In oncology, prognostic markers (also called prognostic factors) are clinical measures used to help elicit an individual patient's risk of a future outcome, such as recurrence of disease after primary treatment. They play a key role in clinical practice, distinguishing patients into different risk groups and thus informing treatment strategies and aiding patient counselling. They can also be used to define strata in clinical trials to ensure comparability of treatment groups. Markers can be simple measures, such as the stage of disease or tumour size, but are often more complex, such as abnormal levels of proteins or genetic mutations. For example, within paediatric oncology, amplification of the MYCN proto-oncogene is a known indicator of poor outcome in neuroblastoma patients (Riley et al, 2004a
). In adult oncology, it is well known that the number of positive lymph nodes has a strong influence on the prognosis of recurrence-free survival time in patients with primary breast cancer (Galea et al, 1992
). For (nearly) all diseases, some important prognostic markers have been well established for a long time, but many more are investigated each year. Few, however, find a role in clinical practice.
Not all potential markers turn out to have prognostic value. Much research effort goes into conducting studies that evaluate the extent to which certain markers help clinical prognosis. Indeed, hundreds of prognostic marker studies are published in cancer journals each year. Unfortunately, though, there is large concern about the quality of such studies and it is clear that large progress is needed to produce clinically relevant results (Hinestrosa et al, 2007
). Over more than 10 years, an increasing body of evidence has signalled that prognostic marker studies are often badly designed (Simon and Altman, 1994
; Altman and Lyman, 1998
), inappropriately analysed (Holländer and Sauerbrei, 2006
), poorly reported (Riley et al, 2003
; Kyzas et al, 2005
), and subject to numerous biases, such as selective reporting (Kyzas et al, 2005
) and ‘optimal' choice of cutpoints (Altman et al, 1994
). Bad quality of design, analysis, and reporting of individual studies result in confusion regarding the prognostic value of a new marker (Sauerbrei, 2005
). Furthermore, evidence-based marker results are a rarity, with systematic reviews and meta-analyses of multiple studies serving only to highlight the serious deficiencies within primary prognostic research (Altman, 2001
). For example, two systematic reviews, one in neuroblastoma (Riley et al, 2004a
) and another in non-small-cell-lung cancer (Brundage et al, 2002
), found that over 100 different prognostic markers have been investigated in each field, yet the median number of articles per marker was only 1. Clinicians are thus faced with a confusing yet ever-growing body of literature, where new markers are regularly investigated but rarely in relation to existing markers. The evidence-base is further blurred by large heterogeneity across studies in the patient population, the method of measuring each marker, the outcomes reported, and the statistical analyses used, among numerous other factors (Riley et al, 2003
), all of which limit a coherent meta-analysis.
In this paper, we examine the key methodological issues of the prognostic marker research process, encouraging a move towards higher standards and clinically useful results. Our aim is to discuss how high-quality evidence can evolve over time from initial exploratory studies, to large protocol-driven primary studies, to meta-analysis and even beyond, to large prospectively planned pooled analyses, and to the initiation of tissue banks. The work extends earlier articles aimed at a more statistical audience (Altman et al, 2006
; Holländer and Sauerbrei, 2006
; Sauerbrei et al, 2006
; Riley et al, 2006a
), and focuses on prognostic, rather than predictive, markers. In cancer, a prognostic marker is one that predicts a patient's clinical course, whereas a predictive marker is one associated with differential responses to a specific treatment (although this terminological distinction is perhaps not widespread across medicine). Predictive marker studies are relatively uncommon, and best embedded within randomised controlled trials. Sargent et al (2005)
discuss how to conduct a clinical trial that assesses the utility of a predictive marker.
Our paper focuses on studies (or meta-analyses) that seek to evaluate a small number of pre-specified markers. In recent years, clinicians have become increasingly enthusiastic about the possibility of improving prognosis and prediction by applying the new technologies of measuring many thousands of genes. However, these technologies create a very large number of potential predictors and require statistical methods to extract a relatively small number of relevant items. We also do not consider the study of prognostic markers from high-dimensional data, such as microarray studies (Tinker et al, 2006
; Dupuy and Simon, 2007
The outline of the paper is as follows. In the section ‘Primary studies of progostic markers', we consider primary studies of prognostic markers and consider their design, analysis, and reporting. In the section ‘Systematic reviews and meta-analysis', we discuss systematic reviews of such studies and explain why meta-analyses that use individual patient data (IPD) are important. In the section ‘Towards multi-centre and collaborative research', we consider why prospectively planned pooled analyses and the initiation of tumour banks may best facilitate evidence-based prognostic marker results. Finally, in the Discussion section, we summarise the key messages for future research.