonal status (primarily ER) and the multiplex biochip assay. The authors concluded that these findings were encouraging.

Methodological detail was lacking for the summarised study and the sample size was very small. Quality assessment could not be undertaken because of the limited detail available. It should also be noted that, although Randox separates luminal A and luminal B groups, the overall reported agreement of 79% was based on luminal A and B combined and hormonal status, and did not differentiate between the subtypes.

Analytical validity

No available evidence.

Clinical validity

Bartlett et al. 124 investigated assignment to risk groups using Mammostrat for 1540 UK patients with LN− tumours. They demonstrated that significantly more cases were assigned to high-risk groups for ER− than for ER+ cases (45% vs. 16%; p < 0.001), but there were no differences between other groups. They also looked at associations between risk scores and DRFS, RFS and OS across the different groups: all cases, all ER+ cases, all ER+ cases treated with tamoxifen only and ER+, LN− cases treated with tamoxifen only. For all groups there were significant associations between risk score and RFS, DRFS and OS (with the exception of the ER+, LN− treated with tamoxifen only group for which there was a trend only for OS). Multivariate analyses for each of the four groups showed that risk score was a significant independent predictor of RFS, DRFS and OS, along with clinicopathological predictors, for all cases and all ER+ cases. Risk score was a significant independent predictor of DRFS and OS with a trend for RFS for ER+ cases treated with tamoxifen only, and there was a trend towards significance for Mammostrat score to predict RFS and DRFS in ER+, LN− cases treated with tamoxifen only. The authors concluded that Mammostrat can act as an independent prognostic tool for ER+, tamoxifen-treated breast cancer and that there is a possible association with outcome regardless of LN status and ER− tumours. This study was rated as being of high quality on the basis of the quality assessment employed here. It also benefits from a large sample size overall, although it should be noted that the subsets analysed had relatively small numbers of patients within each. It had a relatively long follow-up of 9 years, and as it employed UK patients the findings should be applicable to UK practice. The samples used in these analyses were relatively old, dating back to 1981; therefore, there may be differences between the patients in terms of stage at presentation and diagnosis.

In a US-based study, Ring et al. 125 assessed 1109 majority LN−, ER+ patients. Using a training cohort the authors demonstrated that a Cox model identified a group of patients as having either poor or moderate prognosis, with a 5-year DFS rate of approximately 75%, as opposed to patients classified as having good prognosis, who had a 5-year DFS rate of approximately 95% (p < 0.001). In the first independent cohort the model identified poor prognosis patients with a 5-year DFS rate of 50%, compared with approximately 70% for patients classified as having moderate prognosis and 87% for patients classified as having good prognosis (p = 0.008). In the second independent cohort the model distinguished ER+ patients classified as having poor prognosis with OS rates of 55%, compared with 75% for patients classified as having moderate prognosis and 90% for patients classified as having good prognosis (p = 0.0039). In both independent cohorts the model was independent of stage, grade and LN status. In the combined independent cohort, for patients with poor or good prognosis (82%), sensitivity for poor prognosis in predicting disease progression was 38%, whereas specificity was 88%. The PPV of poor prognosis was 38% (95% CI 32% to 44%), whereas the NPV was 88% (95% CI 84% to 92%). The authors concluded that the test can significantly improve on traditional prognosticators in predicting outcome for ER+ breast cancer patients. The quality assessment indicated that this study was of moderate quality and it used a large sample of patients. The test was also validated in an external cohort. However, there are limitations in that the tumour samples used in these analyses date back as far as 1974; therefore, there may be differences between the patients in terms of stage at presentation and diagnosis.

Ross et al. 126 examined the association between the clinical outcomes recurrence-free interval (RFI), DRFI and BCSD and stratification by the Mammostrat test in 711 ER+, LN− tamoxifen-treated patients taken from the NSABP B14 and B20 trials. Of this group approximately 58% were classified as low risk, 21% as moderate risk and 21% as high risk. There was a significant association between patients stratified by the Mammostrat test and RFI (HR 1.3; 95% CI 1.1 to 1.6; p = 0.006). This was not significant in the low-risk group compared with the moderate-risk group (log-rank, p = 0.05) but was significant in the low-risk group compared with the high-risk group (HR 1.8; 95% CI 1.2 to 2.6). The authors reported a significant association between patients stratified by the test and DRFI (HR 1.4, 95% CI 1.1 to 1.7; p = 0.001). In the low-risk group compared with the moderate-risk group this was not significant whereas in the high-risk group compared with the low-risk group it was significant (HR 2.1, 95% CI 1.4 to 3.1; p = 0.0004). They also reported a significant association between patients stratified by the test and BCSD (HR 1.5, 95% CI 1.2 to 1.9; p = 0.0003). In the low-risk group compared with the moderate-risk group this was not significant whereas in the high-risk group compared with the low-risk group it was significant (HR 2.3; 95% CI 1.5 to 3.5; p < 0.0001). The Kaplan–Meier estimate of the proportion of patients recurrence free after 10 years was 82% (95% CI 79% to 85%) for the group overall, 85% (95% CI 81% to 88%) for the low-risk group, 85% (95% CI 80% to 91%) for the moderate-risk group and 73% (95% CI 65% to 80%) for the high-risk group. In multivariate analyses the Mammostrat test had significant prognostic power independent of age and tumour size (HR 1.3; 95% CI 1.1 to 1.6; p = 0.007). The authors concluded that the risk index was significantly associated with clinical outcome among the ER+, LN− tamoxifen-treated patients. Ross et al. 126 also reported data relating to clinical utility, which are detailed in the following section.

Clinical utility

Supplementary evidence

(CIC information has been removed.)

Analytical validity of Mammostrat

No evidence was found on the analytical validity of the test.

Clinical validity (prognostic ability) of Mammostrat

The three studies identified suggest that Mammostrat can act as an independent prognostic tool for ER+, tamoxifen-treated breast cancer. Two of the studies were rated as high quality and one as moderate quality. The test has been validated in an external cohort. Although the evidence base for Mammostrat is relatively immature, these initial studies include a large sample size, appear to be of reasonable quality and, in the case of Bartlett et al. ,127 use a UK-based population.

Clinical utility of Mammostrat

One study reported on clinical utility and was rated as high quality. Initial evidence suggests that low- and high-risk groups benefited from chemotherapy, with high-risk patients benefitting more than low-risk patients. The moderate-risk group did not appear to benefit. There was no published evidence on reclassification of risk groups compared with conventional risk classifiers, and no evidence on the impact of the test on decision-making. Further evidence is required.

Clinical validity

Cuzick et al. 83 reported a study assessing the prognostic value of IHC4. The IHC4 score was created and validated in one cohort (G1) and further validated in an independent cohort (G2). G1 was a retrospective cohort comprising patients from the TransATAC trial (a multinational trial, including the UK). The majority of the 1125 patients in G1 were LN− and hormone receptor positive. In this cohort there was a total of 195 recurrences of which 145 were distant recurrences. In LN− women there were 101 recurrences of which 67 were distant recurrences. The authors determined the value of each of the four IHC markers in three ways: univariately, as an addition to a model containing the classical variables, and when added to a model containing the classical variables and the other three IHC markers; this was carried out for all women and separately for LN− women only. They found that each of the four variables added a significant amount of information. Ki-67 was the most powerful univariately, but not in multivariate analyses because of its correlation with grade. For the multivariate models PR was most prognostic overall, but less so in LN− patients, in whom ER, HER2 and Ki-67 had similar values. The overall contribution of the IHC measurements for distant recurrence was highly significant [_χ2 (4 degrees of freedom, df) = 39.1; p < 0.0001], and it was reported that the median IHC4 score for all patients was −4.2 and the interquartile range (IQR) −29.9 to 29.9. The HR for a change from the 25th to the 75th percentile of the IHC4 score for all patients was 5.7 (95% CI 3.4 to 9.7) in univariate analysis and 3.9 (95% CI 2.4 to 6.7) when added to clinical score. In a second validation cohort of 786 ER+ younger patients treated in the UK (G2), the authors demonstrated that IHC4 score was highly significantly predictive of outcome (HR 4.8; 95% CI 2.2 to 10.2) for a change from the 25th to the 75th percentile in univariate analysis, and gave similar results when added to clinical score (HR 4.4; 95% CI, 2.0 to 9.3; p < 0.0001). The authors concluded that they have created a prognostic model that integrates IHC information with classical clinical and pathological variables and may prove helpful in managing early ER+ breast cancer in postmenopausal patients, but that additional studies are needed to determine the general applicability of the IHC4 score. This study was rated as high quality on the basis of the quality assessment checklist. It has employed a large sample size and the test has been validated in an external cohort of UK patients.

Supplementary information

Supplementary data were provided by the co-investigators of this study84 after request by the External Assessment Group (EAG) (Professor Mitch Dowsett, Royal Marsden Hospital, London, September 2011, personal communication). Two analyses were conducted in the TransATAC trial – among women with LN−, ER+, HER2− early breast cancer (n = 707) and among all women (n = 1117).

Discussion with the co-investigators of the study (Professor Mitch Dowsett, Royal Marsden Hospital, London, July 2011, personal communication) indicated that the test was meant to be used in conjunction with clinicopathological parameters and therefore data for the final algorithm (using age, grade and tumour size) were used in the economic model. They provided data on the reclassification and risk of distant recurrence of patients using IHC4 plus clinical score (for simplicity the term IHC4 will be used in the report but the data refer to the use of the test in conjunction with clinicopathological parameters). Although cut-offs are not available for IHC4, for the purpose of the economic assessment investigators provided risk classification evidence of IHC4 based on low, intermediate and high risk of distant recurrence. Cut-offs were defined using a similar approach to the classification with OncotypeDX (< 10%, 10–20% and > 20% risk of distant recurrence). The cut-offs used for IHC4 are, however, exploratory and were defined only to populate the economic model. More details are available in Chapter 3.

Among the 707 women with LN−, ER+, HER2− early breast cancer, 85.3% of patients (n = 603) were classified as having a low risk of distant recurrence using IHC4. The proportions of patients classified as having an intermediate and a high risk of distant recurrence were 9.9% (n = 70) and 4.8% (n = 34). The risk of distant recurrence for patients classified as low, intermediate or high risk of distant recurrence by IHC4 is shown in Figure 7.

FIGURE 7.

Time to distant recurrence by IHC4 + clinical score group for patients who were LN−, ER+, HER2−.

The reclassification of the three IHC4 risk groups against the two NPI risk groups used in the economic model (NPI ≤ 3.4 and NPI > 3.4) is presented in Table 26.

Analytical validity of IHC4

We found no evidence on analytical validity. Although the use of IHC4 may be extended for use in other laboratories, the included paper suggests that there may be a lack of reproducibility of the test in relation to Ki-67. The authors suggest that, because IHC4 offers the possibility of carrying out the test in local laboratories, full validation would require evaluation of the IHC4 score when carried out in a range of local laboratories. Reproducibility of the test would need to be confirmed and quality assurance programmes put in place.

Clinical validity (prognostic ability) of IHC4

One study on the clinical validity of IHC4 was available, which claims that the IHC4 score is a highly significant predictor of distant recurrence. This initial study included a large sample size and detailed the development of the test in one cohort and the external validation of the test in an independent cohort. The study has been rated as high quality on the basis of the quality assessment employed.

Clinical utility of IHC4

There is currently no evidence on the clinical utility of IHC4 in terms of its ability to change treatment decisions or its ability to predict chemotherapy benefit. Although there are no published data on clinical utility, unpublished data were obtained to populate the economic model.

Supplementary evidence

The manufacturers submitted two draft full papers based on the same data and one draft abstract (of a full paper). The study design and patient characteristics included in these documents are presented in Tables 27 and 28.

Description and critique of Retel et al.:146 cost-effectiveness of the 70-gene signature compared with St Gallen guidelines and Adjuvant! Online for early breast cancer

Description and critique of Chen et al.:132 cost-effectiveness of the 70-gene MammaPrint signature in node-negative breast cancer

Description and critique of Tsoi et al.:135 cost-effectiveness analysis of recurrence score-guided treatment using a 21-gene assay in early breast cancer

Description and critique of OHTA analysis:134,136 gene expression profiling for guiding adjuvant chemotherapy decisions in women with breast cancer

Adjuvant! Online

Despite NICE recommendations to use Adjuvant! Online,7 clinical experts indicated that it is not comprehensively used in the UK for a number of reasons:

• It is based on a US population and there are some difficulties in applying the Adjuvant! Online data to the UK population.

• Although it is a useful aid for discussing risk of recurrence and benefits of chemotherapy with patients, it is viewed by some as complex to use and interpret for decision-making purposes.

• It cannot be used by all NHS trusts as access is blocked by some trusts for information technology security reasons.

The published evidence reports outcomes based only on the risk of recurrence estimated using Adjuvant! Online. However, both the risk of recurrence and predicted impact of adjuvant treatments would be used to inform treatment decisions.

Nottingham Prognostic Index

Nottingham Prognostic Index-based guidelines are widely used in some parts of the UK to inform decisions about adjuvant chemotherapy. The NPI forms part of the National Cancer Dataset for breast cancer so the NPI score should be given in the report of every invasive breast cancer case in the UK. It is simple to use although it may be considered to be less informative and therefore potentially less useful than Adjuvant! Online, particularly when discussing prognosis and potential treatments with patients.

Assignment of patients into different risk groups

OncotypeDX, MammaPrint and Mammostrat assign women into risk groups – high, intermediate and low risk (OncotypeDX and Mammostrat) or good and poor prognosis (MammaPrint). The IHC4 test provides a risk score only; however, patients have been allocated into risk groups, similar to the OncotypeDX risk groups, for the purposes of this assessment (see Model inputs: test-specific parameters for more details).

In the economic model, women were first stratified into two NPI groups (women with a NPI score ≤ 3.4 and women with a NPI score > 3.4). This was carried out to allow the use of the test with different subgroups of patients to be explored and to allow adjustment of non-UK clinical evidence to reflect the NPI distribution in the UK population. This also takes into account the prognostic value of the treatment decision using clinicopathological parameters. Indeed, within the current treatment decision-making process based on clinicopathological parameters, it is possible to identify patients who are at a higher risk of distant recurrence. Within these two NPI groups, patients were further reclassified into low, intermediate or high risk (or low and high risk in the case of MammaPrint) of recurrence according to the outputs of the new tests.

In simple terms, patients are assigned into different boxes, each with a different prognosis. Patients are assigned to the same boxes for the comparator (current practice) arm or the intervention arm (GEP and expanded IHC tests) as the diagnostic tool does not affect the prognosis of those patients if there is no change in the adjuvant treatment.

Identification of women receiving adjuvant chemotherapy on the basis of the test results

Once women have been assigned into the different boxes (with different prognosis) the next step is to identify which women would receive adjuvant chemotherapy.

The aim of categorising patients into risk groups based on distant recurrence with the GEP and expanded IHC tests is to identify patients who have a greater chance of developing a distant recurrence/recurrence. The risk groups identified by the new tests are therefore expected to influence the targeting of chemotherapy. However, other factors will also influence the decision regarding chemotherapy, including clinical and pathological factors, along with patient choice. In clinical practice a proportion of women classified as low risk of distant recurrence using GEP and expanded IHC tests may still receive chemotherapy; similarly, a proportion of women considered to be at high risk may not receive chemotherapy, as shown in Spain173 or in the USA174 for OncotypeDX.

In the intervention arm of the economic model the proportion of patients who would receive chemotherapy is based on the expected interpretation of the test, for example women categorised as high risk of recurrence are more likely to receive chemotherapy than women categorised as low risk. Some previous analyses have assumed that chemotherapy is received based on the risk group only. For instance, all women defined as high risk receive chemotherapy. However, in clinical practice other issues are likely to impact on this decision (clinicopathological factors, age of patient, patient choice, etc.) and it is unlikely that 100% of high-risk patients will receive chemotherapy. An adjustment for such factors was therefore used in the model.

In the comparator (current practice) arm, the proportion of women receiving chemotherapy is based on cancer registry data. Two subgroups are considered: women with a NPI score ≤ 3.4 and women with a NPI score > 3.4. Because the model categorised women into boxes (defined by the new test a posteriori) and the oncologist is blind to the results of the new test, we assumed that the probability of receiving chemotherapy was the same in the current practice arm whether patients were reclassified as low, intermediate or high using GEP and expanded IHC tests. However, it is likely that patients who are classified as high risk by the new test are more likely to have been identified as high risk under current practice and, therefore, are more likely to have received chemotherapy than those patients classified as low or intermediate risk by the new test. To further explore this assumption, data from the Holt et al. study78 were analysed by the EAG to approximate the proportion of patients who were recommended chemotherapy by RS group before knowledge of the OncotypeDX score (analysis conducted by the EAG using individual patient-level data submitted by Genomic Health).

Overall, 30.43%, 30.30% and 68.42% of patients with a low, intermediate and high RS score were recommended chemotherapy before knowledge of the OncotypeDX test results. Preliminary analyses suggested that the proportion is likely to be higher for patients with a high RS, but the sample size was too small (69 in low RS, 33 in intermediate RS and 19 in high RS) to draw any definitive conclusion. While preliminary, this analysis suggested that our assumption (that the probability of receiving chemotherapy in the comparator is constant irrespective of the RS group) might be conservative as current practice using clinicopathological parameters does appear to add some prognostic value.

Natural history of breast cancer

The final part of the model was a Markov model. Patients were able to move between five possible health states: recurrence free (A), distant recurrence (B), local recurrence (C), long-term adverse events after chemotherapy (D) and death (from breast cancer, long-term adverse events or general causes – E).

As shown in Figure 11, patients enter the model in the recurrence-free survival health state (A) and remain in that health state until they develop a distant recurrence (B), have an adverse event after chemotherapy (D) or die from breast cancer or general causes or from their adverse event (E). After a distant recurrence (B), patients remain in this health state until they die from either breast cancer or general causes (E) or develop an adverse event for women treated with chemotherapy (D). Patients developing an adverse event after chemotherapy can remain in that health state, die from their adverse event or die from general causes (E). The estimation of long-term adverse events is simplistic. No distinction was made between patients developing long-term adverse events after a recurrence (B) and patients developing long-term adverse events in the recurrence-free health state (A). Furthermore, patients with a long-term adverse event were assumed to remain in that health state (D) until death (E) and were not allowed to move to other health states.

FIGURE 11.

Schematic of the model structure.

Local/regional recurrences have been modelled by considering the costs and quality of life decrements (disutility) assuming that a proportion of patients entering the distant recurrence state (B) have previously experienced a local recurrence (C). No transition probabilities were used between this health state and death or adverse events. This is simplistic but justified by the fact that the risk categories (used in the economic model) defined by the new tests (OncotypeDX, MammaPrint, IHC4) have been defined according to the risk of developing distant recurrence and there is no robust evidence to accurately model the development of local recurrence and the different transitions between health states for patients reclassified as low, intermediate or high risk with GEP and expanded IHC tests.

Mean age of patients entering the model

The EAG economic assessment focuses on women with ER+, LN−, HER2− who are aged ≤ 75 years. Patients were assumed to enter the economic model at a mean age of 58.3 years based on the average age in the ECRIC dataset of women with ER+, LN−, HER2− aged < 75 years (Eastern Cancer Registration and Information Centre, July 2011, personal communication). Although patient age is not used to determine treatment selection, experts suggested that women aged > 70–75 years are much less likely to be offered chemotherapy because of issues of frailty and comorbidities. Sensitivity analysis was conducted varying the mean baseline age.

Of note, the model does not separate pre- and postmenopausal women but most of the evidence was taken from postmenopausal women. In addition, it was not possible to explore different age thresholds as we did not have access to patient-level data.

Baseline Nottingham Prognostic Index distribution in England and Wales

The economic assessment separates women with a NPI score ≤ 3.4 and women with a NPI score > 3.4. The baseline NPI distribution was extracted from the combined (EAG analysis) ECRIC (2007 onwards) and WMCIU (2007 only) data (West Midland Cancer Intelligence Unit, July 2011, personal communication; Eastern Cancer Registration and Information Centre, July 2011, personal communication). Approximately two-thirds of patients had a NPI score ≤ 3.4 (Table 40).

For the scenario assuming that the test is given to all women with ER+, LN−, HER2− early breast cancer, we modelled patients with a NPI score of ≤ 3.4 and patients with a NPI score > 3.4 separately to account for the prognostic value of the treatment decision using clinicopathological parameters.

Additional sources of evidence for the baseline distribution of NPI were considered in sensitivity analysis (Table 41). These included women with ER+, LN−, HER2− early breast cancer from the TransATAC trial (Professor Mitch Dowsett, Royal Marsden Hospital, London, September 2011, personal communication) and the ER+, LN−, HER2− population from the Holt et al. study78 (analysis conducted by the EAG using individual patient-level data submitted by Genomic Health).

Proportion of patients receiving chemotherapy in current clinical practice in England and Wales

As described in Comparator used in the economic model, cancer registry data were used to reflect the proportion of women with ER+, LN−, HER2− early breast cancer who currently receive chemotherapy in England and Wales. Data from ECRIC (2007 onwards; Eastern Cancer Registration and Information Centre, July 2011, personal communication) and WMCIU (2007 only) (West Midland Cancer Intelligence Unit, July 2011, personal communication; Eastern Cancer Registration and Information Centre, July 2011, personal communication) were combined (EAG analysis) and showed that overall about 14.4% of women aged < 75 years with ER+, LN−, HER2− early breast cancer received chemotherapy. When separating patients with a NPI score of ≤ 3.4 and a NPI score > 3.4, 4.6% and 33.6% of women received chemotherapy respectively (Table 42).

Death from breast cancer causes after a distant recurrence

In the base case, the hazard rate of death after a distant recurrence was taken from Thomas et al. 157 Thomas et al. 157 analysed the time to death from relapse among 77 relapsed breast cancer patients. The first site of relapse was distant in 51 patients (66%) with the remaining patients having locoregional recurrences. The study included a mix of patients with regard to ER status (55% ER+), nodal involvement (66% LN−) and HER2 status (75% HER2−). The study reported a median survival of 40.1 months (equating to an annual hazard rate of about 0.30) and this value was used in the base case. Sensitivity analyses were conducted varying the time spent in the distant recurrence health state within the reported CI in this study. 155

In the base case we assumed that the risk of death after a recurrence was independent of the prognosis of the patient, because of the lack of more informative data. Discussion with clinical experts indicated that it is likely that low- and high-risk patients will spend a different amount of time in the distant recurrence health state. High-risk patients are likely to have more aggressive cancer and are likely to spend less time in recurrence before death. A scenario analysis was therefore explored assuming different times in recurrence between risk groups. Because there are no published data to our knowledge on survival after distant metastasis for patients with different prognosis, assumptions were made to examine the impact of this assumption on the ICER.

Costs

All costs are in 2010 prices.

Endocrine therapy costs

The economic model considers only women with ER+ early breast cancer and assumes that all patients receive endocrine therapy. Five endocrine therapy regimens were considered as per NICE Technology Appraisal 112:162 (1) tamoxifen for 5 years, (2) anastrozole for 5 years, (3) letrozole for 5 years, (4) tamoxifen for 2 years plus exemestane for the final 3 years and (5) tamoxifen for 5 years followed by extended therapy with letrozole for a further 3 years.

Drug costs were taken from BNF 61. 161 It was assumed that each endocrine therapy was given once daily at a 20-mg, 1-mg, 2.5-mg and 25-mg dosage for tamoxifen, anastrozole, letrozole and exemestane respectively. The annual cost for each drug is presented in Table 44.

The probability that a patient will be treated with each regimen was taken from the costing template accompanying TA112. 162 It was assumed that 40% of patients received tamoxifen for 5 years, 20% anastrozole for 5 years, 20% letrozole for 5 years and 20% tamoxifen for 2 years plus exemestane for the final 3 years. It was further assumed that 10% of patients received tamoxifen for 5 years followed by extended therapy with letrozole for a further 3 years. After weighting, the annual drug cost was calculated to be £668.90 for the first 5 years and £110.70 for the remaining 3 years.

In addition to drug costs, monitoring cost were included. We assumed that patients treated with endocrine therapy have two follow-up appointments in the first year and one follow-up appointment every subsequent year (£129 based on NHS reference costs 2009/10,175 370 Medical Oncology). We further assumed that patients had one mammogram every year (£46.37 based on Campbell et al. 176) for a maximum of 5 years.

Chemotherapy costs

It was assumed that all patients received FEC75 as clinical opinion indicated that this is the most commonly used chemotherapy regime for this population (ER+, LN−, HER2−). Sensitivity analysis was carried out varying the cost to explore the sensitivity of the results to this assumption. Note that the choice of chemotherapy (FEC75) in the economic model impacts on cost only as the effect of chemotherapy was taken from a separate source of data49 that uses CMF/MF regimes. No effectiveness data were available for FEC75 for this group of patients.

The cost of the chemotherapy drugs was calculated according to the regime description of FEC75 given by Avon, Somerset and Wiltshire Cancer Services in their chemotherapy protocol documents. 177 All drug costs are taken from BNF 61. 161 We also assumed a dosage per BSA of 1.75 mg/m² based on the value reported by Sacco et al. ,168 estimated in women with breast cancer in the UK.

The chemotherapy drug cost per cycle is summarised in Table 45. No drug wastage was assumed.

Discussion with clinical experts indicated that FEC75 is usually given for six cycles. The number of cycles was varied in sensitivity analysis. In addition to drug costs, we assumed an additional pharmacy cost of £38162 per cycle to account for the chemopharmacy/aseptic costs.

Furthermore, administration costs were assumed to be £270.60 for the first cycle of treatment (NHS reference costs 2009/210,175 S13Z: Deliver more complex Parenteral Chemotherapy at first attendance) and £284.50 for the remaining cycles (NHS reference costs 2009/10,175 SB15Z: Deliver subsequent elements of a Chemotherapy cycle). Patients were also assumed to have a separate outpatient appointment before administration of each cycle of the chemotherapy (NHS reference costs 2009/10,175 370 Medical Oncology – £129).

Patients were also assumed to be monitored and to receive one liver function test (£12.68), a test of urea and electrolytes (£12.30) and a full blood count (£5.81) at each cycle based on 2008/9 data from the Sheffield Teaching Hospital (Sheffield Teaching Hospital Trust, 2007–8, personal communication), uplifted to 2010 prices. 178 Finally, it was further assumed that 25% of patients would have an echocardiogram before starting chemotherapy (NHS reference costs 2009/10,175 RA60Z – £59).

The total cost of chemotherapy including drug acquisition cost, administration and monitoring was calculated to be £4099 for the entire course of chemotherapy (six cycles of treatment).

Costs of short-term adverse events associated with chemotherapy

Short-term adverse events were included for patients receiving chemotherapy.

The probability of short-term grade 3 and grade 4 adverse events was extracted from the PACS-01 clinical trial for patients treated with FEC100 as no data were available for FEC75. 179 This included anaemia, thrombocytopenia, neutropenic infection, nausea/vomiting and stomatitis. Short-term grade 3 and grade 4 adverse events were costed using the NHS reference costs where appropriate. 175 The total cost of treating short-term adverse events was estimated to be £275.61 per patient (Table 46). This excluded the cost associated with the secondary prevention of febrile neutropenia using G-CSF prophylaxis, included separately (see next section).

Costs of the secondary prevention of febrile neutropenia

We assumed that G-CSF prophylaxis was given for the secondary prevention of neutropenia to women receiving chemotherapy only. In the base case it was assumed that about 25% of patients would receive G-CSF prophylaxis (Dr Matthew Winter, Consultant in Medical Oncology, Sheffield Teaching Hospitals NHS Foundation Trust, September 2011, personal communication) and that G-CSF would be given for an average of three cycles. Patients were assumed to receive filgrastim at a dose of 500,000 units/kg daily for six days after each cycle of chemotherapy (maximum three cycles). A mean weight of 66 kg was assumed and the drug cost per injection of 30 million units was assumed to be £59. 161 Filgrastim was assumed to be administrated by a district nurse (£39 per injection) using the cost from the Personal Social Services Research Unit (PSSRU). 178

Overall, the cost associated with the secondary prevention of adverse event for patients receiving chemotherapy was estimated to be £485.30 per patient.

Costs associated with long-term adverse events

Potential long-term adverse events include secondary malignancies and congestive heart failure (CHF). Although CHF is more common than secondary malignancies, the development of cancer is likely to have more serious consequences and to be associated with a higher impact on health-care resources than the management of CHF.

The base-case economic model included acute myeloid leukaemia (AML) as a long-term adverse event after chemotherapy. The probability of developing AML was based on the 8-year cumulative probability of developing AML in women treated with epiribucin extracted from a meta-analysis of 19 trials conducted in early breast cancer. 180

The meta-analysis showed that the 8-year cumulative probability of AML was 0.37% (95% CI 0.13 to 0.61%) in women receiving a cumulative dose of cyclophosphamide < 6300 mg/m² and a cumulative dose of epirubicin < 720 mg/m² (n = 4760). 180

We further assumed that patients spend 8 months on average in the AML health state at a mean cost of £11,500 based on the approximate mean life-years and mean costs estimated by the manufacturer for the NICE technology appraisal of azacitidine for myelodysplastic syndromes. 181 These assumptions were varied in sensitivity analysis.

Costs associated with the management of distant recurrence

The costs associated with distant recurrence were derived from Thomas et al. 157 using a sample of 77 patients with relapsed breast cancer. Costs included active supportive care and end-of-life care. Costs specifically associated with terminal care were removed to avoid double counting as these were included separately in the economic model. (Note that only cost items described as supportive/terminal care were removed.)

After removing cost items that were specific to terminal care, we estimated the 6-monthly cost to be approximately £4082 (uplifted to 2010 prices). 178 We assumed that the cost was constant over time. This is very simplistic as evidence shows that the cost is higher in the first 2 years and decreases thereafter. 182 However, the impact of this assumption is likely to be minimal because this affects both the comparator arm and the intervention arm in the model.

Costs associated with the management of local recurrence

The cost of local recurrence was taken from Karnon et al. 182 and was estimated at £14,132 (uplifted to 2010 prices). 178

Costs associated with the management of death from breast cancer

Finally, the cost associated with terminal care/end of life was taken from Campbell et al. 176 and was assumed to be about £4038. This cost was applied as a one-off in the economic model, immediately before death from breast cancer.

Health-state utilities

Quality of life utility scores were identified from a recent systematic review of utility values in breast cancer. 159 The utility values used in the model are given in Table 47.

Utility values for patients in the recurrence-free and distant recurrence health states were extracted from Lidgren et al. 148 These were EQ-5D values and using this study allowed values for recurrence free (0.824) and distant recurrence (0.685) to be taken from the same study for consistency (see Table 47). The study followed 361 breast cancer patients attending the breast cancer outpatient clinic at Karolinska University Hospital Solna between April and May 2005. The decrement in utility per patient experiencing a local recurrence was taken from Campbell et al. 176 and assumed to be −0.108 in the base case. We assumed that patients with AML have a utility value of 0.26 based on the value used in a previous economic evaluation conducted in Canada. 183 We further assumed that patients receiving chemotherapy have a disutility of 0.038, taken from Campbell et al. 176 for women treated with E-CMF (epirubicin, cyclophosphamide, methotrexate and 5-fluorouracil)/FEC60 in the first year. This is believed to capture the decrement in utility associated with the administration of chemotherapy and related adverse events. Finally, a decrement in utility was applied for patients dying from breast cancer, derived from Campbell et al. 176 and Lidgren et al. 148 Utility values were varied in sensitivity analysis.

Death from causes other than breast cancer

The mortality rate from causes other than breast cancer was extracted from UK life tables (2007–9) for women184 after adjustment to remove death attributable to breast cancer.

Proportion of patients with distant recurrence who have previously experienced a local recurrence

In the base case we assumed that 10.5% of patients entering the distant recurrence state have previously experienced a local recurrence. This is based on an analysis conducted in 3601 women with early breast cancer enrolled in previous European Organisation for Research and Treatment in Cancer (EORTC) trials (10801, 10854 and 10902), which showed that the presence of locoregional recurrence was a significant prognostic risk factor for the occurrence of distant recurrence. 185 The analysis showed that, among the 1224 patients who developed a distant recurrence, 129 patients had a distant recurrence after a locoregional recurrence.

We did not make any assumptions about the time spent in the local recurrence health state. Local/regional recurrences have been modelled by considering the cost and quality of life decrements (disutility), assuming that a proportion of patients entering the distant recurrence state have previously experienced a local recurrence.

Clinical parameters: OncotypeDX and IHC4

The systematic review of evidence indicated that the OncotypeDX test is the furthest along the validation pathway compared with other similar tests, and the evidence base, in particular in relation to the prognostic ability of the test, was reasonably sound. The evidence base for IHC4 is less developed but there is direct evidence relating to the performance of IHC4 compared with that of OncotypeDX and so the clinical evidence relating to these two tests is described together.

The primary analysis compared current clinical practice with treatment guided using OncotypeDX or IHC4 in addition to current practice. This was carried out using evidence that directly compared the test results for OncotypeDX and IHC4. An overview of this evidence was presented in Cuzick et al. 84 However, the specific data used in the economic model for IHC4 were unpublished and were made available to the EAG for the purpose of this assessment (Professor Mitch Dowsett, Royal Marsden Hospital, London, September 2011, personal communication).

Clinical parameters: Mammostrat

Compared with OncotypeDX, the evidence base for Mammostrat was less developed and some gaps were identified by the systematic review of the literature (see Chapter 2, Quality of included studies: Mammostrat test). Most of the evidence available for Mammostrat related to the clinical validity (prognostic ability) of the test. No published analyses of reclassification against Adjuvant! Online or NPI or the impact of the test on decision-making were identified.

Because of the gaps and uncertainties in the data available, an exploratory analysis was carried out to assess the cost-effectiveness of Mammostrat in addition to current clinical practice compared with current clinical practice in England and Wales.

Clinical parameters relating to the three main components of the model are described in turn.

Clinical parameters: MammaPrint

Compared with OncotypeDX, the evidence base for MammaPrint is less well developed and gaps were identified by the systematic review of the literature (see Chapter 2, Results: MammaPrint test). The systematic review of the literature indicated that the evidence base, in particular in relation to the prognostic ability of the test, is developing but is based on cohort studies with small sample sizes. None of the studies is based on UK patients and studies are mainly derived from premenopausal women and so are not representative of the population of women with ER+, LN−, HER2− early breast cancer. No robust evidence on clinical utility was identified. The one study identified that considered chemotherapy benefit by MammaPrint risk group had significant limitations. 111 No UK studies of the impact of the test on clinical decision-making were identified.

Because of the gaps and uncertainties in the data available, an exploratory analysis was carried out looking at the cost-effectiveness of adjuvant chemotherapy guided using MammaPrint in addition to current clinical practice compared with current clinical practice in England and Wales.

Clinical parameters relating to the three main components of the model are described in turn.

Deterministic results

Assuming that the tests were offered to all women with ER+, LN−, HER2− early breast cancer, treatment guided using OncotypeDX was predicted to lead to an increase in the proportion of patients receiving chemotherapy compared with current clinical practice under our base-case assumptions (19.11% vs. 14.42%). On the contrary, treatment guided using IHC4 was predicted to lead to a reduction in the proportion of patients receiving chemotherapy compared with current clinical practice under our base-case assumptions (9.57% vs. 14.42%). More women were classified as high or intermediate risk with OncotypeDX than with IHC4, and were therefore more likely to receive chemotherapy.

For a cohort of 1000 women with ER+, LN−, HER2− early breast cancer, we predicted that 76 distant recurrences would occur under current clinical practice. Treatment guided using OncotypeDX and IHC4 was predicted to reduce the number of distant recurrences to 64 and 71, respectively, under the assumptions used for the base-case analysis.

The mean discounted cost of treatment guided using current clinical practice, OncotypeDX and IHC4 was estimated to be £6519, £9094 and £6340 per patient respectively. The breakdown of costs by category is presented in Table 65. Treatment guided using OncotypeDX was estimated to reduce the costs associated with the management of recurrences (distant and local recurrences, terminal care) but incurred additional costs to perform the test (£2580) compared with current clinical practice. IHC4 also reduced the costs associated with recurrences, but also reduced the costs associated with chemotherapy, for an additional test cost of £150 per patient compared with current clinical practice.

The mean discounted QALYs were 13.44, 13.54 and 13.49 for current clinical practice, OncotypeDX and IHC4 respectively.

Compared with current clinical practice, the incremental cost for treatment guided using OncotypeDX was estimated to be £26,940 per QALY gained. Chemotherapy treatment guided using IHC4 was dominant (i.e. provided more QALYs at a lower cost) compared with current clinical practice. These results are based on the assumption that OncotypeDX has predictive ability. This assumption was tested in sensitivity analysis.

Treatment guided using OncotypeDX, IHC4 and current clinical practice was also compared using incremental analysis, that is, the least effective strategy was compared with the next least effective strategy that was neither dominated nor extendedly dominated. The base-case costs and QALYs are shown in Table 66. Treatment guided using OncotypeDX provided the most benefit (13.54 QALYs) at the highest cost (£9094). The ICER for treatment guided using OncotypeDX compared with treatment guided using IHC4 was £55,406 per QALY gained.

Probabilistic results

The results of PSA using 2500 iterations are shown in Table 67. Treatment guided using IHC4 remained dominant (i.e. provided more QALYs at a lower cost) compared with current clinical practice. The incremental cost for treatment guided using OncotypeDX was £29,503 per QALY gained compared with current clinical practice and £64,111 per QALY gained compared with IHC4 (see Table 67).

Figure 13 shows the cost-effectiveness acceptability curve (CEAC) using results generated over a lifetime horizon. The curve shows the probability of each test being cost-effective for different monetary values that the decision-maker may be willing to pay for an additional QALY. The CEAC shows that treatment guided by IHC4 was the most cost-effective strategy compared with current clinical practice and OncotypeDX when using a willingness-to-pay threshold of £20,000 per QALY gained (in 99.48% of cases). The probability that treatment guided using OncotypeDX was cost-effective at a £20,000 threshold was 0.40% in the incremental analysis and 12.44% compared with current clinical practice alone.

FIGURE 13.

Cost-effectiveness acceptability curve for the primary analysis comparing OncotypeDX and IHC4 with current clinical practice assuming that the test is given to all women with ER+, LN−, HER2− early breast cancer in England and Wales.

Deterministic results

Assuming that the tests were offered only to women with ER+, LN−, HER2− early breast cancer with a NPI score > 3.4, a greater proportion of patients were predicted to receive chemotherapy when using OncotypeDX (under our base-case assumptions) and a lower proportion were predicted to receive chemotherapy when using IHC4 compared with current clinical practice (34.72% vs. 26.31% vs. 33.60% respectively).

Treatment guided using IHC4 was dominant (i.e. provided more QALYs at a lower cost) compared with current clinical practice. The incremental cost for treatment guided using OncotypeDX was £9007 per QALY gained compared with current clinical practice and £26,859 per QALY gained compared with IHC4 (Table 68). This is based on the assumption that OncotypeDX has predictive ability. This assumption was tested in sensitivity analysis.

Probabilistic results

The results of PSA using 2500 iterations are shown in Table 69. Treatment guided using IHC4 remained dominant (i.e. provided more QALYs at a lower cost) compared with current clinical practice. The incremental cost for treatment guided using OncotypeDX was £9774 per QALY gained compared with current clinical practice and £31,125 per QALY gained compared with IHC4 (see Table 69).

The CEAC (Figure 14) shows that treatment guided by IHC4 was the most cost-effective strategy when using a willingness-to-pay threshold of £20,000 per QALY gained (in 81.24% of cases). The probability that treatment guided using OncotypeDX was cost-effective at a £20,000 threshold was 18.60% in the incremental analysis and 91.56% compared with current clinical practice alone.

FIGURE 14.

Cost-effectiveness acceptability curve for the primary analysis comparing OncotypeDX and IHC4 with current clinical practice assuming that the test is given only to women with a NPI score > 3.4.