Tumour profiling tests to guide adjuvant chemotherapy decisions in early breast cancer: a systematic review and economic analysis

Aetiology

The causes of breast cancer are not completely understood. A range of risk factors have been identified, including genetic, hormonal and lifestyle factors. 3

It has been estimated that 12% of women with breast cancer have one affected family member and 1% have two or more affected family members. 4 Genetic predisposition is mediated by high-penetrance genes such as BReast CAncer 1 (BRCA1) and BReast CAncer 2 (BRCA2), which are responsible for around 80–90% of hereditary cancers, and low-penetrance genes, which confer both increased and decreased risk. 3

Environmental and lifestyle factors as well as genetic factors influence breast cancer risk. Asian migrants to the West have increased levels of risk compared with the indigenous population, whereas Asian Americans born in the West have incidence rates approximating the US average. 5 Lifestyle and environmental factors thought to increase risk include hormonal factors such as taking the oral contraceptive pill or hormone replacement therapy, higher age of menopause, early age of menarche, late age of first birth and not giving birth. Factors that decrease risk include higher folate intake, higher number of pregnancies, breastfeeding and younger age at first birth. 3 Obesity increases the risk of breast cancer in postmenopausal women. 6 The picture is less clear for premenopausal women, for whom the risk may be lower but prognosis is poorer. Physical activity in adolescence and young adulthood confers a decreased risk of breast cancer,7 which may be mediated hormonally.

Pathology

Breast cancer starts with genetic changes in a single cell or a small group of cells in the epithelia of the ducts or the lobules of the breast. The genetic change allows cells to reproduce uncontrollably, resulting in a tumour. Tumours that have not yet spread to surrounding tissue are known as ‘carcinoma in situ.’ Once it has spread to the surrounding tissue, a tumour is known as ‘invasive’. More rapid growth and spread occurs once a blood supply is secured. Cancer spreads via the lymphatic system or the bloodstream. Lymphatic spread is usually first to the axillary lymph nodes. Spread via the bloodstream can lead to distant metastases in the bone or viscera that are incurable.

The presence or absence of axillary lymph node metastases is a key indicator of disease and prognosis and adjuvant therapy is, in part, planned based on their presence and extent. 8 They are caused by a single cell or a small number of cells detaching from the main tumour, travelling via the lymphatic system and establishing themselves in the tissue of the axillary lymph nodes. Axillary metastases occur in approximately 41% of cases;9 prognosis is better when there is no axillary spread. When metastases are present, axillary clearance is indicated in order to prevent further spread and ensure local disease control.

Prognosis

Overall, the 5-year, age-standardised survival rate for women with breast cancer is 86.3%. 10 Survival varies with age (Figure 1) and stage of disease (Figure 2).

FIGURE 1.

Five-year net survival, by age, for women in England: 2009–13. Adapted with permission from Cancer Research UK. 11

Adapted with permission from Cancer Research UK.

FIGURE 2.

Five-year relative survival, by stage, for women aged 15–99 years in the former Anglia Cancer Network: 2002–6. Adapted with permission from Cancer Research UK. 2

Adapted with permission from Cancer Research UK.

Other factors can also affect prognosis. Clinicians may use tools such as the Nottingham Prognostic Index (NPI),12 Predict (University of Cambridge, Cambridge, UK) or Adjuvant! Online (AOL) (University of Texas Health Sciences Center, San Antonio, TX, USA) to predict disease course and treatment options, although it should be noted that AOL is in the process of being updated and is not currently available. These tools take into account different patient and tumour factors and may give different risk predictions for the same patient.

In general, good prognosis is associated with small tumour size, lymph node-negative (LN0) status, younger age, oestrogen receptor positive (ER+) status and progesterone receptor positive (PR+) status. Overexpression of human epidermal growth factor receptor 2 (HER2) is associated with poorer prognosis.

Current methods for staging of breast cancer

Three main factors are used to stage breast cancer: (1) tumour size, (2) metastases to the regional lymph nodes and (3) distant metastases. The tumour/node/metastases (TNM) staging system was developed and is maintained by the American Joint Committee on Cancer and the Union for International Cancer Control. 17 The T stage is classified in accordance with the size of the tumour and degree of local infiltration, the N stage is classified in accordance with the number and location of metastases to the lymph nodes in the axilla, between the ribs (internal mammary nodes) and above or below the collarbone (supraclavicular and infraclavicular nodes), and the M stage is classified by the presence of metastases beyond the breast and regional lymph nodes (see Table 56, Appendix 1). Early-stage breast cancer is generally defined as cancer that has not spread beyond the breast or the ipsilateral axillary lymph nodes and is confined to stages I, II or IIIA.

Adjuvant! Online

The AOL computer program is designed to provide estimates of the benefits of adjuvant endocrine therapy and chemotherapy. The current version of AOL does not include HER2 status and the potential benefit of trastuzumab. Patient and tumour characteristics are entered into the program and provide an estimate of the baseline risk of mortality or relapse for patients without adjuvant therapy. Information about the efficacy of different therapy options are derived from the EBCTCG meta-analyses in order to provide estimates of reduction in risk at 10 years of breast-cancer-related death or relapse for selected treatments. These estimates are then provided on printed sheets in simple graphical and text formats to be used in consultations. At the time of writing this report (October 2017), AOL was in the process of being updated and was not accessible.

Nottingham Prognostic Index

The NPI is a composite prognostic parameter involving both time-dependent factors and aspects of biological aggressiveness. The NPI score is based on a combination of tumour grade, lymph node involvement and tumour size. To calculate the score, add numerical grade (1, 2 or 3), lymph node score (negative = 1, 1–3 nodes = 2, > 3 nodes = 3) and 0.2 × tumour size in cm. Patients can be divided into three prognostic groups on the basis of the NPI: a good prognosis group (NPI of ≤ 3.4), a moderate prognosis group (3.4 < NPI < 5.4) and a poor prognosis group (NPI of > 5.4).

Predict (version 2.0)

Predict (version 2.0) is an online computer program designed to help women with breast cancer and their doctors make informed decisions about treatment with chemotherapy or endocrine therapy following breast cancer surgery. Predict version 2.0 was developed using data from > 5000 women with breast cancer from England and has been tested on data from another 23,000 women with breast cancer from around the world. Patient and tumour characteristics are entered into the program, which provides an estimate of the overall survival (OS) for patients with or without adjuvant hormone therapy, adjuvant chemotherapy and trastuzumab.

Clinical opinion suggests that there is variation in clinical practice between NHS trusts in the UK, with some centres using single risk prediction tools and others using multiple tools in combination, in addition to other clinical parameters.

EndoPredict (Myriad Genetics)

EndoPredict is a Conformité Européenne (CE)-marked assay that is designed to assess the risk of distant recurrence within 10 years of initial diagnosis. The test is intended for use in premenopausal and postmenopausal women with early-stage breast cancer with all of the following clinical features:

• oestrogen receptor positive

• human epidermal growth factor receptor 2 negative (HER2–)

• lymph node negative (no positive nodes) or lymph node positive (LN+) (up to three positive nodes).

EndoPredict measures the expression of 12 genes: three proliferation-associated genes, five hormone receptor-associated genes, three reference (normalisation) genes and one control gene.

EndoPredict requires RNA samples extracted from FFPE breast cancer tissue. The test can be conducted in a local laboratory using a VERSANT^® kPCR Amplification Detection Module (Siemens Healthcare Diagnostics Inc, Erlangen, Germany). Alternatively, FFPE samples can be submitted to a Myriad Genetics pathology laboratory in Munich that is accredited by the Deutsche Akkreditierungsstelle, a national accreditation body for Germany.

The test process involves using a reverse transcription-quantitative polymerase chain reaction (RT-qPCR), in which target mRNAs are reverse transcribed, amplified and simultaneously detected. The raw data are then exported to online evaluation software (EndoPredict Report Generator; Myriad Genetics Ltd, London, UK), which performs a quality check and calculates the EndoPredict score and the EndoPredict Clinical (EPClin) score. The EndoPredict score is a number on a scale from 0 to 15, is the molecular score only and is not the final test result. An EndoPredict score of < 5 indicates a low risk of distant disease recurrence in the next 10 years. An EndoPredict score of ≥ 5 indicates a high risk of distant disease recurrence in the next 10 years. The EPClin score is calculated by adding clinical data about tumour size and nodal status to the EndoPredict score. From the EPClin score, the probability of metastasis formation within 10 years is estimated, assuming 5 years of hormonal treatment. The EPClin score (cut-off point of 3.3) provides a single low-/high-risk cut-off point; the threshold was set such that women with a low-risk result (EPClin score of < 3.3) have a < 10% risk of developing distant metastases over the next 10 years. It takes approximately 2 days to obtain the test results if the test is done in-house. If samples are sent away for testing, the turnaround time for the central service is 4 to 5 working days.

MammaPrint (Agendia)

MammaPrint is a CE-marked microarray that is designed to assess the risk of distant recurrence within 5 and 10 years and whether or not a woman would benefit from chemotherapy. The test is intended for use in premenopausal and postmenopausal women with stage I or II breast cancer with the following clinical features:

• tumour size of ≤ 5 cm

• lymph node negative or positive (up to three positive nodes).

The test can be used irrespective of ER and HER2 status; that is, it can be used for tumours that are ER negative (ER–) or ER+ and HER2– or human epidermal growth factor receptor 2 positive (HER2+). MammaPrint measures the expression of 70 genes, including genes associated with seven different parts of the metastatic pathway: (1) growth and proliferation, (2) angiogenesis, (3) local invasion, (4) entering the circulation, (5) survival in the circulation, (6) entering organs from the circulation and (7) adaption to the microenvironment at a secondary site. The MammaPrint test is offered as an off-site service. In Europe, samples are sent for analysis at the Agendia laboratory in Amsterdam, the Netherlands. The test requires a FFPE breast cancer tissue sample from a surgical specimen or core needle biopsy.

The test process involves isolation of RNA from a FFPE sample followed by reverse transcription of the mRNA to get complementary DNA (cDNA). The cDNA is amplified and labelled before being hybridised (bound) to the diagnostic microarray. The microarray is washed and then scanned using an Agilent Technologies, Inc. DNA microarray scanner (Santa Clara, CA, USA). The scan file is analysed using Agilent Technologies, Inc. Feature Extraction Software (Santa Clara, CA, USA) and an algorithm is used to calculate the correlation of the sample profile to a ‘low risk’ template profile on a scale of –1.000 to +1.000 with a cut-off point of 0. The threshold was set such that women with a low-risk result have a 10% risk of developing distant metastases over the next 10 years without any adjuvant hormone therapy or chemotherapy. Test results are available to health-care professionals within 10 days of submitting the sample.

Oncotype DX Breast Recurrence Score (Genomic Health)

Oncotype DX is designed to assess the risk of distant recurrence within 10 years and predict the likelihood of benefit from chemotherapy. The test also reports the underlying tumour biology: ER, progesterone receptor (PR) and HER2 status. The test is intended for use in premenopausal and postmenopausal women with stage I or II breast cancer that has the following clinical features:

• lymph node negative or positive (up to three positive nodes)

• oestrogen receptor positive

• human epidermal growth factor receptor 2 negative.

Oncotype DX quantifies the expression of 21 genes. Of these, 16 are cancer-related genes correlated with distant recurrence/relapse-free survival (DRFS) and five are reference genes for normalising the expression of the cancer-related genes. This information is used to calculate the Breast Recurrence Score.

Oncotype DX is offered as a test service to the NHS. Samples are processed centrally at the Genomic Health laboratory in the USA, which is accredited by the American Association for Laboratory Accreditation and the College of American Pathologists. The test requires a FFPE breast cancer tissue sample from a biopsy or surgical resection, which can be sent as a paraffin-embedded block or as 15 unstained charged slides. The test process is based on RT-qPCR. The test gives a recurrence score of between 0 and 100, which is used to quantify the 10-year risk of distant recurrence, assuming 5 years of hormonal (endocrine) therapy. Based on current cut-off points for the oncotype DX test, a score of < 18 indicates low risk of distant recurrence, a score between 18 and 30 indicates intermediate risk and a score of ≥ 31 indicates high-risk. It should be noted that a number of studies, including the ongoing Trial Assigning Individualized Options for Treatment (Rx) (TAILORx) study,22 are testing the use of different cut-off points for oncotype DX. The recurrence score may also predict the benefit of chemotherapy. The oncotype DX results are typically reported within 7–10 days after the sample is received at the laboratory.

The oncotype DX Breast Recurrence Score can be combined with clinical and pathological factors (tumour size, tumour grade and patient age) using the recurrence score–pathology–clinical (RSPC) calculator; however, this calculator has not been validated.

Prosigna (NanoString Technologies)

Prosigna is a CE-marked assay designed to assess DRFS at 10 years. The test is intended for use in postmenopausal women with early-stage breast cancer that is:

• lymph node negative or positive (up to three positive nodes)

• oestrogen receptor positive

• human epidermal growth factor receptor 2 negative.

The test requires RNA extracted from a FFPE breast tumour tissue sample and is done using the nCounter^® analysis system (NanoString Technologies, Inc., Seattle, WA, USA). The test process involves fluorescent probe pairs that hybridise to the mRNA; the fluorescence is then detected by the nCounter Digital Analyzer (NanoString Technologies, Inc., Seattle, WA, USA).

Prosigna is based on the Prediction Analysis of Microarray 50 (PAM50) gene signature. 23 It measures the expression levels of 50 genes used to classify patients into one of four breast cancer subtypes. It also measures the expression of eight housekeeping genes used for signal normalisation, six positive controls and eight negative controls. Prosigna classifies samples into the following breast cancer subtypes based on their PAM50 gene expression signatures: luminal A, luminal B, HER2-enriched or basal-like. A ROR score, representing the risk of distant recurrence within 10 years (assuming 5 years of hormonal treatment), is then derived from an algorithm based on the results of the PAM50 gene signature plus clinicopathological factors. Four versions of the ROR score exist in the research setting: (1) ROR based on PAM50 subtype information (ROR-S), (2) ROR-S based on PAM50 information plus proliferation score (ROR-P), (3) ROR-S plus tumour size (ROR-T or ROR-C) and (4) ROR-S plus proliferation score and tumour size [PAM50 subtype call, proliferation score and ROR score (ROR-PT)]. The proliferation score is based on a subset of the PAM50 genes associated with the proliferation pathway.

The Prosigna test uses ROR-PT and therefore includes the PAM50 breast cancer subtype, tumour size and proliferation score. Nodal status is also used in converting the score into a risk category. The ROR score is a numerical score on a scale of 0 to 100. Based on this score and the nodal status, samples are classified into risk categories:

• node negative: low risk (score of 0–40), intermediate risk (score of 41–60) or high risk (score of 61–100)

• node positive (up to three positive nodes): low risk (score of 0–15), intermediate risk (score of 16–40) or high risk (score of 41–100).

This assessment includes all studies assessing ROR-PT, whether they use the commercial Prosigna test (using the nCounter system) or other methods (such as RT-qPCR). However, studies assessing ROR-S (subtype), ROR-T/ROR-C (subtype and tumour size) or ROR-P (subtype and proliferation score) are excluded. Studies are also excluded if they only assess PAM50 breast cancer subtypes (luminal A, etc.) rather than ROR-PT score.

IHC4 test

Immunohistochemistry 4 is a laboratory-developed test that combines the results of four IHC-measured parameters. The test can be combined with clinical and pathological features; this is known as IHC4 plus clinical factors (IHC4+C). The test is designed to quantify the risk of distant recurrence in breast cancer patients, assuming 5 years of endocrine therapy. The test is intended for use in postmenopausal women with early-stage breast cancer with the following clinical features:

• oestrogen receptor positive

• lymph node negative or positive (up to three positive nodes).

The components of the test are four immunohistochemical assays: ER, PR, HER2 and the proliferation marker Ki-67. The IHC4 test is currently used within the Royal Marsden Breast Cancer Unit service, but it has been suggested that the test could be run in local NHS laboratories if quality assurance programmes for the individual assays were in place. IHC4 uses FFPE breast tumour tissue samples and IHC techniques that are universally available in NHS pathology departments. ER and HER2 markers are commonly measured in NHS laboratories, although methods for quantifying these markers in the format required for the test may not be routinely available. Although PR and Ki-67 markers are not routinely measured in breast tumour tissue samples, the assays are commonly available for use if needed. The quantitative assessment of Ki-67 required for the IHC4 test is not currently conducted in most NHS laboratories and, therefore, further training for pathologists and biomedical scientists is likely to be needed.

The IHC4+C test involves an algorithm that calculates a risk score for distant recurrence based on the results of the four IHC assays and clinical factors including age, nodal status, tumour size and tumour grade. The algorithm has been published and validated24,25 and is freely available, and a calculator is available for use on request. A distant recurrence score of < 10% is categorised as low risk for distant recurrence at 10 years, a score of ≥ 10% but < 20% is categorised as intermediate risk and a score of ≥ 20% is categorised as high risk for distant recurrence at 10 years. At the Royal Marsden NHS Foundation Trust, the test is processed with an average estimated turnaround time of 1 week; however, results may be made available in 2 working days if they are urgently required.

Electronic database searches

The search strategy comprised medical subject headings (MeSHs) or Emtree thesaurus terms and free-text synonyms for ‘breast cancer’ combined with the individually named tumour profiling tests. Searches were translated across databases and were not limited by language. Searches for oncotype DX, MammaPrint, IHC and Prosigna were limited by publication date from 2011 (the search date in the review by Ward et al. ,1 because these tests were included in this review), whereas no date limits were applied to EndoPredict (as it was not included in the review by Ward et al. 1).

The search strategies are presented in Appendix 2. Literature searching was undertaken in February 2017 in the following electronic databases and trials registries:

• MEDLINE Epub Ahead of Print, In-Process & Other Non-Indexed Citations (via Ovid): 1946 to present.

• EMBASE (via Ovid): 1974 to 24 February 2017.

• Cochrane Database of Systematic Reviews (via Wiley Interscience): 1996 to present.

• Database of Abstracts of Reviews of Effects (via Wiley Interscience): 1995 to 2015 (until close of database).

• Cochrane Central Register of Controlled Trials (via Wiley Interscience): 1995 to present.

• Health Technology Assessment (HTA) Database (via Wiley Interscience): 1995 to 2016 (until close of database).

• NHS Economic Evaluation Database (via Wiley Interscience): 1995 to 2015 (until close of database).

• Science Citation Index Expanded (via Web of Science): 1900 to present.

• Conference Proceedings Citation Index – Science (via Web of Science): 1990 to present.

• World Health Organization International Clinical Trials Registry Platform (http://apps.who.int/trialsearch/; accessed 19 January 2017) (no date limit applied).

• American Society of Clinical Oncology (ASCO) (via Web of Science) (date range searched: 2011–17).

• European Society for Medical Oncology (via Web of Science) (date range searched: 2011–17).

Supplementary searches

References of relevant systematic reviews, primary studies and company submissions were checked to identify additional studies.

Population and setting

The intended population included people with ER+ (and/or PR-positive), HER2–, early-stage breast cancer (stages I or II) with LN0–3.

In practice, it was anticipated that many potentially relevant studies would include a broader population. Therefore, all relevant studies of early-stage breast cancer were eligible for inclusion. When subgroups were reported for the intended population (described above), these were used in the assessment. When no subgroups were reported, the study was included and the findings were interpreted with reference to how closely the study population matched the intended population.

The following subgroups were considered within this assessment:

• people with LN0 cancer, people with micrometastases in the lymph nodes and people with LN1–3

• premenopausal and postmenopausal women

• people predicted to be at low, intermediate or high risk using a risk assessment tool or using clinical and pathological features

• males and females

• people of different ethnicities.

This assessment focuses on the use of tumour profiling tests to guide decisions about adjuvant chemotherapy. The use of these tests to guide endocrine therapy decisions, or decisions about neoadjuvant chemotherapy (to shrink the tumour before surgery), was not evaluated.

Interventions

The following tumour profiling tests were included:

• EndoPredict and EPClin

• MammaPrint

• oncotype DX Breast Recurrence Score and oncotype DX Breast RSPC

• Prosigna (or ROR-PT, which is equivalent)

• IHC4 and IHC4+C.

Comparators

The comparator for the assessment is standard UK practice for chemotherapy decision-making. This was taken to include combinations of clinicopathological factors (e.g. within multivariable models), plus clinicopathological risk tools used in the UK, including Predict, the NPI and AOL. The clinical treatment score (CTS), a combination of commonly used clinicopathological variables, was also included as a comparator even though it is not commonly used in practice as a tool, because it is used in a number of key studies and includes a set of variables that are used in practice.

Other non-UK local or national guidelines, such as St Gallen and the National Comprehensive Cancer Network (NCCN) guidelines, were excluded when a study also reported comparisons with Predict, NPI or AOL, but were otherwise included.

Relevant comparators within individual studies varied in accordance with the study type:

• Studies assessing prognostic performance. No comparator is needed as the aim is to compare outcomes between risks groups for the test being studied.

• Studies assessing prediction of chemotherapy benefit. No comparator is needed as the aim is to compare the effect of chemotherapy between risks groups for the test being studied.

• Clinical utility studies. The relevant comparator is standard clinical practice as defined in the first paragraph of this section.

• Decision impact studies. The relevant comparator is standard clinical practice as above (for pre-test chemotherapy decisions).

Outcomes and study designs

The clinical effectiveness review considers the clinical effectiveness of the tests in relation to the following broad categories:

• Analytic validity (i.e. the ability of the test to accurately and reliably measure the expression of mRNA or proteins by breast cancer tumour cells). Owing to time constraints, it was not possible to conduct a full review of analytic validity for all tests. A rapid review of IHC4 was conducted.

• Prognostic performance (i.e. the degree to which the test can accurately predict the risk of an outcome such as disease recurrence and discriminate patients with different outcomes). This is usually assessed by conducting the test on stored tumour samples for which longer-term patient outcome data are available, but when the test did not influence treatment. Study designs include –

  • Reanalysis of randomised controlled trial (RCT) data.

  • Analysis of prospective or retrospective cohorts in which the test did not influence treatment.

• Prediction of chemotherapy benefit (i.e. whether or not the effect of chemotherapy vs. no chemotherapy on patient outcomes differs between test risk groups). This is usually assessed by conducting the test on stored tumour samples for which longer-term outcome data are available. Study designs include –

  • Reanalysis of RCTs in which some patients received chemotherapy and some did not.

  • Analysis of prospective or retrospective cohorts in which some patients received chemotherapy and some did not. These could include cohorts in which the test did or did not influence practice.

• The difference between absolute and relative benefit should be noted: for a test that is prognostic, a difference in absolute benefit of chemotherapy between groups would be expected, whereas for a test to be considered predictive of chemotherapy benefit, a difference in relative benefit should be seen. As an example, if distant recurrence rates in the test high-risk group were 30% without chemotherapy and 20% with chemotherapy, the absolute benefit of chemotherapy would be 10%. Likewise, if distant recurrence rates in the test low-risk group were 3% without chemotherapy and 2% with chemotherapy, the absolute benefit of chemotherapy would be 1% (i.e. much smaller). However, the relative benefit would be the same in both groups [relative risk (RR) of 0.67, i.e. chemotherapy reduces recurrence by one-third]. If the test is predictive of chemotherapy benefit, the RR would be expected to be different in different risk groups.

• Clinical utility: this is defined differently throughout the prognostic literature. Here, we define clinical utility studies as those that assess the ability of the test to affect patient outcomes (such as recurrence and survival) through the prospective use of the test to guide treatment decisions (the study may be prospective or retrospective, but use of the test should have been prospective, i.e. used in clinical practice rather than conducted on stored tumour samples). Study designs include –

  • RCTs randomising patients to chemotherapy guided by the test or guided by a comparator (e.g. clinical practice).

  • Observational studies reporting clinical outcomes for patients whose treatment was guided by the test. As these studies do not have a comparator, we are primarily interested in outcomes for patients with low-risk disease, who, as a group, have mostly avoided chemotherapy. The observation of good outcomes in this group could, alongside other evidence, support the avoidance of chemotherapy in this group.

• Decision impact (i.e. how the tests influence decision-making in terms of which patients will be offered chemotherapy). Clinical advice to the External Assessment Group (EAG) suggests that chemotherapy rates differ between countries, with lower rates in the UK and Europe than in the USA. The review therefore included only UK and European studies. Study designs include –

  • Studies assessing change in chemotherapy recommendations and/or decisions before and after use of the test (this design does not include follow-up of clinical outcomes such as recurrence or survival).

• Health-related quality of life and anxiety. Study designs include –

  • Studies assessing impact of the test versus usual practice on HRQoL and anxiety.

  • Studies assessing HRQoL and anxiety before and after test use.

• Time-to-test results: studies assessing the time taken to obtain test results.

• Concordance: concordance is defined in this review as the degree to which tests assign the same patients to the same risk groups. Such studies do not report long-term outcomes. A full systematic review of studies that only assess concordance between tests (with no patient outcome data) was beyond the scope of this assessment. However, the Optimal Personalised Treatment of early breast cancer usIng Multi-parameter Analysis preliminary (OPTIMA Prelim) study30 was included as a key example of concordance between tests.

Study characteristics: oncotype DX

Oncotype DX was validated in 11 distinct data sets. Seven were reanalyses of RCTs [the National Surgical Adjuvant Breast and Bowel Project (NSABP) B14,49 B2050 and B2851,52 trials; the Southwest Oncology Group (SWOG) trial 8814;53 the Eastern Cooperative Oncology Group E219754,55 trial; UK patients from the TransATAC37,46 trial; and a French trial, PACS0156] and four were retrospective cohort studies (one from the USA,57 two from China58,59 and one from Japan60). All studies recruited patients with either ER+ or HR+ tumours, but only TransATAC37,46 and one Chinese study58 recruited or reported a subgroup of exclusively HER2– patients. Two studies (TransATAC37,46 and the SWOG 8814 trial)53 recruited only postmenopausal patients.

The test was designed to predict outcomes in patients who received 5 years of endocrine therapy without chemotherapy. Only three studies treated patients with endocrine monotherapy, although it was not always clear if this was for 5 years; of these, one study recruited mixed lymph node status patients and reported separate analyses for LN0 and LN+ patients,37,46 one recruited LN0 patients49 and one recruited LN+ patients. 53 Of the remaining studies, one treated some patients with endocrine monotherapy and some with endocrine therapy and chemotherapy (LN0 patients50), two treated all patients with endocrine therapy and chemotherapy (one study recruited mixed lymph node status patients54,55 and one recruited LN+ patients51,52) and one treated all patients with chemotherapy and some with endocrine therapy (LN+ patients). 56 The retrospective studies treated patients in accordance with usual practice (without oncotype DX) with varying levels of endocrine therapy and chemotherapy. 57–60 The total number of patients included was 4929. A detailed narrative synthesis of study characteristics is provided in Report Supplementary Material 2. Study characteristics data are presented in Report Supplementary Material 2, Table 1.

Two studies did not report how oncotype DX was conducted (PACS01 study56 and Russell et al. 57). In all but three other cases, the test was conducted on fixed, paraffin-embedded tissue by Genomic Health using the commercial oncotype DX assay. The three exceptions were the two studies from China in which the test was not carried out by Genomic Health58,59 and the study by Paik et al. ,49 as Paik et al. (2006)50 described the assay used in Paik et al. (2004)49 as being ‘a preliminary version of the RT-PCR assay (lacking standardized reagents, calibrators, and controls)’. In these three studies, the equivalence of the tests to the commercially offered oncotype DX assay is unknown.

Quality assessment: oncotype DX

Quality assessment is summarised in Report Supplementary Material 2. All studies were validation studies. Only three studies37,46,49,53 used an appropriate study design, as eight50–52,54–59 included patients who had been treated with chemotherapy or did not report the proportion of patients treated with chemotherapy. Undertreatment with endocrine therapy and overtreatment with chemotherapy both have the potential to affect recurrence and may alter the observed HRs for outcomes between risk groups. No studies included all eligible patients and only three37,46,53,54 stated that they blinded test assessors to patient outcomes. A lack of blinding is likely to have a low impact as the test is objective. There are concerns about patient spectrum bias in all studies, mainly owing to the retrospective nature of the studies and the exclusion of tumour samples with insufficient tissue probably leading to the loss of patients with smaller tumours. The potential loss of patients with small tumours is of unknown concern, as it is unknown whether or not the test would have a different prognostic performance in these patients.

Results: oncotype DX

The following is a summary of key results from the review. A detailed narrative synthesis of all study results is provided in Report Supplementary Material 2.

Distribution of patients by risk group: oncotype DX

Data are presented in Table 3. The proportion of patients classified as being at low risk ranged from 48%60 to 64%46 in LN0 cohorts and was generally lower, ranging from 36%51,52 to 57%,46 in LN+ cohorts. The proportion of patients who were classified as being at intermediate risk ranged from 20%60 to 27%46 in LN0 cohorts, and was generally higher in LN+ cohorts, ranging from 30%56 to 34%. 51,52 The proportion of patients who were classified as being at high risk ranged from 9% to 33% in LN0 patients and was similar among LN+ patients, ranging from 11% to 32%. The number of patients who are likely to be prescribed chemotherapy on the basis of their test results will, to a large extent, depend on how intermediate-risk patients are handled and whether or not they would be handled in the same way in the LN0 and LN+ groups.

Prognostic performance, unadjusted analyses: oncotype DX

Data are presented in Table 3. The 10-year DRFI rates in low-risk LN0 patients treated with endocrine monotherapy ranged from 93% to 97% (three studies46,49,60), and was similar when 100% patients received endocrine therapy and chemotherapy (96%, one study50). LN+ patients had much lower 10-year DRFI rates: 82% (one study46) with endocrine monotherapy and 81% (one study51,52) when 100% of patients received endocrine therapy and chemotherapy. The 10-year DRFI rates in LN0 intermediate-risk patients treated with endocrine monotherapy ranged from 86% to 100% (three studies46,49,60), and was similar when 100% of patients received endocrine therapy and chemotherapy (89%, one study50). LN+ intermediate-risk patients had much lower 10-year DRFI rates: 75% (one study46) with endocrine monotherapy and 65% (one study51,52) when 100% of patients received endocrine therapy and chemotherapy. The 10-year DRFI rates in LN0 high-risk patients treated with endocrine monotherapy ranged from 70% to 77% (three studies46,49,60), and was higher when 100% of patients received endocrine therapy and chemotherapy (88%, one study50). LN+ high-risk patients treated with endocrine monotherapy had similar 10-year DRFI rates to LN0 patients, at 69% (one study46); in studies in which all patients received endocrine therapy and some or all received chemotherapy, the DRFI was lower, at 56% (one study51,52). All the DRFI rates in this paragraph exclude one study from China, which appeared an outlier with very low DRFI rates (Sun et al. 59). The study from Japan60 also reported some unusual results in that intermediate-risk patients had better DRFI than low-risk patients (e.g. 10-year DRFI rate was 97% and 100%, respectively). It is unclear if, for both of these studies,59,60 the unusual results are due to small sample sizes (n = 9859 and n = 20060), differences in treatment practices in Japan and China compared with Western countries or differences in ethnicity.

Despite confounding from treatment, the studies reporting prognostic performance data reported largely statistically significant differences between high-risk and low-risk groups, whether through HRs or through analyses of event rates per group, and this was the case regardless of lymph node status. However, differences between the intermediate-risk group and the high- or low-risk groups were not always statistically significant, particularly in the LN+ population (see Table 3).

Two studies reported a C-index (AUC). One study54,55 was in LN+/LN0 patients and the other was in LN0 patients. 58 In both cases, the C-index was 0.69, which indicated that the model was better than chance at placing patients into appropriate risk categories and nearly reaches the cut-off point for a ‘good’ test of 0.7. 48

Data for OS, BCSS, RFS and RFI can be found in Report Supplementary Material 2. These data support observations for DRFS, DRFI and DFS.

Additional prognostic value, adjusted analyses: oncotype DX

Data are presented in Table 4. The analyses that reported multivariable Cox proportional hazards models that were adjusted for clinicopathological variables generally indicated that the prognostic performance of oncotype DX had additional benefit over these factors, as HRs remained significant in most analyses (the exception being RFS and OS analyses by Toi et al. 60). This was consistent regardless of lymph node status and variables adjusted for, which included age, tumour size and LN status (when applicable) in all cases and grade in most cases (Toi et al. 60 and Sun et al. 59 being the exceptions). However, covariates included in multivariate analyses varied, and it is not clear if all important covariates were included in all analyses. The use of the 50-point difference in the adjusted analyses of prognostic performance indicated that the recurrence score may be prognostic over clinicopathological factors, but does not provide information about the clinical significance of the 18–30 recurrence score cut-off points.

The likelihood ratio χ² in the TransATAC data set was statistically significantly higher for oncotype DX than for CTS or NPI only for LN0 patients [10-year DRFI likelihood ratio χ² 10.64 (p = 0.001) and 8.82 (p = 0.003), respectively], while the difference was not statistically significant for LN+ patients [3.56 (p = 0.06) and 2.14 (p = 0.1), respectively]. 46 Compared with AOL, and with a model based on AOL but with 5-year outcomes, oncotype DX also appeared to provide additional prognostic information (see Table 4). 49,61

Oncotype DX RSPC

Data are presented in Table 5. One study (Tang et al. 44) derived the RSPC score in a meta-analysis of NSABP B-14 and TransATAC (LN+/LN0; n = 1735), and performed a limited validation in NSABP B-20 (LN0; n = 625), which included 233 patients used to derive the oncotype DX recurrence score. For this reason, both NSABP B-14 and B-20 data were used to derive part of the algorithm. Based on the NSABP B-14 analysis set, the oncotype DX RSPC algorithm (oncotype DX plus age, tumour size and grade) appeared to provide additional prognostic information over oncotype DX and over clinicopathological variables, and was able to classify more patients into a low-risk category than oncotype DX while maintaining a roughly equivalent rate of distant recurrence in the low-risk group. In the NSABP B-20 cohort, RSPC had prognostic value in a univariate analysis.

Conclusion: oncotype DX and RSPC prognostic performance

Seven reanalyses of RCTs and four retrospective cohort studies were included, with a total of 4929 patients. The generalisability of the evidence base to the decision problem is uncertain owing to the loss of patients with insufficient tumour samples. Generally, when comparing LN0 patients with LN+ patients, similar numbers were at high risk, but more were at low risk in LN0 cohorts, and more at intermediate risk in LN+ cohorts. How many patients would be prescribed chemotherapy would depend in large part on how intermediate patients are handled. The 10-year DRFI rates suggest that patients in the LN0 low-risk group46,49,60 are at very low ROR (10-year DRFI range 93–97%) in the absence of chemotherapy, and patients in the intermediate risk group may be at somewhat higher risk (10-year DRFI range 86–100%). In the LN+ study using endocrine monotherapy, patients were generally at higher risk of recurrence than LN0 patients in both the low and the intermediate categories (10-year DRFI < 85% and ≤ 75%, respectively). Unadjusted analyses indicated that oncotype DX had prognostic power (statistically significant differences between low-risk and high-risk groups) across various recurrence outcomes, regardless of lymph node status. HRs between the intermediate-risk group and the high- or low-risk groups were not always statistically significant. Oncotype DX provided additional prognostic information over the most commonly used clinicopathological variables (age, grade, size and nodal status) regardless of lymph node status (although it was not always clear if all relevant variables had been included in the analyses), and over CTS and NPI in LN0 (but not LN+) patients. On the basis of proportions classified as low risk and DRFI rates, RSPC may outperform oncotype DX in LN0 patients, but these data are from the derivation cohort, with only limited validation data from a cohort (NSABP B-20) that included patients who were used to derive one of its constituent parts (oncotype DX Breast Recurrence Score), and it has not been tested in premenopausal or LN+ patients.

Study design and patient characteristics

Five data sets, reported across 10 published references44,50,53,61,64–69 and one academic-in-confidence (AiC) manuscript (see Report Supplementary Material 3, Table 3), have conducted analyses that assess the ability of oncotype DX to predict the benefit of chemotherapy. Two were reanalyses of RCTs: the American (NSABP) B-20 study (n = 651) in LN0 was reported by Paik et al. ,50 Tang et al. 61 and Tang et al. ;44 and the American SWOG 8814 study (n = 367) in LN+ patients (38% had four or more lymph nodes) was reported by Albain et al. 53 In both trials, patients were randomised to tamoxifen only, or tamoxifen plus cyclophosphamide. NSABP-B20 did not select participants by HER2 status, but an analysis of HER2– patients was provided as a personal communication (Professor Tang via NICE, University of Pittsburgh, personal communication, February 2018) as part of the NICE assessment process. SWOG-881453 recruited patients with any HER2 status (12% were HER2+) and all patients were postmenopausal. Both reanalyses had high attrition rates (72% and 60%, respectively), owing to missing clinical variables, missing samples, insufficient tissue and test failures.

The three remaining studies were observational64–67,70 (total approximately 44,000 with some double-counting). Two LN0 studies were from the USA [MD Anderson Center, n = 1424;64,65 Surveillance, Epidemiology, and End Results (SEER) registry, n = 40,13467] and one study recruiting a mix of LN+ and LN0 patients was from Israel [Clalit Health Services one lymph node micrometastasis (LN1micro)–LN3, n = 620; LN0–LN1micro, n = 159466]. Additional analyses were provided to the EAG as AiC data but cannot be reported here. Patients were treated in accordance with local routine practice in conjunction with their oncotype DX score.

Quality assessment

Report Supplementary Material 3, Table 4, presents the quality assessment of the included studies. Among the RCTs, neither trial scored well on every item. The key concerns included the use of the derivation cohort in Paik et al. 50/Tang et al. 61 and high attrition rates in both trials. The comparison of baseline characteristics between patients included in the analysis and those excluded from the analysis showed differences in patient age at recruitment (Paik et al. 49) and tumour grade. (Paik et al. 50).

The three observational studies64–67,69 are limited by their non-randomised design, whereby patients who received chemotherapy are likely to be systematically different in terms of known (and potentially unknown) prognostic variables (e.g. age) and treatment effect modifiers to those who did not, leading to a high risk of confounding. They also only recruited patients for whom an oncotype DX test had been ordered and it is unclear how this may have affected the patient spectrum and generalisability to the decision problem. However, owing to their prospective use of the test in clinical practice, three studies blinded the test assessors to the long-term outcomes. 64–69

Results

Both RCT reanalysis studies showed that unadjusted HRs for the effect of chemotherapy versus no chemotherapy on survival and recurrence outcomes were most favourable in the higher-risk groups. HRs were generally statistically significant in high-risk groups but not in low- or intermediate-risk groups. In the B20 study (LN0 patients),44,50,61 unadjusted HRs for 10-year DRFI in the low-, intermediate- and high-risk groups were 1.31, 0.61 and 0.26, respectively. HRs restricted to HER2– patients (adjusted and unadjusted) showed the same pattern (Table 6; Professor Tang, personal communication). However, it is interesting to note that absolute differences (for chemotherapy vs. no chemotherapy) were very small in the low- and intermediate-risk groups (1.1% and 1.8%, respectively, both favouring no chemotherapy), although they were greater in the high-risk group (27.6% favouring chemotherapy).

In SWOG-8814 (LN+),53 DRFI was not reported. HRs for 10-year DFS for low-, intermediate- and high-risk groups, adjusted for number of positive nodes, were 1.02, 0.72 and 0.59, respectively.

Unadjusted interaction tests were statistically significant for 10-year DRFI and OS in NSABP B-20 (LN0) (p = 0.031 and p = 0.011, respectively). 50,61 Albain et al. 53 (LN+) found that the effect of recurrence score on treatment varied over time and that recurrence score is a treatment effect modifier in the first 5 years (interaction p-value 0.029) but not after 5 years (interaction p-value 0.580).

Interaction tests adjusted for clinicopathological factors in NSABP-B20 were borderline significant for the full cohort for DRFI (p = 0.035, p = 0.039 and p = 0.068 for different methods of assessing tumour grade),50 (Professor Tang, personal communication), whereas for the HER2– subgroup, the equivalent analyses were statistically significant (p = 0.007, p = 0.018 and p = 0.022) (Professor Tang, personal communication). The EAG report71 stated that it was unclear whether or not all factors were adjusted for simultaneously in B20; however, a personal communication with the biostatistician (Professor Tang, personal communication) confirmed that this was the case. Interaction tests in SWOG-8814 adjusted individually for each of age, ethnicity, tumour size, grade, PR, P53 and HER2 were also statistically significant (p = not reported). Initially, the EAG interpreted this as a model including all clinicopathological variables; however, clarification from the authors in a personal communication (Professor Barlow, University of Washington School of Public Health, personal communication, March 2018) stated that each variable was included in a separate model. An interaction test adjusted for Allred-scored ER status was not significant (p = 0.15). No interaction test that included all clinicopathological variables together was available.

The oncotype DX cut-off point below which chemotherapy could be avoided was reported to be approximately 20 in SWOG-8814,53 but NSABP B-20 authors could not determine a cut-off point as there was no point below which chemotherapy did not confer an advantage. 50,61

The analyses and available data were subject to some criticisms in the context of this decision problem. First, it was not clear whether or not all stratification factors used in randomising patients to treatment were included in the interaction test model. Second, categorising the continuous oncotype Breast Recurrence Score into risk groups may lead to loss of information and has the potential to create spurious interactions between recurrence score and chemotherapy benefit due to imbalances in clinicopathological variables between risk groups. Third, none of the analyses were conducted in the clinically-intermediate-risk patient group of interest to the decision problem; in particular, it is plausible that even if there is no chemotherapy benefit for clinically-low, oncotype DX-low patients, there could be benefit for clinically-intermediate (NPI of > 3.4) oncotype DX-low patients. Fourth, patients from the no-chemotherapy arm of the B20 study were used to derive the oncotype DX score. Therefore, oncotype DX may be overfitted in this study arm (i.e. recurrence rates may be artificially low in oncotype low-risk patients and artificially high in oncotype DX high-risk patients). This could lead to an overestimate of chemotherapy benefit because the chemotherapy arm was not used in derivation; therefore, recurrence rates in this arm may show less separation between the low- and high-risk groups. A more thorough discussion of this and other quality issues is provided in Report Supplementary Material 3.

Other potential biases in the reanalyses of RCTs included attrition of samples, exclusion of patients owing to missing data for covariates and inclusion of HER2+ patients (applies only to SWOG-8814 LN+ patients), who are out of the scope of this assessment.

Observational studies: results

Data are presented in Table 7. From the three observational cohort studies,64–67,69,70 evidence was mixed and at high risk of confounding, because patients who received chemotherapy were likely to be at higher risk of recurrence than patients who did not. Only one study reported an interaction test between recurrence score categories and chemotherapy treatment, and this was statistically significant (p = 0.03), but only adjusted for grade, tumour size, age and race (omitting ER and PR),67,70 and for recurrence score as a continuous variable (p < 0.001). The other two studies only reported HRs for chemotherapy versus no chemotherapy in intermediate- and high-risk patients,64–66,69 and in one study these were statistically non-significant, even after adjustment for confounders. 64,65 In the other study,66,69 patients with LN0/LN1micro did not appear to receive benefit from chemotherapy in the intermediate group (DRFI for chemotherapy vs. no chemotherapy: 94.4% vs. 94.7%, respectively), whereas those in the high-risk group did (DRFI for chemotherapy vs. no chemotherapy: 86.7% vs. 78.9%, respectively). In the LN1micro–LN3 subgroup, patients did appear to receive benefit from chemotherapy in the intermediate-risk group (DRFI for chemotherapy vs. no chemotherapy: 97.8% vs. 90.4%, respectively); data for high risk patients were not reported. A more detailed narrative is provided in Report Supplementary Material 3.

RSPC results

Recurrence score–pathology–clinical results were derived in TransATAC and NSABP B-14,44 and validated in NSABP B-20. 44 Only LN0 data were available. An interaction test was non-significant (p = 0.10), with a standardised HR of 0.65 (95% CI 0.39 to 1.09) (data not tabulated). 44

In practice, it is unlikely that chemotherapy decisions would be made on oncotype DX scores independent of clinicopathological variables. Evidence relating to the ability of the test to predict chemotherapy benefit over and above routinely collected clinicopathological variables was provided in both RCT data sets in the adjusted interaction tests. 50,53,61 Interestingly, Tang et al. 61 tested the ability of AOL to predict benefit from chemotherapy in a large cohort of 1952 patients, and found it to have predictive ability for OS. However, the inclusion of clinicopathological variables alongside recurrence score in the RSPC algorithm resulted in a loss of predictive ability (p = 0.10), indicating that the incorporation of clinicopathological factors may reduce prediction of chemotherapy benefit; therefore, if chemotherapy decisions are based on an informal consideration of clinicopathological factors alongside the oncotype DX score, this may reduce any predictive ability of oncotype DX in clinical practice.

Conclusion: oncotype DX and RSPC chemotherapy benefit

In conclusion, there is some evidence from two reanalyses of RCTs to suggest that oncotype DX may predict benefit from chemotherapy, and that benefit from chemotherapy is highest in oncotype DX high-risk patients. Unadjusted interaction tests between oncotype DX risk group and chemotherapy benefit were mainly statistically significant. Adjusted interaction tests were borderline significant in the LN0 NSABP B20 study (significant in HER2– patients), whereas in the LN+ SWOG-8814 study they were significant when adjusted for some clinicopathological variables individually, but not when adjusting for ER determined by Allred status. Only data from the derivation cohort (NSABP B-20) were available for LN0 patients and this may have biased the results. The RSPC algorithm (oncotype DX plus age, tumour size and grade) showed a non-significant interaction test between chemotherapy benefit and RSPC risk group, and also used the NSABP B-20 data set. Three observational cohort studies were at high risk of confounding; one reported a statistically significant interaction test but this was only adjusted for limited factors. If predictive ability was assumed, it is unclear below which exact cut-off point patients could avoid chemotherapy (although one study suggests that this is a recurrence score of 20), as chemotherapy benefit is uncertain in the intermediate-risk group. Although TAILORx22 (an ongoing RCT comparing long-term outcomes in patients treated with usual care vs. patients treated with usual care in conjunction with oncotype DX) will address the issue of whether or not low- and intermediate-risk patients can avoid chemotherapy, it is unclear to what extent it will address the question of whether or not the test can predict chemotherapy benefit. Considering the limitations of the available data, the EAG concludes that there remains uncertainty surrounding whether or not oncotype DX is associated with a predictive benefit of chemotherapy (i.e. a difference in relative effect by genomic risk group) and, if so, that there is uncertainty in the likely magnitude of this predictive effect within the clinical subgroups considered in this appraisal.

The TAILORx study22 reported during the course of the assessment, and included some data relating to chemotherapy benefit prediction. The EAG prepared a preliminary appraisal of this evidence, which can be accessed via the NICE website. 72

Study and patient characteristics

Five data sets reported across nine publications22,65,66,68,69,73–75 and one AiC manuscript were included. Study characteristics are presented in Report Supplementary Material 4. Two studies had a prospective trial design (within RCTs),22,73,74 and the remainder were retrospective analyses of observational data. One further study, the SEER registry,67,68 did not meet the inclusion criteria for the review in that only survival outcomes with < 5 years follow-up were reported, but it reported useful data on certain subgroups of interest to the review (micrometastases and race) and is presented here due to the paucity of other data relating to there characteristics.

Only one study (TAILORx)22 randomised patients (HR+, HER2– and LN0) to treatment guided by the test or treatment in accordance with usual practice. Women with recurrence scores of < 11 were assigned to endocrine therapy only, whereas women with recurrence scores of 11–25 were randomised to either endocrine therapy plus chemotherapy or endocrine therapy only. As of July 2017, this study had only reported results for the low-risk (recurrence score of < 11) group (n = 1626). Data for this group are effectively prospective observational data. The study reported in full in June 201876 and the EAG prepared a preliminary appraisal of this evidence, which can be accessed via the NICE website. 72

The West German Study Group (WSG) PlanB73,74,77 trial (n = 3198) is also a prospective RCT, but does not aim to assess the clinical utility of oncotype DX, as it randomises clinically high-risk patients {pathological tumour stage (pT)1–T4c; LN+, or LN0 with a risk factor [≥ pT2, grade 2/3, high urokinase plasminogen activator (uPA)/plasminogen activator inhibitor-1 (PAI-1), aged < 35 years or HR-negative]} with 0–3 positive lymph nodes with a recurrence score of ≥ 12 to two different sorts of chemotherapy. However, a translational research aim was to assess the ROR in patients with recurrence scores of < 12 who were not treated with adjuvant chemotherapy. This group is, again, effectively a prospective observational cohort.

There were three retrospective cohorts in which patients were treated using oncotype DX,65,66,69,75 and one further retrospective cohort that did not meet the inclusion criteria for the review. 67,70 The total number of patients included in these retrospective analyses is ≈54,000 (some double-counting from the Clalit Health Services cohort). 66,69 Patients were ER+, HER2– and must have had an oncotype DX test (it was unclear how patients were selected for testing, and this may have introduced spectrum bias). Two studies recruited only LN0 to lymph node micrometastases (LNmicro) patients,65,75 and one reported LN0-LNmicro69 and patients with micro metastases or between one and three lymph node metastases (LNmicro–LN3). 66,69 The study that did not meet the inclusion criteria recruited patients with LN0–3, and is included as it subgrouped patients in accordance with age (40–85 years), lymph node status (LN0, LNmicro–LN3, LNmicro only) and race (black, white, other). 67,70

Quality assessment

The highest level of evidence for clinical utility is a RCT of treatment guided by the test versus treatment guided in accordance with usual practice. Assessment with the Cochrane risk-of-bias tool for RCTs indicates that all studies are of too poor quality to meet this aim.

Results: oncotype DX clinical utility

Data relating to the clinical utility of oncotype DX are presented in Table 8. A more detailed narrative synthesis is provided in Report Supplementary Material 4. All studies report data relating to recurrence or survival, but differences in cut-off points (recurrence score of < 11, < 12 and < 18), patient populations (clinically high risk, LN0 or LN+), treatment regimens (some patients had chemotherapy in some studies) and outcome measures (DRFS, DFS, DRFI, BCSS and OS) precluded a meaningful meta-analysis.

Across the studies in which chemotherapy was delivered in accordance with usual practice using oncotype DX, chemotherapy rates in patients at low risk (recurrence score of < 18) ranged from 1%69 to 12%75 (four studies65,67,69,70,75) in LN0 patients and from 7% to 23% (two studies66,67,70) in LN+ patients. In intermediate-risk (recurrence score of 18–30) patients, chemotherapy rates ranged from 26%69 to 43%65 (three studies65,69,75) in LN0 patients and from 40%66 to 47%67,70 (two studies66,67,69) in LN+ patients. These data perhaps indicate that lymph node status was considered in treatment decisions, although no formal comparison has been made. In high-risk patients, chemotherapy rates were similar in LN0 (90%65 and 89%69) and LN+ patients (90%). 66

Studies generally reported different outcomes [5-year DRFS (n = 2),22,65 DRFI (n = 2),66,69,75 IDFS (n = 2),22,73,74,77 BCSS (n = 3)66,67,69,70 and OS (n = 1)22], making comparisons across studies difficult. For outcomes including recurrence (DRFS, DRFI and IDFS), low-risk patients with recurrence scores of < 1865,66,69,75 had outcomes ranging from 95.9%65 (5-year DRFI) to 99.6%75 (5-year DRFI) in LN0 patients (n = 3) and 97%66 (5-year DRFI) in LN+ patients, whereas low-risk patients with recurrence scores of < 11 had outcomes of 94%22 (5-year IDFS) and 99.9%75 (5-year DRFI) in LN0 (n = 2)22,75 and 95%66 (5-year DRFI) in LN+ patients (n = 1). 66 Clinical advice to the EAG suggests that these levels of recurrence are acceptable in a low-risk population.

It was beyond the scope of the assessment to determine whether the newer cut-off points (of 11–25) should be used, or whether the original cut-off points of recurrence score 18–30 would be preferable. Data relating to this are summarised in the narrative synthesis (see Report Supplementary Material 4) and Table 8, and the general observation can be made that although use of lower cut-off points may result in better outcomes in the low-risk group (although data are mixed on this point), it would also result in fewer patients being classified as low risk.

Data relating to intermediate- and high-risk patients are included in the narrative synthesis (see Report Supplementary Material 4) and appear in Table 8. It was not possible to draw any conclusions regarding whether or not patients in intermediate- and high-risk categories had better outcomes as a result of using oncotype DX to guide treatment, as there were no comparator (no-oncotype DX) groups.

The data on micrometastases are difficult to interpret as there is no analysis that reports all nodal statuses in the same patient group (i.e. LN0, LNmicro, LN1–3). The analyses that have been done show that the trend for worse outcomes with increasing risk group holds true in patients with micrometastases. 66,67

The data relating to the performance of the test in patients of different races showed that BCSS survival differed similarly in accordance with risk categories in all race categories. 67

Conclusions

Without the highest level of evidence, it is not possible to conclude whether or not patient outcomes would be affected by use of the test in a clinical setting. In LN0 patients, use of the test in clinical practice appears to result in low rates of chemotherapy use in low-risk patients (1–12%), with acceptable outcomes (DRFS/DRFI/IDFS 96–99.6%). Rates of chemotherapy use increased with increasing risk category, and were generally higher in LN+ patients; only one study reported DRFS/DRFI/IDFS for LN+ patients, which was 97% (7% received chemotherapy). It was not possible to draw any conclusions regarding whether or not patients in intermediate- and high-risk categories had better outcomes as a result of using oncotype DX owing to the observational nature of the studies.

Study characteristics

Several publications describe validation of the prognostic value of MammaPrint. Many include overlapping cohorts of patients, sometimes pooled with other cohorts, sometimes focusing on patient subgroups (e.g. ER+ or LN0/LN+), sometimes updating the data with longer follow-up and reporting a range of different outcomes. Therefore, it should be noted that there is some overlap between patient cohorts within the references included here. Table 9 shows both the study reference(s) (column 1) and the cohort(s) (column 2) used for each analysis, and Report Supplementary Material 5 describes the cohorts in more detail.

There were nine main cohorts, which were small, retrospective analyses of consecutive patient series: four were from the Netherlands,79,80,84,87,89 two were multi-European,85,88 two were from the USA90,96 and one was from Japan95 (total n = 1805). They covered a mix of LN0 and LN+ patients, with variable proportions receiving endocrine therapy and chemotherapy. In most studies, around 70–80% were ER+, and HER2 status was not well reported. Four analyses pooling some of the above cohorts78,83,87,93 are also included due to their focus on specific subgroups. In addition, there was one reanalysis of a RCT [the Swedish Stockholm Tamoxifen-3 (STO-3) trial; n = 538], of which a subgroup had endocrine monotherapy. 78,91,92,94,97

Quality assessment

One study (van de Vijver et al. 79) included a small proportion of patients from the derivation cohort, and may therefore overestimate prognostic performance, although a ‘leave-one-out’ analysis was used to mitigate this problem to some extent. Most analyses excluded some patients from the original cohort, some because of insufficient tumour sample, which may introduce bias due to attrition of patients with smaller tumours. Blinding of test assessors to outcomes was reported in around half of the studies. Outcomes did not always match standardised definitions and it was not always possible to tell from the publication whether or not all deaths and breast cancer deaths were counted as events or were censored. Different levels of treatment use in the high- and low-risk groups may confound results by reducing separation between the groups in recurrence rates, and selection of patients in accordance with treatment may introduce spectrum bias because these patients may be systematically different from the whole population. Many studies included a proportion of patients who were not in the scope of this research.

Distribution of patients by risk group

The percentage of patients categorised as low risk ranged from 20% to 71% and the percentage of those categorised as high risk ranged from 29% to 80% across seven analyses of LN0 patients. 84,86,88,89,91,95,96 In two analyses of LN+ patients,85,89 the percentages categorised as low risk were 38% and 41%, whereas the percentages categorised as high risk were 59% and 62%.

Prognostic performance: unadjusted analyses

Prognostic data for MammaPrint is provided in Tables 10 and 11. Among LN0/LN+ studies, Mook et al. 87 pooled 964 patients from seven series79–81,84–86,88 (one-third had endocrine therapy and one-quarter had chemotherapy) and showed that MammaPrint was statistically significantly prognostic for 10-year DRFS (HR 2.70, 95% CI 1.88 to 3.88; p < 0.0001), with 10-year DRFS of 87% in low-risk patients. 87 Knauer et al. 83 pooled 541 patients (restricted to LN0–3 patients, 100% endocrine therapy, 42% chemotherapy) from six of these series, and also reported statistically significant results. In terms of longer follow-up of the original van de Vijver et al. 79 cohort (51% LN0, 37% had chemotherapy and 14% had endocrine therapy), MammaPrint was statistically significantly prognostic in an unadjusted analysis of DRFS over 0–25 years89 in a LN0/LN+ cohort;79 most of this difference was in the first 5 years (HR 9.6, 95% CI 4.2 to 22.1). A separate US series (Yao et al. ;90 72% LN0, 43% had chemotherapy and 87% had endocrine therapy) also showed statistically significant prognostic ability for DRFS at 10 years (HR 2.91, 95% CI 0.97 to 8.68; p = 0.045) (see Table 9) with DRFS rates in the low-risk group of 96% at 10 years; results were similar (low-risk 10-year DRFS 98%) in a subset with no chemotherapy.

Among LN0 patients, in the only reanalysis of a RCT the STO-3 trial (van ‘t Veer et al. 91) reported 10-year DRFS rates (93% in low-risk patients, 85% in high-risk patients) (see Table 9) but no statistically significant levels were reported. 91 Four out of five retrospective LN0 cohorts reported statistically significant prognostic performance of MammaPrint for DRFS/DRFI at varying time points, based on unadjusted HRs between risk groups. 84,86,88,89 The 10-year DRFS/DRFI rates in low-risk patients ranged from 80% to 90% across three analyses (with varying rates of endocrine therapy and chemotherapy use). 84,86,88

Three of the LN0 cohorts84,88,89 included comparisons with AOL and NPI, which appeared to have less prognostic value than MammaPrint, although no statistical comparisons were reported. There were no comparisons with other risk tools such as Predict or Modified Adjuvant! Online (mAOL). Among LN+ patients, two cohorts reported statistically significant prognostic performance of MammaPrint based on unadjusted HRs between risk groups, with 10-year DRFS rates in low-risk patients of 79% and 91% (with varying rates of endocrine therapy and chemotherapy use). 85,89

See Report Supplementary Material 5 for subgroup analyses in patients with low or high clinical risk and in patients with lobular breast cancer, and for OS and BCSS outcomes.

Additional prognostic value

Several studies reported adjusted analyses relating to the additional prognostic value of MammaPrint over existing clinicopathological risk scores and clinicopathological variables. A pooled analysis of LN0/LN+ patients from seven series79–81,84,85,87,88 showed that MammaPrint was statistically significantly prognostic for 10-year DRFS in a multivariable analysis adjusting for clinicopathological variables. However, in the US series (Yao et al. 90), MammaPrint was borderline non-statistically significant in a multivariable analysis (p = 0.08). Among LN0 patients, MammaPrint was statistically significantly prognostic for DRFI when adjusted for AOL or NPI in three cohorts reported across two publications. 84,88 C-indices from two79,84 of these cohorts showed higher values when MammaPrint was included alongside clinicopathological factors than for either alone, although differences were not statistically compared. 84 In one analysis of LN+ patients, MammaPrint was borderline statistically significant for 10-year DRFS and statistically significantly prognostic for 10-year BCSS,85 although in another analysis79 BCSS at 10 years was borderline non-statistically significant.

Conclusions: MammaPrint prognostic performance

The prognostic value of MammaPrint is based on nine retrospective analyses, four pooled analyses (including six of the nine retrospective series and one prospective series) and one reanalysis of a RCT. Studies were variable in terms of nodal status, ER status and receipt of endocrine therapy and chemotherapy. The percentage of LN0 patients categorised as low risk ranged from 20% to 71% and the percentage of those categorised as high risk ranged from 29% to 80%. In LN+ patients, the percentage categorised as low risk was 38% to 41% and the percentage categorised as high risk ranged from 59% to 62%. MammaPrint was statistically significantly prognostic for 10-year DRFS in almost all unadjusted analyses of LN0 and LN+ patients (as well as in pooled analyses). For LN0 patients, 10-year DRFS/DRFI rates for low-risk patients ranged from 80% to 90% (with varying rates of endocrine therapy and chemotherapy use), whereas the reanalysis of a RCT reported 10-year DRFS of 93% with endocrine monotherapy and 83% without endocrine therapy or chemotherapy. Interestingly, although on the whole MammaPrint low-risk 10-year DRFS rates are lower than those for the other in-scope tests, the 93% figure for patients having endocrine monotherapy is more in line with other tests and may better reflect the population used in studies of other tests (ER+, endocrine monotherapy). For LN+ patients, 10-year DRFS rates in low-risk patients ranged from 79% to 91% (with varying rates of endocrine therapy and chemotherapy use). In terms of additional prognostic value, MammaPrint was statistically significantly prognostic for 10-year DRFS/DRFI in multivariable analyses adjusted for clinicopathological risk tools (AOL and NPI) and various combinations of clinicopathological variables in LN0/LN+ and LN0 cohorts, whereas adjusted analyses in LN+ cohorts were statistically significant or borderline significant.

Study designs and patients

Two publications have reported the ability of MammaPrint to predict the benefit of chemotherapy. Knauer et al. 83 reported a pooled analysis of 541 patients (100% received endocrine therapy, 42% received chemotherapy) from six consecutive patient series (see Report Supplementary Material 5, Table 7). Overall, 90% were ER+ and 89% were HER2–, and half were LN0 and half had one to three positive nodes (LN1–3). In addition, the article by Mook et al. 85 reported a pooled analysis of two out of the six patient series from Knauer et al. 83 (see Report Supplementary Material 5, Table 7), with an extended follow-up (10 years), but restricted to LN1–3 patients (including micrometastases).

Quality assessment

Both studies were pooled retrospective cohorts in which patients were treated in accordance with usual practice [in addition, one of the six cohorts in Knauer et al. 83 was the prospective MicroarRAy-prognoSTics-in-breast-cancER (RASTER) study,81 in which patients were treated in accordance with usual practice plus MammaPrint]. For this reason, those who received chemotherapy are likely to be systematically different in terms of known (and potentially unknown) prognostic factors to those who did not, leading to a high risk of confounding. Both studies blinded the test assessors to clinical outcomes, and both used standard outcome definitions. Both studies included a proportion of patients outside the scope (ER– and/or HER2+). See Report Supplementary Material 5, Table 7.

Results

The pooled analysis of six consecutive series by Knauer et al. 83 reported that at 5 years, there was a statistically significant effect of chemotherapy in the MammaPrint high-risk group but no statistically significant effect in the low-risk group, although HRs favoured chemotherapy in both groups (Table 12). Unadjusted HRs for DRFS (for no chemotherapy vs. chemotherapy) were 0.26 (95% CI 0.03 to 2.02; p = 0.20) in the low-risk group and 0.35 (95% CI 0.17 to 0.71; p < 0.01) in the high-risk group, whereas unadjusted HRs for BCSS were 0.58 (95% CI 0.07 to 4.98; p = 0.62) in the low-risk group and 0.21 (95% CI 0.07 to 0.59; p < 0.01) in the high-risk group. Multivariable analyses of the effect of chemotherapy on 5-year BCSS were again statistically significant in the high-risk group (HR 0.21, 95% CI 0.06 to 0.80; p = 0.02) but not in the low-risk group (HR not estimable; p = 0.98) (see Table 12). However, the interaction test for chemotherapy treatment and risk group was not statistically significant (p = 0.45; the interaction test appears to relate to 5-year BCSS as opposed to DRFS but this is unclear in the publication83). This indicates that the effect of chemotherapy versus no chemotherapy on 5-year BCSS was not statistically significantly different between risk groups. It is unclear whether this interaction test relates to the adjusted or unadjusted analysis.

For the two pooled LNmicro–3 cohorts reported by Mook et al. 85 (these were subsets of two of the six cohorts pooled in Knauer et al. 83), the only evidence relating to prediction of chemotherapy benefit was a test of the interaction between chemotherapy treatment and risk group (within a multivariable analysis of 10-year BCSS), which was not statistically significant (p = 0.95) (see Table 12).

In the analysis of six series,83 it was unclear whether the interaction test was unadjusted or adjusted, and, if so, for which factors. In the analysis of LN1–3 patients from two series,85 the interaction test was conducted within a multivariable analysis adjusted for clinicopathological variables.

Conclusions: MammaPrint chemotherapy benefit

In a pooled analysis of 541 patients (half LN0, half LN1–3) in which patients were treated in accordance with usual practice, the effect of chemotherapy versus no chemotherapy on 5-year DRFS and BCSS was statistically significant in the MammaPrint high-risk group but not in the low-risk group in unadjusted analyses for 5-year DRFS and BCSS and in adjusted analyses for 5-year BCSS. However, the interaction test for chemotherapy treatment and risk group (for 5-year BCSS) was non-significant (p = 0.45). A further pooled analysis of two of the above series, restricted to LN1–3 patients, also reported a statistically non-significant interaction between chemotherapy treatment and risk group for 10-year BCSS (p = 0.95). The evidence for the ability of MammaPrint to predict chemotherapy benefit is extremely limited; although unadjusted analyses suggest a greater effect of chemotherapy in high-risk groups, adjusted analyses were only reported for one outcome, and the non-significant interaction tests suggest that there was no statistically significant difference in effect of chemotherapy between risk groups.

Overview

Two studies reported evidence relating to clinical utility of MammaPrint (the impact of prospective use of the test on clinical outcomes). Microarray In Node-negative Disease may Avoid ChemoTherapy (MINDACT) is a RCT of MammaPrint versus clinical practice. 98 RASTER81,99,100 is a prospective observational study in which patients were treated in accordance with usual practice plus MammaPrint. As these two studies are very different in design, they are reported separately in the following sections.

This section of the report summarises the main points; a full description can be found in Report Supplementary Material 5.

Clinical utility randomised controlled trial: MINDACT

Results

Conclusions: randomised controlled trial of clinical utility for MammaPrint (MINDACT)

MINDACT randomised patients with discordant MammaPrint and mAOL risks to chemotherapy or no chemotherapy. For patients who were high-clinical, low-MammaPrint risk, 5-year DMFS was 95.9% with chemotherapy and 94.4% without chemotherapy, an absolute difference of 1.5%. This raises the possibility of avoiding chemotherapy in these patients. In patients who were low-clinical, high-MammaPrint risk, 5-year DMFS was 95.8% with chemotherapy and 95.0% without chemotherapy, an absolute difference of 0.8%. This could be interpreted as showing that MammaPrint may not be useful in this group as it would increase chemotherapy rates without improving outcomes. However, the comparator was mAOL, and it is unclear whether or not the same would be true for other clinical risk measures.

Clinical utility observational study: RASTER

Results for lymph node-negative patients

Results for lymph node-positive patients

MammaPrint was retrospectively conducted in 164 LN+ patients (Table 16). Over 95% of patients received chemotherapy. MammaPrint categorised 48% of LN1–3 patients as low risk. The 5-year DRFI was 98.4% for low-risk and 86.9% for high-risk patients, whereas 10-year DRFI was 94.9% for low-risk and 80.7% for high-risk patients, showing a statistically significant difference between groups (HR 4.7, 95% CI 1.3 to 16.2). A comparison was made with the mAOL, although this analysis included 30 additional patients with LN > 3 who were automatically classed as high risk. The mAOL categorised only 14% of patients as being at low risk; 10-year DRFI was 94.4% for low-risk and 85.8% for high-risk patients, which was not statistically significantly different (HR 3.7, 95% CI 0.5 to 28.5). Within the mAOL high-risk group, 10-year DRFI was statistically significantly better in MammaPrint low-risk (95.2%) than high-risk (79.6%) patients (HR 4.8, 95% CI 1.1 to 21.4). 101

Conclusions: observational study of clinical utility for MammaPrint (RASTER)

RASTER is a prospective observational study in which patients were treated in accordance with MammaPrint plus usual clinical practice (LN0) or in accordance with usual clinical practice (LN+). The 5-year DRFI for LN0 patients was 97.0% for low-risk patients (15% had chemotherapy) and 91.7% for high-risk patients (81% received chemotherapy). The 10-year DRFI for LN0 patients was 93.7% for low-risk and 86.8% for high-risk patients. The DRFI rates in the MammaPrint low-risk group may be considered sufficiently low for these patients to avoid chemotherapy. MammaPrint provided additional prognostic information over AOL and NPI, but not over Predict Plus when considering C-indices. Estimates of prognostic performance between risk groups are likely to be affected by the differing rates of chemotherapy in each group and the fact that chemotherapy use was influenced by MammaPrint.

Study designs: Prosigna prognostic performance

Eight data sets were used to assess the prognostic performance of ROR-PT (see Report Supplementary Material 6, Table 1). These included six reanalyses of RCTs [TransATAC,38,46 Austrian Breast and Colorectal Cancer Study Group (ABCSG) 8,104,105 Cancer and Leukemia Group B (CALGB 9741),106 National Cancer Institute of Canada (NCIC) MA.21,107 Grupo Español de Investigación en Cáncer de Mama (GEICAM) 9906108,109 and NCIC MA.12110] and two retrospective analyses of prospective cohorts [the Danish Breast Cancer Cooperative Group (DBCG) cohort111–114 and two analyses of the British Columbia cohort115,116].

Patients: Prosigna prognostic performance

Two of the RCTs (TransATAC38,46 and ABCSG 8;104,105 total n = 2252) and the two retrospective analyses (total n = 3508) included patients who were all/mostly ER+ and HER2– and received endocrine monotherapy. The other four RCTs106–110 (total n = 3358) included higher-risk patients (not restricted to ER+ and HER2–, higher proportion LN+) and all received chemotherapy (see Report Supplementary Material 6, Table 1). Two studies recruited only LN+ patients (CALGB 9741106 and GEICAM 9906108,109), one recruited LN0 patients (DBCG111–114) and the remainder recruited both LN0 and LN+ patients.

Tests and comparators: Prosigna prognostic performance

Four analyses of RCTs38,46,104–107 and two analyses of prospective cohorts111–115 measured ROR-PT using the nCounter device, and two analyses of RCTs108–110 and one of a prospective cohort116 used RT-qPCR (see Report Supplementary Material 6, Table 1). The cut-off points used to define risk groups varied across studies, and some analyses assessed ROR-PT as a continuous score (see Report Supplementary Material 6, Table 1).

Quality assessment: Prosigna prognostic performance

See Report Supplementary Material 6, Table 1. All analyses excluded some patients recruited to the original trial or cohort. Blinding of test assessors to outcomes was reported in five analyses. All used standardised outcomes.

Results: Prosigna prognostic performance

Conclusions: Prosigna prognostic performance

Based on six reanalyses of RCTs and two retrospective analyses of prospective cohorts, Prosigna/ROR-PT was statistically significantly prognostic for unadjusted analyses of 10-year DRFS/DRFI in LN0 and LN+ patients. Among LN0 patients, approximately 50% were categorised as low risk, 30% as intermediate risk and 15% to 20% as high risk. Among LN+ patients, 4% to 25% were at low risk, 27% to 56% were at intermediate risk and 26% to 62% were at high risk. The 10-year DRFS/DRFI rates for low-risk patients were 95% to 97% in three studies of LN0 patients (all endocrine therapy only), and in LN+ patients these were 100% in two studies (endocrine therapy only) and 92% in one study (all endocrine therapy and chemotherapy). ROR-PT added prognostic information over clinicopathological variables or CTS/CLP/NPI in three studies; this was statistically significant in LN0 patients and either significant or borderline significant in LN+ patients.

Study characteristics

The prognostic value of EPClin was assessed in three reanalyses of RCTs,36,46,108,109,118–120 which included four RCTs (total n = 3135): UK patients from TransATAC (n = 878), the ABCSG 6 and ABCSG 8 (total n = 1702) and the Spanish GEICAM 9906 trial (n = 555). All recruited only, or reported a subgroup of, patients who were ER+ and HER2–. One reported on LN0 patients (total n = 680)36,39,46 and two36,46,108,109 on LN+ patients (total n = 753; one108,109 included 36% patients with at least three positive nodes). One reported on patients unselected by lymph node status,120 and one on both LN0 and LN1–3 together;46 additional analyses118 were provided to the EAG as commercial-in-confidence data and cannot be reported here. Patients received endocrine monotherapy in two trials36,46,120 and all patients received endocrine therapy and chemotherapy in the GEICAM trial. 108,109 All excluded some original trial participants (or this was unclear). A detailed narrative synthesis of study characteristics is provided in Report Supplementary Material 7. This supplement also contains relevant data in tables.

Quality assessment

The EAG’s assessment of study quality is provided in Report Supplementary Material 7, Table 2. All analyses excluded some original trial patients (or this was unclear), sometimes owing to insufficient tumour samples, which may introduce bias owing to attrition of patients with smaller tumours. Blinding of test assessors to outcomes was reported in two analyses. 36,108,109 All used standardised outcomes. A discussion of how these factors might affect results is given in Results: oncotype DX.

Results

A summary of key results from the review is provided in this section. A detailed narrative synthesis of all study results is provided in Report Supplementary Material 7.

Conclusions: EndoPredict and EndoPredict Clinical prognostic performance

Based on three reanalyses of RCTs (total n = 3135) in ER+, HER2–, endocrine-treated patients, EPClin was statistically significantly prognostic for unadjusted analyses of 10-year DRFS/DRFI in LN0 and LN+ patients. The percentage of patients categorised as EPClin low risk was 73% for LN0 patients and was 13% and 24% for LN+ patients. The 10-year DRFS/DRFI rates for low-risk patients were approximately 95% in LN0 and LN+ patients receiving endocrine therapy only. EPClin added statistically significant information over CTS/NPI in LN0 and LN+ patients in TransATAC, and in two further studies the EndoPredict score added statistically significant information over clinicopathological variables in mixed LN0/LN+ and LN+ patients (no data for EPClin). There was no evidence relating to chemotherapy benefit or clinical utility for EndoPredict or EPClin.

Analytic validity

A rapid review of the analytic validity of IHC4 can be found in Report Supplementary Material 8.

Study designs, patients and treatments: IHC4 prognostic performance

Five cohorts [Tamoxifen vs Exemestane Adjuvant Multinational (TEAM), the WSG PlanB, Intergroup Exemestane Study (IES), GEICAM 9906 and West German Study Group epirubicine and cyclophosphamide-Doc (WSG-AGO-Doc)]24,73,74,77,122,123,126,127,129 were reanalyses of RCT data (three LN+/LN0 studies, n = 8496; no LN0 studies; two LN+ studies, n = 1705) and six25,58,124–126,128 were reanalyses of routinely collected data when patients were treated in accordance with usual practice without the use of IHC4 or IHC4+C (five LN+/LN0 studies, n = 3128; one LN0 study, n = 105; no LN+ studies). All studies recruited HR+ or ER+ patients except the IES RCT127 and the study from Taiwan. 124 TransATAC,46 WSG PlanB,73,74,77 GEICAM 9906,129 WSG-AGO-Doc,123 the Kaiser Permanente cohort,125 the Institut Curie128 cohort and the Chinese58 cohort all recruited or reported a subgroup of HER2– patients. Only one validation cohort (Stephen et al. 126) treated 100% of patients with endocrine monotherapy, and the remainder treated varying proportions of patients with endocrine therapy and chemotherapy. Two observational studies were from East Asia;58,124 clinical advice received by the EAG suggests that these two East Asian studies may be less generalisable to the English context because (1) patients were treated in accordance with usual clinical practice and this may differ between these countries and England enough to affect prognostic outcomes and (2) it is possible that people of different ethnicities have different underlying risk profiles and disease natural histories.

A detailed narrative synthesis of patient characteristics and treatments is provided in Report Supplementary Material 8. Overall, only the derivation cohort (TransATAC)46 reported an analysis of 100% ER+, HER2–, LN0–3 patients who had not undergone chemotherapy but had received 5 years of endocrine therapy. For this reason, most of the evidence base has low generalisability to the decision problem.

IHC4 methodology and cut-off points: IHC4 and IHC4+C prognostic performance

The methodology for conducting IHC4 is not standardised outside the Royal Marsden Hospital. Report Supplementary Material 8 details the methods reported in the included studies, and a judgement provided by personal communication with the IHC4 methodologists [Andrew Dodson, National External Quality Assessment Service (UK), personal communication, September 2017] regarding whether or not the methods used are compatible with their own method is provided in Report Supplementary Material 8 and Table 21. Seven data sets were analysed using IHC4 methodologies that were the same as or very similar to the IHC4 team’s own methodology (referred to from here on in as the ‘standard IHC4 methodology’) [TransATAC,46 TEAM,24,122,126 the Nottingham cohort,25 the Edinburgh Breast Conservation Series (BCS) cohort,126 the Institut Curie128 cohort, GEICAM 9906129 and WSG-AGO-Doc123], whereas the remaining five data sets were analysed using methodologies that were unclear or dissimilar to the IHC4 team’s methods (WSG PlanB,73,74,77 the Kaiser Permanente cohort,125 IES,127 the Chinese cohort58 and the Taiwanese cohort124). Results have not been excluded by IHC4 methodology, as methodologies are not currently standardised and, for this reason, all data are of some relevance.