PSA

The most well-known prognostic marker that has been used to assess prognosis (as well as detection of early disease) is PSA. PSA is a 30- to 33-kDa protease belonging to the kallikrein family, which is made up of 15 serine proteases encoded by a cluster of genes on chromosome 19q3.18 The earliest reported investigations of tissue-specific antigens in the human prostate were conducted by Ablin and colleagues in 1970.19 Further investigations resulted in the discovery of prostatic antigens in seminal plasma.20,21 Sensabaugh and Crim22 went on to characterise and isolate PSA from human seminal plasma during investigations into potential markers to aid detection of rape crimes. Wang and colleagues23 purified and isolated an antigen from prostate tissue that was considered to be prostate specific in nature. A large number of men are being diagnosed with early-stage prostate cancer as a result of the increasing use of PSA testing.24

Stage

In the TNM system, the extent of primary tumour (T category), regional lymph node involvement (N category) and distant metastasis (M category) are determined. The TNM system for classifying the anatomic extent of disease in cancer has been in existence for more than 50 years.25 Over time the TNM classification has evolved to accommodate new knowledge from the growth in medical research to improve its prognostic ability and keep pace with the demands of clinical practice.26 The TNM system was last updated in 2002.27 The latest version of the TNM staging system is used to stage prostate cancer (Table 2).28 Two main changes have been made to the new TNM classification system compared with the older versions: (1) subdivision of T2 disease into three clinical substages and (2) the recommendation that the Gleason scoring system is used for grading.

The clinical stage is based on information obtained before surgery to remove the tumour. The pathological stage provides additional information from the examination of the tumour microscopically. Pathological staging provides a more direct examination of the tumour and its spread, whereas clinical staging can be limited as the information is obtained by making an indirect assessment of the tumour whilst it is still in the patient. In Europe the TNM staging system is most commonly used. In stage T1 the tumour is located within the prostate gland only and is too small to be felt on DRE. In stage T2 the tumour is still located only within the prostate but it can be felt on DRE. In stage T3 the tumour has spread from the prostate into the immediate surrounding tissue. The seminal vesicles may be included. In stage T4 the tumour is still within the pelvic region but may have spread to other areas, i.e. metastatic disease may be present. Both T3 and T4 are often referred to as locally advanced disease. However, it should be noted that, for the purposes of this review, despite being interested only in early localised prostate cancer, we shall still evaluate stages T1, T2 and T3 with no lymph node involvement or metastases.

Although the TNM system stages are universally used, a similar system called the Jewett–Whitmore system is sometimes used in the US (Table 3). This has more specific alphanumeric subcategories. The Jewett–Whitmore system classifies prostate cancer first into stages A, B, C or D. Stages A and B are considered curable, whereas stages C and D are treatable. A number is given to describe a condition within each stage.

It is important to recognise that patients may move stages over the course of disease progression. Upstaging or downstaging has been found following treatment and also stage classification can depend on the imaging procedure used.30

Gleason

The most commonly used scheme for reporting histological grade is the Gleason score. Within this scheme there are five possible tissue patterns with 1 being well differentiated (good prognosis) and 5 being poorly differentiated (poor prognosis). The two most frequent patterns are added together to give a score. Albertsen31 reported that over the last 20 years there has been a significant shift in the use of the Gleason scoring system: tumours scored as Gleason 2–5 a decade ago are more likely to be scored as Gleason 6 tumours today. Men with high-grade prostate cancers (Gleason scores 7–10) appear to be at greater risk of disease progression and death if managed expectantly, whereas for men with low-grade prostate cancers (Gleason scores 6 or less) the outcome is unclear.

Surgical margins

A positive margin of resection means that the tumour extends to the inked surface of the prostate specimen removed by the surgeon.32 Although this definition is useful it presents difficulties in terms of its practical application as the prostate is surrounded by many structures that limit its the radical removal. There appear to be two main causes of positive margins: (1) non-iatrogenic and (2) transection of intraprostatic tumour (capsular incision).32 The incidence of positive margins following radical prostatectomy (RP) has significantly decreased over the last decade.33–35 Although this may be partly the result of improvements in surgical techniques, it is likely that the majority of the decrease is due to stage migration and careful patient selection.32 It has been reported that patients with positive margins have an increased risk of progression compared with patients with negative margins.33,36 These studies by Epstein and colleagues found that the probability of being progression free at 5 years following RP ranged from approximately 81% to 83% for margin-negative disease and from 58% to 64% for margin-positive disease.

Survival

Few studies report survival outcomes, mainly because patients diagnosed with low-stage localised prostate cancer typically survive for several years and in fact many will die of other causes. This demonstrates the importance of an adequate length of follow-up, although even then the number of events may be small. Those studies that do report survival outcomes vary in their definitions of survival.

The most reliable outcome in prostate cancer is all-cause mortality, but as most patients with prostate cancer do not die of the disease it is not a sensitive measure and is also highly dependent on the age distribution of the study population.

Prostate cancer survival is a more sensitive measure of prostate cancer outcome than all-cause mortality; however, a potential problem with prostate cancer survival as an outcome is ensuring that cause of death has been accurately determined.72,73

Clinical failure

Clinical failure may refer to local disease recurrence, the development of metastatic disease, or both. For patients who do not have radical treatment for prostate cancer there is no definition of biochemical failure, and disease progression is usually measured in terms of those developing symptomatic or metastatic disease. There are variations between studies in the frequency of follow-up and methods for identifying and confirming disease recurrence that may affect this outcome measure. Clinical failure may be biased if prognostic factors influence the frequency of follow-up.

Biochemical failure

As prostate cancer is a slowly progressive disease and has many competing causes of death, the development of biochemical failure may not necessarily be associated with prostate cancer mortality or clinical failure. There has been a surge of interest in attempting to identify a definition of biochemical failure after RP or radiation therapy that is both sensitive and specific in predicting subsequent clinically significant failure. Although the principle of using biochemical failure is a useful one, in practice it has proved difficult to determine an appropriate definition of what constitutes failure. For example, there is a difference in PSA behaviour following different treatment modalities. In principle, PSA levels fall to zero after a few weeks’ washout period following prostatectomy. Subsequent re-emergence of detectable PSA is interpreted as disease recurrence. However, radiotherapy does not necessarily destroy the entire prostate and it may take several months for PSA levels to reach the lowest point or ‘nadir’. Other treatments such as brachytherapy are also now available and each has a differing effect on subsequent PSA behaviour.

Following a consensus conference in 1996 the American Society for Therapeutic Radiology and Oncology (ASTRO) established a definition of biochemical failure following radiotherapy.74 The definition was three consecutive rises in PSA after a nadir, with the date of failure defined as a point half-way between the nadir date and the first rise, or any rise great enough to provoke initiation of salvage therapy. It was also recommended that a minimum period of follow-up of 2 years after therapy was required. Problems subsequently emerged with this definition, including the non-comparability of survival estimates based on different follow-up periods, as the backdating in the definition biases the survival estimates, the bias being worse the shorter the follow-up: results change dramatically if follow-up is only 3 years compared with 6 years. Another criticism of the 1996 definition of biochemical failure was that there had been no attempt to link it to clinical outcomes. To resolve these issues a second ASTRO consensus conference was held in 2005. A new definition of biochemical failure following radiotherapy, to be known as the ‘Phoenix definition’, was agreed: an increase of 2 ng/ml or more above the nadir PSA (lowest PSA attained following treatment). Data presented at the conference suggest that this definition yields a sensitivity and specificity of 66% and 77% for predicting clinical failure at 10 years. Patients who undergo salvage therapies without meeting the PSA failure definition should also be counted as failures at the time of positive biopsy or salvage treatment, whichever is first. A further recommendation of the conference was that control rates should be quoted at a time 2 years before the median follow-up to avoid the artefacts that may result from a short follow-up, including the backdating issue of the first ASTRO definition and the more favourable short-term outcomes that result from using the new Phoenix definition of PSA failure compared with the original ASTRO definition. However, it was emphasised that these definitions of PSA failure do not address the issue of cure rates, for which more data and longer follow-up are needed. As the new Phoenix definition was only published in 2006 it is unlikely that it will be used in many of the studies included in this review.

Cooksonet al.75 recently reviewed the variability in published definitions of biochemical recurrence and provided recommendations for a standard definition in patients treated with RP. Their review followed the American Urological Association (AUA) Prostate Guideline Update Panel being given the task of updating the guidelines for clinically localised prostate cancer. It became clear to the AUA that there were a substantial number of definitions being used to describe biochemical recurrence. Cookson and colleagues found 13,800 citations between 1991 and 2004 that included the terms prostate cancer and prostatic neoplasm, with 436 articles dealing with the clinical T1–T2N0M0 prostate definition of biochemical recurrence. Of these, 145 articles contained 53 different definitions of biochemical recurrence for those treated with RP. The most common definition after RP was a PSA of > 0.2 ng/ml or a slight variation of this. For radiation therapy, 208 articles were found reporting 99 varying definitions of biochemical failure. The most common definition for radiation failure was the ASTRO definition, three consecutive rises in PSA after a nadir. Overall, 166 different definitions of biochemical failure were found. The review shows the high degree of variability that is being used in the definition of biochemical recurrence following treatment for localised prostate cancer. These differences in definition can have a considerable effect on failure rates, as illustrated in a study by Amlinget al.76 For thresholds of 0.2 ng/ml and 0.5 ng/ml, biochemical survival was 62% and 78%, respectively, at 5 years. The authors concluded that strict definitions for biochemical recurrence are necessary to identify men at risk for disease progression and to allow reliable comparisons among patients treated similarly.

Following RP, the AUA recommends defining biochemical recurrence as an initial serum PSA of ≥ 0.2 ng/ml or more, with a second confirmatory PSA level of > 0.2 ng/ml. The panel recommended the use of the ASTRO criteria for patients treated with radiation therapy but recognised that these criteria will soon be updated.75

Review of novel markers

To be included the article had to report a primary prognostic study of (a) novel marker(s). Novel markers were defined as all disease-specific markers other than those previously defined as classical markers (clinical or pathological stage, total Gleason score, single pretreatment PSA measurement, surgical margins) but excluding epidemiological markers or measures of co-morbidity.

Review of prognostic models

To be included the article had to report a primary study or validation of a prognostic model. A prognostic model is defined as a model developed using statistical methodology to combine two or more factors to predict a relevant prostate cancer outcome. It should be noted that, although the statistical methods used to test the novel prognostic markers and to develop prognostic models are the same, to be classified as a review of a model the study needed to present predicted outcomes for different prognostic groups based on a multivariate analysis. Model articles that included novel markers were also included in the novel marker review.

Population

Males with a diagnosis of early localised prostate cancer (i.e. clinical or pathological stage TI/T2/T3N0M0 or Jewett–Whitmore system stages A, B, C) before treatment (radical or not) or at the time of radical treatment (prognostic markers taken before or at treatment). Studies were included if at least 80% of the study sample were in the target patient group.

Study end points

All reported measures of the prognostic value of individual or combinations of markers that predict the following outcomes:

• overall survival

• disease-specific survival

• disease-free survival

• biochemical (PSA) recurrence

• biochemical (PSA) freedom from recurrence

• clinical recurrence.

Exclusion of studies with fewer than 200 patients in the target group (T1/T2/T3N0M0)

Given the large volume of literature that the scoping literature searches indicated would be identified, we needed a simple method that would enable us to quickly identify the higher quality studies. Studies with a low number of outcome events (death or clinical/biochemical recurrence) tend to yield statistically weak analyses. It is recommended that analyses should have at least ten events per variable (EPV), if not 20,92 and so, with at least three (or four if pathological variables are included) classical variables that should be included in any multivariate analysis, as well as any novel markers, the very minimum number of events is 40–50. However, the number of events is often not reported and the reporting of the number of EPV is even more rare. The EPV can sometimes be estimated if sufficient information is presented, but this is often difficult to locate in an article. It was therefore decided that it was not practical to use number of events or EPV as a study inclusion criterion. Instead, a minimum number of patients used in the analysis was specified as an inclusion criterion for the review. This allowed small studies to be sifted out relatively quickly. The minimum was set at 200 based on an approximate calculation of the number of outcome events expected with a median follow-up of 5 years. This was carried out as follows. The outcome with the highest event rate is biochemical recurrence. Approximately 30% of patients suffer biochemical recurrence at 5 years following radiotherapy, with a similar proportion following surgery, dependent on the definition of biochemical recurrence.76,93 Approximately 10% of treated patients with localised prostate cancer will die within 5 years86 and we allowed a further 10% loss to follow-up. Thus, after 5 years in a cohort of 100 patients, 24 events {30 × [1.0–(0.1 + 0.1)]} might be expected. As a minimum of 40–50 events are required, a cohort of 200 was specified as an inclusion criterion. Note that other prostate cancer outcomes have much lower event rates and therefore need much larger cohorts to achieve 40–50 events. For the outcomes of local progression and prostate cancer death with cumulative incidence rates of 8.1% and 2.3% respectively,86 similar calculations to that shown above suggest that cohort sizes of at least 600 and 2000 respectively are required to obtain the same number of events.

Length of follow-up

Patients diagnosed with localised prostate cancer usually live for several years with their disease and are more likely to die of other causes. For those who have radical treatment, approximately 8.1% and 19.2% will have experienced local recurrence at 5 and 10 years respectively. Prostate cancer mortality at the same time intervals is 2.3% and 9.6% respectively.86 Clearly, studies with a follow-up of only a few months will identify only a small proportion of those who will eventually experience disease recurrence and almost none of those who will die of prostate cancer. In a study of radiotherapy94 24% of recurrences were recorded after 5 years of follow-up (median 6 years’ follow-up, maximum 11). This study quotes results from a study of prostatectomy95 showing that the proportion is similar following this mode of treatment: 27% of all recurrences occurred after 5 years in a series with a median follow-up of 8.8 years. They argued in favour of a follow-up period of at least 5 years following radiation therapy. In an editorial comment concerning a review of prognostic models used in prostate cancer61 it was noted that PSA recurrence in the reviewed nomograms was reported at between 2 and 6 years, ‘which is too short to be definitive’.

Another issue in determining the length of follow-up that is adequate for prognostic studies is the phenomenon of PSA ‘bounce’, which may occur following radiotherapy treatment. This is a temporary rise in PSA level, which with a short follow-up period may appear to be a failure. The American Society of Clinical Oncology (ASCO) recommends a minimum follow-up period of 2 years following radiotherapy.74

On the basis of the above discussion one might argue that the prognostic studies should have a follow-up of several years. However, there must be a balance between a sufficiently long follow-up, so that a significant proportion of those destined to suffer disease progression have done so, and the relevance of studies conducted several years previously when screening, diagnosis and treatments will have been different.

Scanning the literature indicated that using a minimum follow-up period as an inclusion criterion for the review would not be useful, as most studies do not report this statistic. Those that do report a measure of the follow-up period usually give a mean or median. Similarly, relying on the timing of the reported outcome (e.g. 5-year progression-free survival) was also unsatisfactory for two reasons. First, not all studies report the outcome in this way and, second, for those that do, it was clear that in some studies median follow-up represented only a fraction of the time to the reported outcome, suggesting a low level of events at this time and therefore potentially unreliable results.

It was decided pragmatically to apply a mean or median follow-up of 5 years as an inclusion criterion. Clearly the two measures are not the same as the distribution of follow-up time is often skewed, but as many studies report only one measure this was a practical method of eliminating studies with the shortest follow-up times.

All articles produced by the searches were entered into a Reference Manager database. All identified titles were screened by at least one of three reviewers (PS, SH, ES). If there was any doubt as to the relevance of the article to the review the article was included at this stage. All abstracts were read by at least two reviewers and consensus obtained. The reviewers held regular meetings to discuss the review process and the assessment of the literature.

Study population

It was clear from the outset that the studies were not reporting on entirely homogeneous populations. Rather than defining some theoretical ideal population and then determining how actual study populations would be biased to representations of that ideal, it was decided that the most important factor was that studies reported sufficient information on the principal factors known to affect patient prognosis so that it would be clear to which population the results were applicable.

The key factors known to affect patient outcome, and which were considered essential to report for the population studied, were treatment, recruitment dates and the established prognostic markers of PSA, clinical or pathological stage, biopsy or pathological Gleason grade, and surgical margins (where relevant). A TNM stage of T1–T3N0M0 or stage A–C on the Jewett–Whitmore system was an inclusion criterion so that, as a minimum, all studies included in the review reported clinical or pathological stage.

Treatment

It was noted whether the principal treatment (usually surgery, radiotherapy or watchful waiting) and also the proportion of patients who had had adjuvant or neoadjuvant treatment were recorded. Note that in none of the studies were patients randomised to treatment and it is likely that there are differences between populations selected for the different treatment modes.

Recruitment dates

Many factors that affect prognosis may change with time. A particular example in prostate cancer is the introduction of PSA testing, which has considerably changed the population of patients newly diagnosed with prostate cancer, who on average have lower-stage cancers than those diagnosed before the introduction of PSA testing.103 Biopsy methods and surgical techniques have also evolved. The staging classifications used in the TNM system have also undergone several minor changes. It is therefore important to know over what period of time the patients were recruited. The more recent studies are likely to be most relevant to new patients.

Baseline characteristics

It is important to describe the study population with regard to known prognostic factors. In particular, there were differences between studies in terms of the stages of the cancers included and whether postoperatively those who had had positive surgical margins were included or not. The availability of PSA measurements was also an indication, together with the recruitment dates, of whether the patient population may have been initially identified through PSA screening.

The reporting of diagnostic methods and ‘time zero’ were not recorded. For both issues the differences in populations arising through variations in these factors were considered to be small in comparison to those resulting from the advent of PSA screening, which has resulted in younger patients being diagnosed with lower-stage cancers. Furthermore, time zero, where stated, is generally defined as the start of treatment. In the traditional model of care the decision of whether to have radical treatment or not is made close to the time of diagnosis. It is only more recently that a different model of care has emerged, in which a patient is monitored and is possibly offered radical treatment at a later date, and this model is still unusual. Thus, generally, it is unlikely that there will be large discrepancies between the approaches to the definition of time zero.

Study attrition

It was apparent that the majority of studies were going to be retrospective and so the assessment of attrition had to be relevant to this type of study.For these studies, loss to follow-up was not the only issue to consider; the selection of cases was also important, on the basis of either complete follow-up data or complete baseline data. The question regarding baseline information was awarded a ‘yes’ if the total number of patients from which the study population was selected was given, together with reasons for patient exclusion. If some of this information was given, the question was ranked ‘partly’. Similarly, with loss to follow-up, a ‘yes’ was given only if either the number or the percentage lost to follow-up was reported or if the number of patients at risk was recorded at least one time point after time zero.

Biases due to such selection are difficult to assess from a publication. Ideally, the authors discussed what biases such selection may have introduced and we recorded whether they had done so.

Prognostic factor measurement

For a prognostic marker to be useful its measurement must be consistent. This means that there must be a well-defined and reproducible method of extraction and measurement. Some markers may be affected by how they are stored before measurement and so it is important to know that studies have considered this issue. We looked for a description of the measurement of the prognostic markers, with a particular emphasis on the novel markers. A full description of measurement methods was considered less important for the classical markers, for which methods are more established, although for PSA measurements there are different assays in use. Hayden and colleagues102 also consider the issue of how continuous variables are treated in the analysis in this section and we followed suit. In summary, categorising continuous variables leads to the loss of statistical power, and data-dependent categorisation leads to overoptimism. In the latter case, studies were graded ‘no’ on this issue. If the data were categorised, but using well-established groups such as are often used for PSA, the study was graded as ‘partly’ satisfying this question.

Outcome measurement

The most reliable outcome in prostate cancer is all-cause mortality but as most patients with prostate cancer do not die of the disease it is not a sensitive measure and is also highly dependent on the age distribution of the study population. The potential problem with prostate cancer survival as an outcome is ensuring that cause of death has been accurately determined.72,73

Because of the long average survival time of prostate cancer patients most studies in fact use freedom from biochemical (PSA) recurrence as the outcome measure. As discussed inChapter 1 (see sectionBiochemical failure), with PSA being a continuous measure the problem is the definition of PSA recurrence. There are, however, consensus recommendations for the definition of PSA recurrence following surgery and radiotherapy, and we recorded whether these had been used. Two definitions were allowed following radiotherapy as the original 1996 recommendation was changed in 2005.

It was also recorded whether a unique definition of PSA recurrence was used: it is important that the outcome is defined consistently so that the predicted outcomes are unambiguous.

Length of follow-up was not included in the quality assessment as this was an inclusion criterion for the review.

Confounding measurements

The most important confounders were considered to be the classical markers. In this section it was noted whether a multivariate analysis was reported that included all appropriate classical markers (dependent on whether the model was pretreatment or at surgery). At pretreatment the markers should include clinical stage, PSA and biopsy or pathological Gleason score. At treatment (only relevant for surgery) the markers should include clinical or pathological stage, pretreatment PSA, biopsy or pathological Gleason score and positive or negative surgical margins.

Treatment was another potential confounder but in the majority of studies all patients had the same principal treatment (usually surgery). Ideally, if some patients have had adjuvant or neoadjuvant treatment this should be included as a confounding variable, as should age if the end point is all-cause mortality. A recent review104 concluded that age is not a prognostic factor for prostate cancer outcome.

Analysis

In addition to an adequate description of the analysis, to determine whether there were sufficient data to assess the quality of the study the reporting that a univariate analysis had been undertaken was considered essential; this resulted in a ‘yes’ score and was used as an indication that the authors had undertaken a systematic analysis of their data.

The question regarding model building was relevant only to the multivariate models. Although there is some controversy regarding the optimum method of developing multivariate regression models all reasonable approaches were accepted (forward and backward removal of variables, all plausible variables), as long as variables were not introduced that were not included in the univariate analysis.

For a model to be considered adequate it had to include a time-to-survival analysis such as the Cox regression and have no other major inadequacies. Ideally, a multivariate analysis with novel and established markers was sought. Thus, if only a log-rank test of difference between survival curves was used (a univariate analysis) instead of multivariate regression analysis the maximum score was ‘partly’. Division of patients into groups and testing of survival differences using at-test were considered inadequate.

In total, there were 23 questions. Each question was scored as yes (y), no (n), partly clear (p), unsure (?) or not applicable (na). There was also an overall question on the conclusion for each domain.

The quality of each study was assessed by at least two of the three members of the research team (PS, SH, ES). There is an element of subjectivity in quality assessment, as well as a need for attention to detail as reporting methods and formats vary widely, so disagreement between the two reviewers was common. Regular discussion meetings were arranged to resolve uncertainty between the two members who had completed the assessment. The third team member attended the meetings when agreement could not be reached. A statistician (TY) provided additional support for the interpretation of the statistical models and validation of the quality assessment scores assigned by the two reviewers. It was always possible to reach a consensus among the team members.

It is important to recognise that, as with all forms of systematic review, our review may be influenced by publication bias. By this we mean that the findings from the individual studies that have been published might be different from the findings of individual studies that have not been published. The exclusion of smaller studies may have reduced the possibility of publication bias, but with the literature comprising retrospective case series the possibility of publication bias remains considerable. Furthermore, with several possible outcome measures available there is scope for selective outcome reporting.

Brief description of the prognostic marker

β-Catenin is an intracellular protein that is involved in intercellular adhesion at the cellular membrane and cell signalling in the nucleus. It has been implicated in prostate carcinogenesis primarily through modulation of androgen receptor activity. The loss of expression of membrane β-catenin has been associated with progression from benign to malignant prostate pathology.133 The definition of the marker and its distribution in the population studied are shown inTable 7.

Brief description of the objectives of the individual study identified

The primary aim of the identified study was to assess β-catenin as a prognostic marker in patients with localised prostate cancer treated with RP. Horvathet al.108 chose to investigate β-catenin expression as it is thought to have a significant role as a signal transduction molecule in both in vitro and in vivo models of prostate cancer. They attempted to define the pattern of β-catenin protein expression in the nuclei of normal, hyperplastic and malignant human prostate tissue to evaluate whether differences in expression in patients with cancer were related to disease progression. The basic study design characteristics are summarised inTable 8.

Quality of the individual study identified

Although the statistical analysis in this study is appropriate and the multivariate model includes the recognised classical markers, a weakness of the study is that the cut-point for differentiating between high and low β-catenin levels was determined within the data. This means that the results are likely to be overoptimistic as the β-catenin variable has been optimised to the data. At a value of 10 EPV the model just meets the minimum criterion in the quality assessment. However, with most of the variables entered into the model as dichotomous rather than continuous variables, an EPV of 10 is low and may lead to unreliable results. The overall concluding questions for each of the six subheadings are presented inTable 9.

Summary of the baseline characteristics of the sample

Horvath and colleagues used a sample of 232 participants who had had RP, 22% of whom also had some form of adjuvant therapy (hormone therapy, radiotherapy or orchidectomy). Participants all had clinically localised cancers and were pathological T1/T2 (47%) or T3/T4 (53%). The Gleason scores and PSA distributions appeared to be within the usual range. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual study identified

Table 10 presents a summary of the main statistical findings from the single study included in this section.

In a Cox univariate analysis β-catenin was found to be significantly prognostic for biochemical recurrence (p = 0.008. However, in a Cox multivariate analysis including the classical markers it was not (HR 1.4, 95% CI 0.8–2.3,p = 0.2).

Overall conclusions based on the results and quality of the findings

The results of this study indicate that, although β-catenin may be prognostic for biochemical recurrence following RP, its association with the existing widely used PSA marker means that it would not provide additional prognostic information. In addition, the quality issues raised above mean that the results are inconclusive.

Brief description of the prognostic marker

Prostatic acid phosphatase (PAP) is an enzyme produced by the prostate. Serum ACP was used as a marker for prostate cancer before the 1980s.134 However, with the development of assays for PSA, the use of ACP has diminished. The measurement methods, definitions and distributions of the marker in the populations studied are compared inTable 11.

Note that the proportion of patients in the elevated PAP groups, however defined, is relatively small, varying from 6.7%110 to 25%.112 With the exception of Hanet al.,110 all studies used a binary measure for ACP, sometimes resulting in a relatively small number of patients in the elevated group (e.g.n = 47109), and probably a small number of outcome events, making the results of the analyses less reliable.

Brief description of the objectives of the individual studies identified

Only three of the studies109–111 had a primary aim of assessing ACP as a prognostic marker. The aims of these studies were to: (1) identify those patients at most risk for local failure;109 (2) investigate the prognostic value of preoperative serum ACP in men with localised prostate cancer following radical retropubic prostatectomy;110 and (3) identify prognostic factors for prostate cancer treated by external beam radiation.111 Of the other studies, one112 was concerned with long-term survival in patients treated with radiotherapy and one,113 although concerned with prognostic factors in prostate cancer, did not specifically investigate ACP. The basic study design characteristics are summarised inTable 12.

Quality of the individual studies identified

The five studies varied in quality. The overall concluding questions for each of the six subheadings are presented inTable 13. The study considered to be of the highest quality for this novel prognostic marker was conducted by Hanet al.110 This was the most recent study involving ACP. Most of the other studies,109,111,113 being older, do not report PSA measurements and do not have this measurement available to enter as a covariate in multivariate models. Some also omit grade109,113 or stage.111 The only study to report a multivariate analysis including all classical markers was that of Hanet al.110 Some of the models also have a low number of events, for example that of Anscheret al.109 has only six. Perezet al.111 did not state the number of events but with a patient sample of 328 and 12 variables in their model the EPV is likely to be low.

Summary of the baseline characteristics of the sample

In only two studies109,110 did most patients (> 95%) have clinically organ-confined disease. In these two studies patients were treated with surgery. The other studies111–113 are all atypical of the majority of studies in this review in that most of the patients did not have organ-confined tumours; in one study all patients had extraprostatic disease.111 Two studies111,112 report relatively high proportions of patients with high-grade tumours (31% and 28% respectively), whereas one113 does not report grade. In all three studies with high proportions of patients with non-organ-confined disease, patients were treated with radiotherapy. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from individual studies identified

Table 14 presents a summary of the main statistical findings from the five studies included in this section.

Most of the univariate analyses on ACP level as a prognostic marker found it to be significantly associated with outcome (local recurrence,109 survival from metastatic failure and disease-free survival112,113), and some found it to be highly so (prostate cancer survival,p = 0.0001;112 survival from metastatic failure and disease-free survival, bothp < 0.001113). All of these last three analyses have a large number of outcome events. In three univariate analyses, ACP failed to reach significance at the 95% confidence level (metastases,109 local recurrence and any death113). These analyses include patients treated both with RP and with radiotherapy.

None of the multivariate analyses for which the outcome was survival from all causes of death showed ACP to be a statistically significant marker of outcome,111–113 but, as many patients will die from causes other than prostate cancer, the outcome is not highly sensitive to prostate cancer-specific markers. In the study by Zagarset al.,113 ACP was also not found to be significant in the multivariate analysis with an outcome of local recurrence.

In the other multivariate analyses with prostate cancer-specific outcome events – biochemical recurrence110 or local or distant failure109,111,113 or prostate cancer death112 – ACP was shown to be a significant prognostic marker in all with the exception of that of Perezet al.111 (p = 0.23). This analysis may be statistically weak. Although the EPV is not reported the number of patients (n = 328) and the number of variables in the model (n = 12) suggest that it may be low. This may also be a problem with one of the studies that found a positive result109 (EPV = 6), and although the EPV is large in the study by Zagarset al.113 the number of events in the elevated ACP group is likely to be very small as only 43 of 357 cases were in this category. It should also be noted that only one of these studies included all of the classical markers in the model110 and so the prognostic value of ACP in addition to that of the classical markers has only been demonstrated in one study. In this study ACP was found to be a highly significant marker (p < 0.001) for biochemical recurrence in patients who had RP.

Overall conclusions based on the results and quality of the findings

The studies for this marker are particularly heterogeneous, with two109,110 of the five studies based on patients with organ-confined tumours and the rest with all, or the majority of, patients with non-organ-confined tumours. In the former studies patients were treated with surgery, whereas in the latter patients were treated with radiotherapy. However, the results do not appear to be dependent on these factors. In the multivariate analyses four of five analyses that had prostate cancer-specific outcomes found ACP to be a statistically significant marker. However, only one of these analyses110 included all of the classical markers in the multivariate model. Although the number of events for this analysis was not stated, the large sample size and the fact that ACP was entered in the model as a continuous variable suggest that the study was statistically well powered. Thus, although the direction of evidence from several studies suggests that ACP is prognostic of prostate cancer outcomes, there is only one study that shows that it is prognostic independently of the established markers.

Brief description of the prognostic marker

Androgen function is mediated by the androgen receptor, which is a ligand-dependent steroid hormone transactivation factor located on the X chromosome.115 Namet al.114 hypothesised that CAG repeats may be associated with prognosis as it has been shown in other studies that men with ≤ 18 CAG repeats have an increased risk for developing prostate cancer compared with men with a longer CAG sequence and also have a 2.1-fold increased risk for developing advanced-stage or high-grade prostate cancer.135 The measurement methods, definitions and distributions of the marker in the populations studied are compared inTable 15.

Note that the proportion of patients with ≤ 18 CAG repeats in the study by Namet al.114 is relatively small (n = 39). In the study by Powellet al.115 the distribution of the marker according to the groups used in the analysis is not stated, but if the three groups are of similar size this should not be a problem as there are 711 patients in total.

Brief description of the objectives of the individual studies identified

Both studies had the primary aim of assessing the prognostic marker. Namet al.114 examined the significance of the CAG repeat polymorphism of the androgen receptor gene for predicting biochemical progression among patients treated by RP for clinically localised prostate cancer. The hypothesis was that a high level of androgen receptor activity associated with short CAG repeats may be important in prostate cancer progression. Powellet al.115 also examined the impact of the number of CAG repeats in the androgen receptor on disease progression (not defined) among men with prostate carcinoma following prostatectomy. The basic study design characteristics are summarised inTable 16.

Quality of the individual studies identified

A summary of the quality assessment for the studies is shown inTable 17. Both studies were of reasonable quality. However, in the study by Namet al.114 there are only a small number of patients with ≤ 18 CAG repeats. This weakens their analysis and is a particular issue in the model in which CAG repeats is used as a binary variable. In the study by Powellet al.115 it is not clear exactly what the end point is: biochemical recurrence or biochemical or clinical recurrence.

Summary of the baseline characteristics of the sample

The patient populations appear similar with all of the patients having clinically localised cancers, just over 40% of patients having pathologically organ-confined tumours, and around 14% having high-grade tumours (Gleason score 8–10), although for Powellet al.115 the Gleason score is pathological rather than clinical. In both studies patients were treated with RP. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual studies identified

Table 18 presents a summary of the main statistical findings from the two studies included in this section.

In the univariate analysis, Namet al.114 did not find the number of CAG repeats to be prognostic for biochemical recurrence-free survival (p = 0.80). Both studies present multivariate analyses. Both include the classical markers of PSA, Gleason grade and stage. Both studies also present two analyses, with the number of CAG repeats entered into the models in a different form. Namet al.114 entered CAG repeats as a dichotomous variable and as a continuous variable. In neither analysis was it a significant predictor of outcome. Powellet al.115 used the same two categories as Namet al.114 for CAG repeats but with the opposite category entered as the baseline. Thus, the direction of the risk reduction is actually the same as for Namet al.:114 those with ≤ 18 CAG repeats are at lower risk for disease recurrence and this result was statistically significant at the 95% confidence level (p = 0.03). The fact that this result was significant, whereas that for Namet al.114 was not, may be due to the larger sample size. The results of the other analysis by Powellet al.,115 which examined the increase in risk for each category of CAG repeats (≤ 18, 19 22 and ≥ 22), were not significant (p = 0.32). This analysis may be considered less reliable as it treats three categories of the CAG repeat variable as a continuous variable in the analysis.

Overall conclusions based on the results and quality of the findings

Although otherwise of reasonable quality, the results of the study by Namet al.114 might be considered less reliable because of the small number of patients with short CAG repeats (≤ 18 CAG repeats). In the study by Powellet al.115 with a larger patient sample, and possibly a larger proportion in the group with ≤ 18 repeats, an analysis with the number of CAG repeats entered as a binary variable did show a significant association between this marker and disease progression. Another analysis by Powellet al. in which the marker was entered in a different format did not show a significant association but this may be less reliable. The results are inconclusive as to whether the number of CAG repeats is prognostic of prostate cancer outcome.

Brief description of the prognostic marker

Creatinine is a by-product of muscle metabolism. It is widely used to measure kidney function. It was hypothesised by Merseburger116 that in localised disease creatinine could be associated with good prognosis as a high proportion of low-volume cancers are in enlarged glands, which may be associated with renal insufficiency and creatinine elevation. The definitions and distributions of the marker in the populations studied are shown inTable 19.

Note that in both studies the proportion of patients with a high level of creatinine (> 1.3 mg/dl,116 > 1.5 mg/dl117) is relatively small. This is an issue, particularly in the analyses carried out by Zagarset al.117 and in a univariate analysis by Merseburgeret al.,116 in which patients are grouped according to their level of creatinine, with only a very small number of patients in the elevated creatinine group.

Brief description of the objectives of the individual studies identified

Only the study by Merseburger116 had a primary aim of assessing this prognostic marker. Merseburger116 investigated the ability of creatinine to predict PSA recurrence using Cox regression analysis. Zagarset al.117 studied outcomes for patients with stage C cancer. The basic study design characteristics are summarised inTable 20.

Quality of the individual studies identified

The two included studies varied in quality (Table 21). Zagarset al.117 did not conduct a multivariate analysis but rather compared survival curves for patients with normal and elevated creatinine. There were only 28 patients in the elevated creatinine group and so the number of events is likely to be very low. Merseburger116 did undertake multivariate analysis that included several covariates including Gleason grade, PSA and stage. It did not, however, include surgical margins. The multivariate model was not fully presented and it is not entirely clear exactly which covariates were included in the model; therefore, although there are a reasonable number of outcome events (n = 130) the EPV may be below 10.

Summary of the baseline characteristics of the sample

The clinical stage of the participants was very different in the two studies. Merseburger116 used a sample that was almost entirely clinically organ confined, whereas the participants in the Zagarset al.117 study were all stage C or non-organ confined. We were unable to compare the participants according to Gleason score or PSA level as these were not reported by Zagarset al.117 The patients in the Merseburger116 study were treated with surgery where those in the Zagarset al.117 study were treated with radiotherapy. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual studies identified

Table 22 presents a summary of the main statistical findings from the two studies included in this section.

Zagarset al.117 conducted three univariate analyses using the log-rank statistic to compare survival curves with three different end points: all deaths, any disease relapse and local control. As previously discussed there were only a small number of patients in the elevated creatinine group (n = 28) and so the results may be unreliable. Of these three analyses only one, that with any disease relapse as the outcome measure, showed a statistically significant association between elevated creatinine and outcome (p = 0.05).

Merseburger116 also reported a log-rank analysis to compare survival by creatinine stratified into three groups. The curves were not statistically significantly different (p = 0.845). Again, there were only a small number of patients in the elevated creatinine group (n = 42). In the multivariate analysis with creatinine entered into the analysis as a continuous variable with several other covariates including PSA, Gleason grade and stage, Merseburger116 found no significant effect of creatinine on PSA recurrence (p-value not stated). The analysis may be statistically weak with a low EPV.

Overall conclusions based on the results and quality of the findings

These two studies were carried out on different patient groups (organ confined and non-organ confined) and patients had different treatments. The results of neither study indicate that creatinine is a useful prognostic marker for prostate cancer. However, the results cannot be considered conclusive as both studies had statistical weaknesses.

Brief description of the prognostic marker

Cytochrome P450 3A4 (CYP3A4) is a member of the cytochrome P450 supergene group. It is thought to be involved in the oxidative deactivation of testosterone to biologically less active metabolites. Testosterone is a major contributor to prostate cancer progression. A germline genetic variant in the 5′ regulatory region of theCYP3A4 gene (A to G transition) on chromosome 7 has been reported and named asCYP3A4*1B (otherwise known in the literature as –392A > G andCYP3A4-V). ThisCYP3A4 genetic variant was the prognostic factor of consideration in this section. The definition and distribution of the marker in the population studied are shown inTable 23.

Brief description of the objectives of the individual study identified

The primary aim of this study was to assessCYP3A4 genotypes as prognostic markers. The study examined the survival of men with localised prostate cancer who had undergone RP to evaluate whetherCYP3A4*1B was associated with disease progression and whether it was independently prognostic of outcome. The basic study design characteristics are summarised inTable 24.

Quality of the individual study identified

An important quality item that needs to be considered in the interpretation of the study results is that the number of EPV is unknown. In common with many studies there was poor reporting of the effects of missing data on the results, the authors did not use the internationally agreed definitions of PSA recurrence after prostatectomy and the methods of storage of materials were not reported. Generally the study was of adequate quality. The overall concluding questions to each of the six subheadings are presented inTable 25.

Summary of the baseline characteristics of the sample

Powell and colleagues used a sample of 737 participants in the analysis, all treated with RP. Participants were all clinical stages T1/T2. Pathologically, 50% of the white men (WM) and 37% of the African American men (AAM) had organ-confined tumours. More of the AAM than the WM had high-grade (Gleason score 8–10) tumours (17% and 13% respectively) and fewer had low (≤ 10 ng/ml) preoperative PSA levels (WM 67%; AAM 57%). Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual study identified

The association betweenCYP3A4 genotypes and biochemical progression was examined using a multivariate Cox proportional hazards model that included the classical prognostic markers. Although a model including both WM and AAM is presented, the authors argue that the strong association betweenCYP3A4 genotype and race means that race-stratified models should be used to avoid co-linearity. These are also presented.Table 26 presents a summary of the main statistical findings from this study.

Powellet al.118 report several analyses that look at the effect of the G alleles in different ways. The analyses including all men showed a significant association between theCYP3A4*1B genotype and progression-free survival, with the most statistically significant result obtained with the number of copies of G allele (p = 0.0049). The presentation of race-stratified results is justified by the author by the strong association found between the AA, AG and GG alleles and race (p = 0.00002). They suggest that the G allele was not associated with biochemical progression-free survival in AAM. In WM some of the associations were of marginal significance at the 95% confidence level: the number of copies of the G allele in a dose model (p = 0.03) and the comparison of men with the AA genotype versus men with AG and GG (p = 0.04).

Overall conclusions based on the results and quality of the findings

This single study presents some evidence in support ofCYP3A4 genotype as a prognostic marker in localised prostate cancer. TheCYP3A4 variant was shown to be significantly more prevalent among AAM but was not prognostic in this group.

Brief description of the prognostic marker

DNA ploidy is a test to measure the DNA content within tumour cells. The definitions and distributions of the marker in the populations studied are shown inTable 27.

Brief description of the objectives of the individual studies identified

The study by Lieber and colleagues106 had the primary objective of investigating whether measurement of DNA ploidy provided additional unique prognostic information beyond the customary parameters of tumour stage and grade for patients with prostate cancer. Blute and colleagues105 were interested in predicting biochemical failure following prostatectomy, and the main aim of the study by Siddiqui and colleagues119 was to assess whether age at treatment was a predictor of survival following prostatectomy. The basic study design characteristics are summarised inTable 28.

Quality of the individual studies identified

The principal limitation of all of these studies is that an absolute measure of PSA is not included in any of the multivariate models, thus limiting the conclusions that can be reached regarding the prognostic value of DNA ploidy in the presence of established markers. The Lieberet al. study106 pre-dates routine PSA measurement, but it is not clear why it was omitted from the models of Bluteet al.105 and Siddiquiet al.119 The Bluteet al.105 model does, however, include a measure of PSA doubling. Two of the studies105,119 have a very large number of participants and therefore should give good statistical power, although the number of outcome events is not reported by Siddiquiet al.119 The Lieberet al. study106 is smaller than the other two studies but reports an adequate number of events, and, in a rare example of good practice, also reports the number of patients and events in each marker category. Thus, we know that 283, 181 and 30 patients had diploid, tetraploid and aneuploid tumours respectively, with 60, 90 and 24 respectively experiencing disease progression.

A major drawback of the Siddiquiet al. study is that it is not clear in what form ploidy is entered into the statistical analysis (i.e. diploid/non-diploid), which means that the results are difficult to interpret. The overall concluding questions to each of the six subheadings are presented inTable 29.

Summary of the baseline characteristics of the sample

In all three studies patients had been treated with RP. However, the clinical stage of the patients in the Lieberet al. study106 was more advanced, with only 52% having organ-confined tumours compared with around 90% for those in the Bluteet al.105 and Siddiquiet al.119 studies. The proportion of patients with pathologically high-grade cancers was not dissimilar across the studies, ranging from 4%105 to 9%.106 Additional summary characteristics are provided inAppendix 7.

Brief description of the results from individual studies identified

Table 30 presents a summary of the main statistical findings from the three studies included in this section.

In the univariate analyses of Bluteet al.105 and Lieberet al.106 tetraploid and aneuploid tumours are compared with diploid tumours, and Bluteet al. also carry out this comparison in multivariate analysis. In the multivariate analysis Lieberet al. enter a binary ploidy variable (non-diploid versus diploid). In the Siddiquiet al.119 study only one ploidy variable is entered into the analyses and this is not defined. Lieberet al. and Siddiquiet al. both examine ploidy as a prognostic marker for survival from clinical progression (although not necessarily similarly defined) and prostate cancer death, whereas the end point for the Bluteet al. study is biochemical or clinical (local or distant) progression. Lieberet al. also use crude survival as an end point.

All studies present univariate analyses and for all studies and all outcomes ploidy was found to be a significant predictor, in many analyses highly so (seeTable 30).

In the multivariate analyses two studies106,119 found ploidy to be highly significantly prognostic for clinical progression and prostate cancer death (p-value ranged from 0.0011 to < 0.0001). The Lieberet al.106 model included grade and stage, and the Siddiquiet al.119 model grade and pathological variables including stage T3. Neither study included PSA. An analysis by Lieberet al.106 did not find ploidy to be prognostic for all-cause death, but this outcome is less sensitive to prostate cancer markers than the others.

Bluteet al.105 found ploidy to be significantly prognostic for biochemical or clinical recurrence, but marginally so at the 95% confidence level (tetraploid versus diploid,p = 0.05, anueploid versus diploid,p = 0.04). This analysis included similar covariates to that of Siddiquiet al.119 but with the addition of PSA doubling.

Overall conclusions based on the results and quality of the findings

Although two studies106,119 found DNA ploidy to be highly significantly prognostic for prostate cancer outcomes, another105 found it to be only marginally significant. The fact that the data used in the study by Blute and colleagues105 are probably included in the analysis of Siddiquiet al.119 makes this more puzzling. All three studies are large and so are more likely to be statistically reliable than many other studies included in this review.

The most obvious differences between the analyses of Bluteet al.105 and Siddiquiet al.119 are that Siddiquiet al. had no measure of PSA in their analysis and used clinical outcomes only whereas Bluteet al. included a measure of PSA (PSA doubling) and used an outcome of biochemical or clinical progression. Vollmeret al.107 suggest that pathological variables may be better at predicting clinical outcomes, whereas PSA is a better predictor of biochemical recurrence. This might explain the results. Neither analysis includes the usual absolute measure of preoperative PSA, although these data are presented in the baseline statistics and therefore must be available in the data set. The relationship between DNA ploidy and clinical and biochemical outcomes with and without PSA as a covariate could be explored in this data set (Siddiquiet al.119 and/or Bluteet al.105 if not the same) and might resolve the contradictions apparent from the current analyses.

Brief description of the prognostic marker

Vitamin D binds to the vitamin D receptor in the prostate and forms a complex with other factors such as retinoid X receptors. It is believed that this complex binds to vitamin D response elements on DNA and regulates the transcription of a number of genes involved in cell growth, differentiation and metastasis. Prostate cancer mortality rates appear to increase significantly with decreased ultraviolet radiation exposure, which decreases vitamin synthesis in the skin. This has led to the hypothesis that those men with a vitamin D deficiency might be at increased risk of prostate cancer. The definition and distribution of the marker in the population studied are shown inTable 31.

Brief description of the objectives of the individual study identified

Williamset al.120 aimed to analyse the associations between germline genetic variation in the vitamin D receptor with clinical and pathological factors at the time of prostate cancer diagnosis and progression after RP. The basic study design characteristics are summarised inTable 32.

Quality of the individual study identified

In general this is a good quality study but there are some issues that need to be considered when interpreting the results. First, the end point, disease recurrence, is not defined. It is not even clear if a consistent definition was used. Also, the number of events is not stated. It is possible that there is a low EPV rate, particularly in the second analysis, which is conducted on white men only with separate models for organ-confined and locally advanced tumours. The patient samples in these two models were 213 and 215 respectively. The overall concluding questions to each of the six subheadings are presented inTable 33.

Summary of the baseline characteristics of the sample

Williamset al.120 used a sample of 738 participants in the analysis (428 WM and 310 AAM), all of whom were treated with RP. Participants were all clinical stages T1/T2. More of the AAM than the WM had high-grade (Gleason score 8–10) tumours (16.5% and 12.7% respectively) and more also had pathologically non-confined tumours (WM: 50.2%,n = 213; AAM: 62.6%,n = 215) and high (≥ 20 ng/ml) preoperative PSA levels (WM 10.3%; AAM 22.9%). Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual study identified

The association between Bsm1 genotypes and progression was examined using a multivariate Cox proportional hazards model. The model was stratified by race to avoid multicolinearity effects between race and genotype, as the two were associated.Table 34 presents a summary of the main statistical findings from this study.

In neither model were Bsm1 genotypes significant predictors of progression; however, they were classified [according to the number of copies of the B allele (allele dose); the individual genotypes included in the same model (genotype specific); comparing bb with Bb plus BB (dominant effect of B); and comparing bb plus Bb with BB (recessive effect of B)].

A graphical analysis had suggested a differential effect of Bsm1 by pathological stage. In a further exploratory analysis a Cox regression model on WM was stratified by organ-confined status. In this analysis Bsm1 status showed high HRs for WM with organ-confined tumours, although they were not significant. For men with locally advanced tumours, the B allele was associated with a lower recurrence risk, with the HRs of marginal significance at the 95% confidence level.

It was reported that similar results were obtained for the Taq1 genotype but none of the analyses were shown.

Overall conclusions based on the results and quality of the findings

The primary analysis indicated that vitamin D receptor gene polymorphisms are not prognostic in prostate cancer. A secondary analysis on WM stratified by pathological organ-confined status did yield statistically significant associations between the Bsm1 genotype classifications and progression, with the B allele having an opposite effect in the two groups, but the statistical power of the analysis may have been weak. The authors claim that the complexity of the biological effects of vitamin D in experimental studies supports the possibility of complex clinical effects. The plausibility of such effects would need to be considered before pursuing vitamin D receptor gene polymorphisms as a prognostic marker in prostate cancer.

Brief description of the prognostic marker

Two studies122,123 examined whether the primary Gleason grade could differentiate between the prognostic outcomes of patients with a Gleason score of 7, a patient group that has particularly heterogeneous outcomes, i.e. whether there was a difference between patients whose Gleason pattern was 4 + 3 and those whose pattern was 3 + 4. These studies are shown inTable 35.

Three studies107,121,124 examined whether some measure of the amount of high-grade cancer was prognostic of outcomes. The measures included percentage of tumour grade 4 or 5,121,124 length of high-grade tumour124 and the presence or not of grade 5 cancer in the primary and secondary prostatectomy specimens.107 Samples were taken from TURP, biopsy and prostatectomy specimens. Details, as far as provided by the study authors, of the different definitions and measurement methods of these different measures of high-grade cancer are shown inTable 36.

Egevadet al.121 also calculated a modified Gleason score, which was the sum of the dominant (primary) and worst Gleason grades.

Brief description of the objectives of the individual studies identified

Four of the studies121–124 had a primary aim of assessing the prognostic value of different methods of measurement or scoring of Gleason grade assessments of tumour differentiation.

Two studies122,123 examined whether the primary Gleason grade could differentiate between the prognostic outcomes of patients with Gleason score 7, a patient group that has particularly heterogeneous outcomes, i.e. whether there was a difference between patients whose Gleason pattern was 4 + 3 and those whose pattern was 3 + 4. Note that Gonzalgoet al.122 selected a population who were all biopsy Gleason score 7, whereas Tollefsonet al.123 selected a population who were all pathologically Gleason score 7. Egevadet al.121 also included an analysis of Gleason pattern in Gleason score 7 patients but as this analysis had fewer than 200 participants it did not meet the inclusion criteria for Gleason score 7.

Both Egevadet al.121 and Viset al.124 had the aim of examining the amount of high-grade cancer as a prognostic factor, whereas Vollmeret al.107 was interested in the relative importance of anatomic and PSA factors for prostate cancer outcomes.

The basic study design characteristics are summarised inTable 37.

Quality of the individual studies identified

Perhaps because the focus of most of these studies was on different measures of Gleason grade, only one study123 reports a multivariate analysis including ‘known risk factors’ as well as the novel Gleason measure, although the former are not specified. The statistical analysis in two of the studies122,123 is also poorly reported and therefore difficult to assess. The number of events or EPV is low in some studies. Both the Vis and Vollmer studies have adequate EPV in their final models according to our criteria but that is only because they have removed most variables. The initial models that were used to select variables for the final model will have had low EPV and therefore may not have been reliable. In the analysis by Gonzalgoet al.122 the number of events is not stated, but there are relatively small numbers of patients in two of the four groups (C:n = 32; D:n = 36) and so there are likely few events for these patients on which to base the analysis. The EPV is adequate in the study by Egevadet al.121 and although the number of events is not stated by Tollefsonet al.123 the large sample size suggests that it is also adequate. The overall concluding questions to each of the six subheadings are presented inTable 38.

Summary of the baseline characteristics of the sample

With the exception of Egevadet al.121 the patients in all of the studies had more than 90% organ-confined tumours. The study population in Egevadet al.121 was different from the others, with prostate cancer diagnosed at TURP because of obstructive symptoms. In total, 83% of these patients had organ-confined tumours. These patients also had a high proportion of high-grade cancers (31% pathologically Gleason score 8–10). The Gonzalgoet al.122 and Tollefsonet al.123 studies included only patients with Gleason score 7. The patients in all studies, with the exception of those in the Egevadet al.121 study who had deferred treatment following TURP, were treated with RP.

Brief description of the results from the individual studies identified

Table 39 presents a summary of the main statistical findings from the two studies on Gleason patterns 3 + 4 and 4 + 3 included in this section.

Primary Gleason pattern in Gleason score 7 patients

In the study by Gonzalgoet al.122 patients (all biopsy Gleason score 7) were divided into four groups according to whether they were Gleason pattern 3 + 4 or 4 + 3 at biopsy and after prostatectomy. The prognosis of these four groups in terms of freedom from biochemical recurrence was compared using a log-rank test to test the significance of differences between pairs of the four survival curves, and also using an overall test of the four curves. Survival at 5 years ranged from 89% for group A to 55% for group D. Not all of the pairs of curves were significantly different from each other (seeTable 39), but groups A and B (both biopsy Gleason pattern 3 + 4) had significantly different outcomes (p = 0.002) as did groups C and D (both biopsy Gleason pattern 4 + 3) (p = 0.03). The latter analysis may be unreliable because of the small numbers of patients in groups B and C. The overall log-rank statistic for all curves was significant (p < 0.0001). A comparison between all those with clinical Gleason pattern 3 + 4 and those with pattern 4 + 3 was not made.

In a univariate analysis Tollefsonet al.123 found significant differences in prognosis between patients with biopsy Gleason pattern 3 + 4 and those with Gleason pattern 4 + 3 with outcomes of biochemical recurrence-free survival (p < 0.0001), systemic recurrence-free survival (p < 0.002) and cancer-specific survival (p = 0.013). In a multivariate analysis ‘correcting for known risk factors’, primary Gleason score was an independent significant predictor of biochemical failure (p < 0.0001), systemic recurrence (p = 0.002) and cancer-specific survival (p = 0.029). The lowerp-values for the relationship between primary Gleason score and outcome in both univariate and multivariate analyses when the outcome was survival rather than disease recurrence (even biochemical or systemic) may be due to the lower number of events for the survival outcome compared with the recurrence outcomes, rather than any difference in the strength of the relationship. The number of events is not reported in the study but, after 10 years, although around 95% of patients have survived prostate cancer death, only around 50% are biochemical progression free.Table 40 presents a summary of the main statistical findings from the three studies included in this section on the amount of high-grade cancer.

Amount of high-grade tumour

In univariate analysis both Egevadet al.121 and Viset al.124 found the percentage of high-grade tumour to be significantly prognostic for prostate cancer death (p < 0.001) and biochemical progression (p < 0.001) respectively. Using multivariate analysis Egevadet al.121 examined the performance of the percentage of high-grade tumour in a model with Gleason score but no other covariates, in which it was significant (p = 0.002). Viset al.124 found the percentage of high-grade tumour to be significantly prognostic for biochemical progression (p < 0.001) in a multivariate model that included PSA. Gleason score was removed from the model because of non-significance.

Viset al.124 also tested a variable of length of high-grade cancer from the biopsy core. In univariate analysis it was significant for the outcomes of survival from biochemical and clinical progression. In multivariate analysis it was significant for biochemical survival with PSA as the only covariate, but for the outcome of clinical recurrence all of the other covariates were removed from the model using a stepwise process and so the result reported is the same as that for the univariate analysis.

Vollmeret al.107 found the presence of Gleason grade 5 in either the primary or secondary prostatectomy specimen to be significantly prognostic for prostate cancer death (p = 0.0096) in a multivariate model with no classical markers but with percentage of tumour in the prostate.

Modified Gleason score

Egevadet al.121 also found a modified Gleason score [sum of the dominant (primary) and worst Gleason grades] to be prognostic of prostate cancer death in univariate analysis (p < 0.001) and in a multivariate model with Gleason score (p < 0.001).

Overall conclusions based on the results and quality of the findings

Two studies122,123 showed that primary Gleason grade in Gleason score 7 patients was prognostic, although Gonzalgoet al.122 report only a univariate analysis. In the multivariate analysis reported by Tollefsonet al.123 primary Gleason grade was prognostic for biochemical failure (p < 0.0001), systemic recurrence (p = 0.002) and cancer-specific survival (p = 0.029). This study was likely to have been adequately powered but poor reporting of the analysis makes it difficult to assess. The results needed to be confirmed.

Gleason pattern has already been used by Hanet al.140 in a prognostic model, which is discussed inChapter 6. If further prognostic information could be derived from what is routinely collected data this would clearly be advantageous.

Two studies121,124 found the percentage of high-grade tumour to be prognostic for prostate cancer death and biochemical progression respectively, and in both it appeared to outperform Gleason score. In neither study was percentage of high-grade tumour tested in a multivariate model with all of the established markers and so its additional prognostic value is not established. Viset al.124 also found length of high-grade cancer to be prognostic in univariate and multivariate analysis, but most covariates were removed from the analysis and so its performance in the presence of the classical markers is not shown. Vollmeret al.107 found the presence of Gleason grade 5 to be significantly prognostic for prostate cancer death (p = 0.0096), but this marker also was not tested in a multivariate model with classical markers. Thus, although measured differently, all measures of amount of high-grade cancer were found to be prognostic, but none was tested in models including all of the established markers.

One study121 found a modified Gleason score [sum of the dominant (primary) and worst Gleason grades] to be prognostic of prostate cancer death.

All of the studies in this section report a variety of novel Gleason measures to be significantly prognostic of various prostate cancer outcomes. However, only one study123 was (probably) tested in models including all of the established markers and the quality of the studies was generally worse than average. The positive results, combined with the relative ease with which some of these measures could be applied as the data are currently collected, suggest that more rigorous studies would be worth undertaking.

Brief description of the prognostic markers

Tissue microarrays are emerging as powerful tools to rapidly analyse the clinical significance of new molecular markers in human tumours. Ki67 LI (labelling index) is a nuclear antigen that is present throughout the cell cycle but not at rest (GO phase) or in the early G1 phase.141 Antibodies to the p53 protein bind both normal (wild type) and mutant forms.141 The Bcl-2 oncoprotein inhibits apoptosis, such that its overexpression leads to increased cell growth.141 Syndecan-1 (also known as CD138, CD138 antigen, SDC, SYND1, syndecan-1 precursor) is a multifunctional transmembrane heparan sulfate proteoglycan that is present on many cell types and which mediates growth factor binding.142

The definitions and distributions of the markers in the population studied are shown inTable 41.

Brief description of the objectives of the individual study identified

The study examined the expression of the molecular markers Ki67, Bcl-2, p53, syndecan-1 and CD10 for prognostic significance. The basic study design characteristics are summarised inTable 42.

Quality of the individual study identified

The study does poorly on many quality assessment criteria. One important issue is recognised by the authors, that is the heterogeneity of the study cohort. Participants were accrued over a considerable period of time between 1971 and 1996. This means that there were different staging, treatment and follow-up methods. There is also heterogeneity in how disease progression was defined, with it being defined clinically in some patients and biochemically (by PSA) in others. Furthermore, the definition of PSA failure is not given and may have been variable.

The statistical analysis may also be weak as there are relatively small numbers of patients in each of the ‘high-risk’ marker categories and thus the number of events in these groups is likely to be small (Table 43). With the exception of pathological grade, classical markers were not included in the model and therefore the prognostic significance of these markers over those in current use is not demonstrated.

Summary of the baseline characteristics of the sample

The study involved 551 participants who had been treated with RP or TURP. All participants were organ confined at clinical stage. At pathological stage there were still a greater number of organ-confined (71.9%) compared with non-organ-confined participants (18.5%), with a small number of participants having missing data (9.6%). Only Gleason grade (as opposed to Gleason score) was reported because of the small size of the specimens. PSA levels were not reported. The failure to measure and report this information limits the comparison of this study with other prognostic studies involving other types of markers. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual study identified

Table 44 presents a summary of the main statistical findings from the single study included in this section.

Zellwegeret al.125 reports thep-values of the markers (Ki67, Bcl-2, p53, syndecan-1 and CD10) in three different Cox regression models, each with a different end point: progression, overall survival and tumour-specific survival. Markers were only introduced into the multivariate model if they were found to be statistically significant predictors of that outcome in univariate analysis. Gleason grade was the only classical marker entered into the statistical model. Marker Ki67 LI (p = 0.023) was the only marker found to be statistically significant for all end points in univariate analysis. It remained significant in multivariate analysis for the end points of overall survival and tumour-specific survival, but with Gleason score as the only classical marker in the model. CD10 was not significant in any of the univariate analyses and thus was not tested in the multivariate models.

Bcl-2 and p53 were not significant in any of the multivariate analyses. The marker syndecan-1 was of marginal significance for tumour-specific survival (p = 0.051).

It should be noted that Zellwegeret al.125 reported many significant associations between the markers and this may have affected their individual performances in the multivariate models.

Overall conclusions based on the results and quality of the findings

The weaknesses of this study make the results inconclusive. Of the markers studied Ki67 LI appeared to be the most strongly associated with the study end points and in particular tumour-specific survival (p = 0.023). p53 was of marginal significance for this end point (p = 0.051).

Brief description of the prognostic marker

These studies all used some measure of the proportion of the prostate affected by cancer as a prognostic marker. Four studies124,126–128 achieved this by counting the number of biopsy cores containing cancer, usually expressing this as a proportion of cores affected. Two studies107,121 used a measure of the percentage of the prostate involved with cancer, estimated from the surgical specimens; however, the Egevadet al.121 study used TURP specimens whereas in the Vollmeret al.107 study patients had RP. The definitions and the marker distributions in the different studies are shown inTable 45.

Brief description of the objectives of the individual studies identified

It is important to note that only two of the studies126,128 had a primary aim of assessing positive biopsy cores as a prognostic marker. Antuneset al.126 evaluated the prognostic value of the percentage of positive biopsy cores (PPBC) in determining the pathological features and biochemical outcome of patients with prostate cancer treated by R.P. Seleket al.128 aimed to determine the utility of the PPBC in predicting PSA outcome after external beam radiotherapy alone. Potterset al.127 assessed the outcomes of men undergoing prostate brachytherapy and evaluated factors that could impact on disease-specific survival. Viset al.124 and Egevadet al.121 investigated the predictive value of the amount of high-grade cancer (Gleason growth patterns 4/5) in the biopsy following RP and TURP, respectively. Vollmeret al.107 compared anatomic and PSA factors as prognostic markers.

Quality of the individual studies identified

One of the key failings amongst these studies is the omission of classical markers in the reported multivariate models,107,121,124,128 usually because of stepwise removal of variables rather than lack of data. The statistical power of some of the studies107,124,128 in terms of EPV may also be weak, although in the case of Seleket al.128 and Viset al.124 the assessment criterion of an EPV of at least 10 in the final model was met. The study by Antuneset al.126 avoids both of these issues and is overall probably the best quality study for this marker. In the four studies that had an end point of biochemical recurrence124,126–128 only one used a recognised definition;128 the definition therefore varied across the studies, although at least all of the studies were internally consistent. The overall concluding questions to each of the six subheadings are presented inTable 46.

Two studies107,128 failed to present sufficient data to assess the adequacy of the analysis.

Summary of the baseline characteristics of the sample

Three of the studies107,124,126 used RP treatment. Potterset al.127 used brachytherapy (some in combination with radiotherapy), Seleket al.128 used radiotherapy alone and Egevadet al.121 used TURP. The studies varied in population size ranging from 203 to 1449 (Table 47). The largest study was conducted by Potterset al.127 and the smallest by Vollmeret al.107

In evaluating the results of the six studies it is important to consider the differences in sample characteristics (e.g. stage, Gleason score and PSA distributions). The clinical stage of the participants was provided in all six studies. More than 98% of the samples in five of the studies were organ-confined cancers at clinical stage. The exception was the study of Egevadet al.,121 in which 17% of cancers were non-organ confined and whose participants also had a high proportion of high-grade cancers (35% Gleason score 8–10). This study pre-dates PSA screening and the patients had their tumours detected on TURP carried out for obstructive symptoms. The distributions of Gleason and PSA scores (where reported) were similar across the other studies. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual studies identified

Table 48 presents a summary of the main statistical findings from the six studies included in this section.

All of the studies provided a Cox multivariate analysis of the data. As shown inTable 48 all studies used an end point of biochemical recurrence but the definition varied between studies, and in the Seleket al.128 study patients were treated with radiotherapy and so PSA behaviour following treatment is different from that in the other studies. Viset al.124 also used an outcome of clinical progression.Table 48 shows the different clinical and pathological classical markers entered into the statistical models across the four studies: all included the classical markers in their models with the exception of the Selek analysis, which does not include stage.

All of the studies that reported a univariate analysis124,126,128 found PPBC to be prognostic. However, only two studies126,127 showed PPBC to be prognostic in multivariate analysis, both for PSA survival. Of these, one126 has a large EPV ratio (30) suggesting a statistically strong analysis and the other,127 although it is not stated, is likely to be more than adequate because of the sample size (n = 1449). The studies of Antuneset al.126 and Potterset al.127 both also include all of the classical prognostic markers, suggesting that the proportion of positive biopsy cores may add prognostic value to that of the established markers.

The multivariate results of three analyses in two studies124,128 indicate that PPBC is not prognostic. The study end points were biochemical progression and clinical progression. The number of events in both of these studies may have been low, making the analyses less reliable. The analyses of Seleket al.128 and Viset al.124 met the quality criterion of an EPV of at least 10, but for Seleket al.128 it was only 13 and not all continuous variables were treated as continuous, thus weakening the analysis. Viset al.124 achieved adequate EPV in their final models by eliminating most variables. However, there were only 39 events in total and so the EPV for the full models (when the number of positive cores would have been eliminated for non-significance) would have been low.

Table 49 presents a summary of the results of the studies concerning the percentage of cancer in the specimen.

Brief description of the prognostic markers

Both studies used linear regression to calculate the rate of rise in the PSA level (PSAV) in the year before diagnosis129 or 2 years before treatment130 using all available values. PSADT is the time that it takes for the PSA value to double; this was calculated by Senguptaet al.130 using log-linear regression. The definitions and the marker distributions are shown inTable 50.

Brief description of the objectives of the individual studies identified

Both of the included studies had a primary aim of assessing PSA kinetics as a prognostic marker. D’Amicoet al.129 evaluated whether the rate of rise in the PSA level (i.e. PSAV) during the year before diagnosis could predict PSA recurrence, prostate cancer mortality and all-cause mortality. Senguptaet al.130 also used three separate end points for different analyses: PSA recurrence, clinical recurrence and prostate cancer mortality. In both studies two models are presented for each end point, the first using only clinical variables and the second including pathological variables. Senguptaet al.130 assessed preoperative PSADT as a predictor of outcome following RP.

Quality of the individual studies identified

Both studies are large and of good quality. However, they both determined the cut-point for differentiating between high and low PSAV within their respective data sets. The same applies to the doubling time (18 months) used by Senguptaet al.130 This means that the results are likely to be over-optimistic as the PSAV and PSADT variables have been optimised to the data. The overall concluding questions to each of the six subheadings are presented inTable 51.

Summary of the baseline characteristics of the sample

The two studies both had over 1000 participants, with almost all (> 95%) having clinically organ-confined tumours. In the largest study Senguptaet al.130 evaluated 2290 men who were treated with RP for prostate cancer between 1990 and 1999, with multiple preoperative PSA measurements available. In the study by D’Amicoet al.129 patients were also treated by RP (Table 52).

The distributions of Gleason and PSA scores (where reported) were similar across studies. Although different cut-points were used in the two studies for PSAV, the proportions in the high-velocity groups were similar at 20.1% and 23.9% respectively. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual studies identified

Table 53 presents a summary of the main statistical findings from the two studies included in this section.

Both studies report a Cox multivariate analysis of the data.Table 53 shows the different clinical and pathological classical markers entered into the statistical models across the studies, together with the results of each analysis.

Senguptaet al.130 calculated PSADT by log-linear regression and PSAV by linear regression. Each of these parameters was used in preoperative and postoperative multivariate models for the end points of biochemical and clinical progression, and cancer death, but only one remained in each model. PSAV appeared to be a better predictor of biochemical progression, and PSADT of clinical progression and death. Of all the predicted outcomes the association with cancer death appeared to be the strongest. In the clinical model the HR for death from prostate cancer was 6.18 (95% CI 2.75–13.88,p < 0.0001) in men with a PSADT of less than 18 months versus men with a PSADT of 18 months or more; similarly, the HR was 3.92 (95% CI 1.95–7.85,p = 0.0001) in the pathological model.

D’Amicoet al.129 also reports a particularly strong association between PSAV and prostate cancer death in both clinical and pathological models. In the clinical model the HR for death from prostate cancer was 9.8 (95% CI 2.8–34.3,p < 0.001) in men with an annual PSAV of more than 2 ng/ml versus an annual PSAV of 2 ng/ml or less; similarly, the HR was 12.8 (95% CI 3.7–43.7,p < 0.001) in the pathological model.

Overall conclusions based on the results and quality of the findings

Both of these large, good-quality studies report compelling results showing an association between PSA kinetics and prostate cancer outcomes, and in particular cause-specific mortality. This result remained significant in the presence of other clinical and pathological variables. However, with both studies using data-dependent cut-points to define high and low PSAV the results will be over-optimistic. Whereas D’Amicoet al.129 derived an optimum cut-point of 2.0 ng/ml/year, Senguptaet al.130 found 3.4 ng/ml/year gave the best results. Use of the other cut-points in the two data sets would give more realistic estimates of how this prognostic marker would perform in practice. A review of monitoring protocols for men with localised prostate cancer143 showed that in some research protocols PSAV and PSADT were already used, in conjunction with other factors, to identify disease progression that might require radical treatment. Note that in the UK regular measurements of PSA are not routinely available before diagnosis as was the case in these two studies, as regular PSA screening is not normal practice.

Senguptaet al.130 concluded that, although PSADT may perform more accurately and strongly in multivariate analysis than PSAV, PSAV is simpler to derive and therefore more easily used in clinical practice.

Brief description of the prognostic marker

Signal transducer and activator of transcription-5 (Stat5) is a signalling protein that is activated by prolactin in normal and malignant prostates. The definition of the marker and its distribution in the population studied are shown inTable 54.

Brief description of the objectives of the individual study identified

The study aimed to investigate whether activation of Stat5 in prostate cancer was linked to clinical outcome with disease recurrence as an end point. The basic study design characteristics are summarised inTable 55.

Quality of the individual study identified

In general this was a good quality study. Unusually it was very specific as to the events that were included as the end points, but the number of events was not stated and so the EPV is unknown. In interpreting the results the omission of PSA from the multivariate analysis must be considered. As with many prognostic studies in this systematic review the study did not provide details about the storage of materials, although it was clear that the study was based on archival specimens. The overall concluding questions to each of the six subheadings are presented inTable 56.

Summary of the baseline characteristics of the sample

The study involved 357 participants who had been treated with RP or TURP. At pathological stage there were still a greater number of organ-confined (79.5%) than non-organ-confined participants (19.7%), with a small number of participants having missing data (0.7%). The Gleason scores ranged between 2 and 5 but PSA levels were not reported. The failure to measure and report this information limits the ability to compare this study with other prognostic studies involving other types of markers. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual study identified

Liet al.131 provided a multivariate analysis of the data. Non-significant factors were removed from the multivariate model. The end point was progression-free survival, with clinical recurrence, PSA recurrence and prostate cancer deaths all treated as events. The HRs andp-values are shown for the univariate analyses and for the variables kept in the multivariate model. Univariate analysis showed that Stat5 activation was associated with early disease recurrence (p = 0.04). However, in multivariate analysis Stat5 activation status only reached borderline significance in its association with progression-free survival (HR 1.63; 95% CI 0.99–2.69;p = 0.057) in a model that included Gleason grade and stage but not PSA. The effect size (HR = 1.6) was similar to that for grade (HR = 2.0) and stage (HR = 2.0). A subgroup analysis of patients with intermediate Gleason grade prostate cancers (3 and 4; 325 of the total patient sample of 357) showed similar results.Table 57 presents a summary of the main statistical findings from this study.

Overall conclusions based on the results and quality of the findings

Although the current study was found to be adequate in terms of key quality factors considered to be important when evaluating prognostic studies, there were shortcomings that make the result inconclusive: the absence of PSA from the analysis and the uncertain (possibly inadequate) number of EPV needed to give a statistically reliable result. To establish whether Stat5 really adds prognostic value to the established markers it needs to be tested in a study that addresses these issues. The authors claim that the predictive value of active Stat5 in prostate cancers of intermediate and low histological grades might be improved by an analysis of other prognostic markers in conjunction with active Stat5 (e.g. Ki67, p53, Bcl-2, syndecan-1125). This hypothesis needs to be tested.

Brief description of the prognostic marker

Two principal approaches have been used to estimate tumour size: tumour volume and maximum tumour dimension. The estimate used in each study together with the measurement methods and values are shown inTable 58.

It is not clear whether any of the measures are the same, but the values for tumour volume reported by Lieberet al.106 and Salomonet al.132 appear consistent with each other. Note that the measure of tumour dimension used by Viset al.124 is clearly different to those used by Bluteet al.105 and Senguptaet al.,130 being from biopsy cores rather than from the pathological specimen.

Brief description of the objectives of the individual studies identified

Only one of the studies had a primary objective of assessing the prognostic significance of tumour size.132 Salomonet al.132 aimed to evaluate the association between Gleason score, stage and status of surgical margins and tumour volume in prostate cancer progression after RP. Three studies had the objective of investigating other novel markers,106,124,130 and one developed a prognostic model.105

Quality of the individual studies identified

The overall concluding questions to each of the six subheadings are presented inTable 59.

The principal weakness present in all of these studies is that the classical markers were not present or kept in all analyses and so the additional prognostic value of tumour size in the presence of known markers is not clear. In particular, several analyses omitted PSA, a classical marker that may be associated with tumour volume. The only study that had the assessment of tumour size as its main objective132 did not use a time to failure analysis (Cox regression) and so the statistical analysis is weak.

Summary of the baseline characteristics of the sample

The five studies included a wide range of samples sizes, from 281124 to 2290.130 All five studies were based on patients who had received RP treatment (Table 60).

In evaluating the results of the five studies it is important to consider the differences in sample characteristics (e.g. stage, Gleason score and PSA distributions). The clinical stage of the participants was provided in four of the five studies (not that of Lieberet al.106). More than 90% of the samples in the four studies were made up of organ-confined participants at clinical stage. Lieberet al.106 had 18% of patients who were found pathologically to have positive regional lymph nodes, which is high compared with the other studies in this group. The distributions of Gleason and PSA scores (where reported) were similar across studies. Additional summary characteristics are provided inAppendix 7.

Brief description of the results from the individual studies identified

Tables 61 and62 present a summary of the main statistical findings from the five studies included in this section.

Maximum tumour dimension

Two studies105,130 report analyses of maximum tumour dimension with PSA recurrence, clinical recurrence and prostate cancer death all used as outcomes in different analyses. In both studies maximum tumour dimension was found to be significant in univariate analysis but not in multivariate analysis. With biochemical progression as the outcome, Viset al.124 found length of tumour in biopsy cores significant in univariate and multivariate analysis (p = 0.04), but the multivariate analysis included only one of the classical markers, PSA. With the outcome of clinical progression, length of tumour in biopsy cores was not significant in univariate or multivariate analysis.

Tumour volume

Four studies106,124,130,132 report several analyses of this marker with different end points: PSA recurrence, clinical recurrence, prostate cancer death and all deaths. In univariate analyses, except that with all deaths as the outcome,106 tumour volume was reported to be significant. In multivariate analysis it was not found to be significant in the studies of Lieberet al.106, Salomonet al.132 or Viset al.124 Senguptaet al.130 did not find it to be significant in an analysis with biochemical recurrence as the end point but did find it to be a significant predictor of clinical progression (p = 0.0008) and prostate cancer death (p = 0.003). It may be of note that PSA and stage were included in the first analysis but were not in the last two analyses (i.e. tumour volume was only significant in the absence of PSA and stage in the model). The association between tumour volume and PSA may account for the results of Senguptaet al.130

Overall conclusions based on the results and quality of the findings

All of these studies have weaknesses that make their individual results inconclusive with respect to the significance of tumour size as a prognostic indicator; however, the direction of evidence suggests that maximum tumour dimension, length of tumour in the biopsy core and tumour volume are not independent prognostic parameters after other routinely assessed variables are accounted for. Tumour volume was only found to be significant in multivariate models that did not include PSA or stage.130

Study populations

All of the studies made clear statements about the patients included and the dates that marked the start and finish of patient recruitment, with the exception of Vollmeret al.107 These were the principal criteria for the quality assessment. Only two reported on the setting, one reported zero time (Lieberet al.106) and none mentioned diagnostic methods.

Specification of the principal treatment was a condition for inclusion in the review. All models applied to patients treated with RP except that of Cowenet al.,150 in which patients had a mixture of prostatectomy, radiotherapy and ‘other treatment’, the last being principally watchful waiting. Two studies did not specify if any patients had had adjuvant or neoadjuvant treatment,107,150 and Hanet al.140 excluded such patients from their analysis. The patient cohort of Bluteet al.105 comprised 15% who had had adjuvant therapy, a group that they considered excluding ‘but thought it would have resulted in a lower risk cohort that would not be reflective of our practice’. Instead they included adjuvant therapy as a covariate in their models. A total of 17% of the patients in the Lieberet al.106 cohort had had ‘early endocrine therapy’, but as this factor was not statistically significant it was not included in the final model.

The studies in general gave good descriptions of the key characteristics, as demonstrated inTables 66,67 and68, which show the study populations by stage, Gleason grade and PSA respectively. As far as it is possible to tell from the different statistics reported for these factors it appears that the study populations are broadly similar.

The stage distribution of the Lieberet al.106 study population is not comparable with that of the other studies as only pathological stage was reported. Many patients have their tumours upstaged on surgery. Of the studies that reported pathological stage as well as clinical stage, Hanet al.140 reported that 50% of study patients had pathologically non-organ-confined tumours and 5% had positive lymph nodes; Vollmeret al.107 and Bluteet al.105 reported 43% and 13% extracapsular tumours respectively. This demonstrates the differences that may be found between clinical and pathological staging, but there also appear to be differences in the accuracy of clinical staging, although study exclusion criteria (for example Bluteet al. excluded patients with pathologically positive lymph nodes) may be the reason for this.

The Gleason distributions of Cowenet al.150 and Hanet al.140 are not strictly comparable with those of the other studies as many patients’ Gleason scores are upgraded when pathological specimens are available. This may explain the relatively high proportion of patients with low-grade cancers in the Cowen study. Low Gleason scores (2–4) are usually no longer assigned to biopsy specimens, which may explain their absence in the study of Hanet al. Of the studies that report pathological Gleason scores the populations appear similar on this factor.

The Lieberet al.106 study is based on a pre-PSA era cohort of patients, and the very high proportion of missing PSA values in the Cowenet al.150 study may be for the same reason. The distributions in the other studies appear comparable, with the median PSA in the 4.1–10 ng/ml range.

Study attrition

Study attrition included both the omission of patients because of the lack of baseline variables and loss to follow-up. Although most studies stated the total population from which the study sample was drawn, together with reasons for exclusions, none reported the extent of loss to follow-up. However, Lieberet al.106 showed the number at risk for the three different outcome measures used in their models for all three factors in the models at 10 years. Two studies105,140 reported how loss to follow-up was dealt with in the analyses. None discussed the biases that may have been introduced from the loss of patients from the analyses, although Cowenet al.150 did discuss the potential effect of a high proportion of missing PSA data on their results.

Prognostic factor measurement

Most studies gave some information regarding the measurement of some of the prognostic markers used. Both of the studies that included the novel ploidy marker described its measurement;105,106 however, only two studies reported the PSA assay that was used,140,150 although there are several. Material storage was only described in two studies, i.e. those in which ploidy was measured105,106 There was no evidence of data-dependent cut-points being used for any continuous variables in the studies, but in two of the five studies continuous variables were categorised106,140 and in a further study it was not clear what was done.107

Outcome measurement

The end points used in the studies, together with some of their properties, are shown inTable 69. Four different end points for the outcome measurement (all deaths, prostate cancer deaths, clinical recurrence and biochemical recurrence) were used in the eight models. Of these, only all-cause death was unambiguously defined.106,150 Lieberet al.106 and Vollmeret al.107 report models with prostate cancer death as the end point, but they do not report how attribution of cause of death was made. The Lieber study also uses clinical recurrence as a model end point, but, although reporting tests that were given to patients to establish recurrence, the frequency of follow-up is not stated. This outcome is now used more rarely and has generally been superseded by PSA recurrence, which was used by Hanet al.140 and Bluteet al.105 Both used a unique definition of PSA recurrence, but only the study of Hanet al. used the consensus definition of 0.2 ng/ml. In none of the three studies in which recurrence was an outcome105,106,140 was it clear whether deaths were treated as events or censored.

Confounding measurement

Confounding measurement, considered principally as the inclusion of the classical markers in the models, was also dealt with poorly in the studies. Only two models included all confounders in their analysis,140,150 and in one instance this was not a deliberate choice but the result of all of the established markers remaining significant in the stepwise variable selection process.140 In the Cowenet al.150 study all potential covariates were kept in the model but most patients had missing data on a key confounding variable, PSA, and so the study could not be awarded a ‘yes’ for this category. None of the other studies forced known confounders into their analysis, although omitting them can result in a misleading model. The inclusion of the classical markers in the prognostic models is shown inTable 70. Note that the inclusion of other factors is also relevant in particular circumstances, such as age for an end point of all-cause mortality and treatment when this varied (seeTable 61).

Statistical analysis

All of the models included in the review were developed using a multivariate Cox proportional hazards regression. None of the studies reports testing the proportionality assumption, although Hanet al.140 tried parametric (Weibull, lognormal and gamma) Cox models. They selected the proportional hazards model on the basis of a comparison of actual and predicted survival curves (calibration) for four risk groups.

All of the models used were considered to be methodologically adequate and all had at least 10 EPV in the multivariate model.

In general the statistical methods used were well reported, although presentation of the univariate results was not universal. Univariate analysis was reported to have been carried out in three studies,105,106,140 was presented in two,105,106 but was only used in one140 to select variables to enter into the multivariate model. There was further heterogeneity in the methods used to select variables for the final models presented. Three studies106,107,140 appear to have used a stepwise process, either forwards106 or backwards.140 The method used by Vollmeret al.107 was not specified. Cowenet al.150 state that the variables for their model were chosen on a ‘conceptual basis’. Bluteet al.105 start with ‘established predictors’ in their model and then add and remove variables to determine the effect on the predictive power of the model, as judged by theC-statistic. When model predictive power was similar despite the inclusion or exclusion of variables, these variables were removed from the model. These variable selection processes, as well as the lack of availability of data, resulted in well-established markers [Gleason score, PSA, stage (or organ-confined status) and surgical margins (when relevant)] being omitted from all but two of the eight final models, as discussed above.

Study heterogeneity

The heterogeneity between studies precluded the use of meta-analysis. One of the main sources of heterogeneity was in the measures of outcome, with all-cause mortality, prostate cancer mortality, and clinical and biochemical recurrence all being used, with the definition of the last two also varying. Other important differences between studies were the covariates included in the multivariate analyses and the marker measurement methods and cut-points used to define prognostic groups. As well as the heterogeneity in study design and analysis methods, the poor reporting of models and particularly the lack of HRs sometimes made meta-analysis impossible. Methods are available to estimate HRs from other results presented, but this would have been possible in a limited number of cases and would not have affected the possibility of undertaking meta-analysis because of the other sources of heterogeneity. Similarly, if more articles had been included in this review it is very unlikely to have affected the ability to have undertaken meta-analyses.

The heterogeneity between studies, poor quality of studies and the limited number of studies for each marker also mean that the classification of markers into ‘promising’ and ‘not promising’ groups can be considered indicative only, based on the generally weak evidence available. Other reviews of prognostic markers and models, not only in cancer, have also commented on the generally poor quality of studies in this field92,97,99,100 and the issues have been more generally discussed in the literature.96,155,156

There is increasing interest in meta-analysis using pooled individual patient data from different studies.156–158 This method allows differences in statistical models, and particularly differences in the treatment of covariates and marker cut-points in reported studies, to be standardised in a single analysis (assuming covariate data are available) and reduces the potential for misleading results.158 However, not all differences between studies can be retrospectively overcome through uniform analysis. Some of these differences are common to all prognostic marker studies, such as the different (or unspecified) definitions and measurement methods of novel markers. For prostate cancer studies a particular issue is the variation in definition of PSA failure, as failure may result in different patient treatment and so different failure thresholds cannot be applied retrospectively.

Publication and reporting bias

There was only a small number of studies, or sometimes only a single study, for each marker. It was not possible to examine the potential issues of publication bias or selective outcome reporting. The exclusion of smaller studies may have reduced the possibility of publication bias, but with the literature comprising retrospective case series the possibility of publication bias remains considerable. Furthermore, with several possible outcome measures available there is scope for selective outcome reporting. Kyzaset al.159 evaluated publication bias and variation in outcome definitions in the literature on prognostic factors for head and neck squamous cell cancer. Their analysis showed that these biases may inflate the apparent importance of prognostic markers. This must be considered in the interpretation of the results of this review. It is possible for many markers that a single unpublished study could have altered the conclusions considerably.

Prognostic or predictive marker?

In none of the novel marker studies was it considered whether a marker was prognostic or predictive. Given that in the majority of studies patients all had the same principal treatment this was not possible to assess. Before a marker is adopted it needs to be considered whether it is truly prognostic or whether it may be predictive, i.e. whether there is an interaction with any particular treatment.

Economic evaluation

This study did not include an economic evaluation of the use of novel markers. The clinical and financial consequences of the use of prognostic markers will be known only if research is carried out to show which prognostic groups are likely to benefit from radical treatment. Currently most men who are otherwise healthy have radical treatment. The consequences of introducing a novel prognostic marker will depend on whether some men opt not to have such treatment as a result of the test and how their disease subsequently progresses compared with men of the same prognostic status who do have treatment. The advantage of immediate radical treatment compared with active monitoring is not yet fully understood for the prognostic groups defined by the classical markers in current use.