Survival estimates stratified by the Nottingham Prognostic Index for early breast cancer: a systematic review and meta-analysis of observational studies

The Nottingham Prognostic Index (NPI) is a commonly used, clinically relevant and internationally validated [1, 2] system for classifying early and locally advanced breast cancer cases (TNM stages I, II, and III [3]) into three or more prognostic groups [4, 5]. Although more advanced models [6, 7] have superseded NPI in some applications, it continues to have an active role in clinical practice and research [8]. In clinical practice, NPI and related prognostic models have an important application for patients and clinicians to inform the decision of whether or not to undergo adjuvant chemotherapy following surgery [9], an early example of what is commonly called precision medicine [10].

The NPI can be used to provide prognostic information by assigning individuals into prognostic categories and then applying the survival estimates from a previous cohort study [11]. As well as providing prognostic estimates for individual patients, the NPI is useful in the context of economic evaluations. In economic evaluation of new interventions, overall survival is a key input for decision-analytic models designed to quantify the incremental costs and benefits and inform if, and how, to allocate finite healthcare budgets towards new screening and management options [12]. Decision-analytic models have a key role in the evaluation of new technologies when clinical trials are not likely to be feasible or timely in terms of producing robust evidence of the impact of introducing new programmes or changing existing programmes [13].

The NPI categories are a linear combination of three prognostic factors: tumour size (maximum diameter in millimetres), histological grade (1- to 3-point scale) and lymph node staging (1- to 3-point scale). Standardized published cut-offs (see Additional file 1: Appendix 1) are used to form prognostic categories within a population (e.g. good, moderate and poor prognostic groups).

This study aimed to identify and describe all published observational studies reporting NPI-category-specific overall survival following a diagnosis of early and/or locally advanced breast cancer. A meta-analysis was used to explore the between-study heterogeneity in survival estimates.

A systematic review and meta-analysis, following published recommendations [14, 15], were used to identify all published cohort studies investigating the survival of early and locally advanced breast cancer patients stratified into prognostic groups using the NPI and to synthesize the results. This systematic review was not registered with PROSPERO.

A total of 28 studies were suitable for inclusion in the review (see Fig. 1) and 19 of these studies were included in the meta-analysis (see Table 1). Of the 28 studies, two studies [2, 26] provided the results of survival analysis for two data series; therefore, a total of 30 datasets were included in the review and 20 in the meta-analysis. Three studies [27‐29] provided only data up to 5-year survival and were excluded from the meta-analysis of 10-year survival. Nine studies (containing ten datasets) were excluded from any kind of meta-analysis due to the following reasons. The quality of the graphical presentation in two studies [30, 31] did not allow for extraction of the survival data. Three further studies [2, 32, 33] did not report numbers of patients in each NPI category. One [34] used different cut-off points for NPI that meant the categories were not comparable with the majority of the studies. Two studies [35, 36] reported insufficient data. One study [37] excluded patients who died in the first months after the surgery, and therefore, the provided survival estimates were incomplete. Eight data series from seven studies [1, 5, 26, 38‐41] reported results from the original Nottingham case series and/or West Midlands Cancer registry (to which Nottingham contributes). These studies report partially overlapping cohorts of patients (see Fig. 2). To minimize ‘double-counting’, four studies were excluded from the meta-analysis which meant that the remaining included studies [1, 5, 41] contained minimal overlap in the reported cohorts. Among the remaining studies, Blamey et al. [1] does overlap with Allgood et al. [41]; however, the proportion of data that was shared was a relatively small proportion of the data because the first study reports a pooled European cohort of which approximately 15% of the data is from Nottingham, while the second study reports data from the West Midlands Cancer Registry to which Nottingham contributes only a minority of cases.

This review identified a substantial number of observational studies that aimed to generate estimates of the overall survival of women diagnosed with early and/or locally advanced breast cancer stratified by prognostic category using NPI.

The strengths of this study include using a comprehensive, systematic review of the literature underpinned by robust review methods (see Additional file 2: PRISMA checklist). A broad electronic search strategy, supplemented with hand searching of reference lists of identified studies, to collate all the relevant studies and datasets. The search strategy and inclusion criteria achieved good sensitivity in identifying relevant studies. The identified studies included patient cohorts from a broad range of time periods for data collection, countries and age groups. Critical appraisal following a standard protocol found study reporting quality to be generally satisfactory. A number of relevant identified studies had to be excluded from the meta-analysis of overall survival as the relevant data could not be extracted from the published manuscript and authors were not able to provide the data required. The studies excluded at this stage did not appear to suggest different outcomes for survival compared to the included studies based on the more limited information available. We focussed on cohort studies and did not include randomized controlled trials (RCT). RCT studies were considered unlikely to report NPI-stratified survival estimates and are likely to have applied more extensive exclusion criteria than cohort studies. No formal quantitative assessment of publication bias was made. It is not clear in this context the mechanism or result of publication bias that should be expected because each study was testing different hypothesis/hypotheses and in which the data extracted for this study were presented descriptively.

A particular concern in evidence synthesis is whether it was valid to pool data across multiple studies [20].This exploratory meta-analysis investigated study heterogeneity and demonstrated that this was too severe for simple pooling of survival estimates to be appropriate for the purposes of estimating survival in the current population. It is of interest to understand the factors driving the heterogeneity in survival estimates; however, these data are insufficient to allow a robust meta-regression or multilevel analysis to test hypotheses about such determinants. There are only a small number of studies and many possible explanatory variables, including some for which only weak surrogate markers would be available from study level data. Investigation of the sources of between-study heterogeneity could be best achieved using an individual patient data meta-analysis to explore factors such as the year of diagnosis of patients, country, screen detection rate and age.

This study had a number of limitations. To provide quantitative estimates of overall survival by NPI category, given the relative age of some of the identified studies, it was necessary to collate some data by visual inspection of published survival curves. This process will have introduced some measurement error. Due to lack of detailed reporting of losses to follow-up it was also necessary to assume that all studies had no loss to follow-up, which could result in different degrees of selection bias in the results if there were differential losses to follow-up across studies. Based on the published study protocols and the routine and often statutory collection of mortality data, we believe that losses to follow-up are likely to be minimal. Due to limitations of available resources, non-English language studies and those with insufficient reporting of data were excluded. Furthermore, it was considered infeasible based on available resources and likely data sharing restrictions, to attempt to collate individual patient data to address some of these limitations. The resulting exclusions may introduce some selection bias, but it is not possible to know the impact of such a bias.

There were some observed differences in included studies in the methods used to assess lymph node status which is a key input to define NPI category. Originally, node involvement for selected NPI category was assessed using biopsy results of a lower axillary node, an apical axillary node and a node from the internal mammary chain. Node status was specified as one if the tumour was absent from all three nodes, two if tumour cells were found in lower axillary node only, and three if tumour cells were in the apical and/or internal mammary node [4]. To increase usability of the NPI, Galea et al. [5] suggested staging based on the number of involved nodes from the single location as a viable alternative, which was subsequently used in later studies. However, we believe this change in assessment of node status was sufficiently similar to not introduce bias in the analysis.

Precision medicine promises to deliver improved patient outcomes through better understanding of patient and disease heterogeneity, and ultimately better targeted screening and therapeutic strategies. Moving the concept of precision medicine into clinical practice requires mechanisms to identify and stratify eligible patient populations. Prognostic models are an important component in such a mechanism because they allow an individualized quantitative estimate of potential treatment benefit.

This systematic review and meta-analysis of NPI stratified estimates of survival revealed a key challenge in the design and evaluation of precision medicine interventions; that substantial differences may exist between studies estimating the relationship between a marker or score and the outcome of interest. Heterogeneity of study estimates should be investigated carefully for all such interventions. This finding suggests a potential additional burden on shared decision-making between clinicians and patients. Achieving a better informed decision is only possible if all of the available evidence is correctly synthesized and clearly presented.

The results of this systematic review and meta-analysis have implications for patients, clinicians, decision analysts and policy analysts requiring a prediction of long-term overall survival, in women diagnosed with early and/or locally advanced breast cancer. The observed heterogeneity in overall survival estimates indicates that it is important to make use of survival data from an appropriate setting to provide reliable prognostic information for use in clinical practice, and also in decision-analytic model-based economic evaluations of new treatments, screening programmes or prevention strategies to inform health care resource allocation decisions.