Prostate cancer screening: what can we learn from randomised trials?

In this article, the principle of randomised trials are first described and then prostate cancer screening trials published to date are evaluated based on these principles. A summary of the randomised prostate cancer screening is provided. The conclusion that can be made from the results of the screening trials, as well as limitations of the evidence and open questions are outlined in the end.

Introduction

Modern medicine requires that management decisions are based on solid science. The question that we seek to answer is whether an intervention can produce the defined goals. In the spirit of evidence-based medicine and effectiveness or outcomes research, claims of effectiveness must be derived from head-to-head comparisons of therapeutic alternatives, with direct observation in comparable patients and conditions with controlled design. A valid study design allows us to infer what would have happened to this group of patients subjected to an intervention, if they had received an alternative intervention (ceteris paribus).

Principles of randomised trials

A randomised trial is the gold standard for evaluating the effectiveness of any medical intervention (1). Its chief virtue is that it can eliminate bias from treatment assignments. By virtue of randomisation, they can create comparable groups at baseline and avoid any imbalance in the key determinants of the outcomes studied and hence minimise confounding and selection bias.

This requires, however, careful selection of subjects in terms of eligibility and exclusion criteria (2,3) (). Recruitment can already be a hurdle, if potential participants are not willing to join a trial. Population-based effectiveness trials that identify and randomise subjects prior to consent have at least an apparent advantage in this respect. However, low compliance or substantial contamination can abolish that benefit. Informed consent is crucial in a randomised trial, where participants are subjected to an intervention that usually also carries a risk of adverse effects. Also, information on the features defining eligibility should be readily available and up to date. For instance, in a screening trial the participants must be free from the target condition, and ideally recruitment attempts should exclude people with the disease at the outset, based on case lists from comprehensive cancer registries.

Table 1

Key features affecting the relevance and quality of randomised trials

Feature	Effect
Equipoise	Relevant study question, avoid excessive risks
Outcome	Clinically important end-point
Comparator	Relevant study question
Fair randomisation	Avoid selection bias and confounding
Blinding	Avoid information bias
Adequate sample size	Sufficient statistical power
Intention-to-treat analysis	Maintain balanced distribution

In addition, the randomisation procedure must be valid, i.e., based on genuinely unpredictable allocation to guarantee allocation concealment (4,5). If the assignment of a particular subject can be foreseen, allocation can become biased by certain subjects being preferably assigned to a given trial arm, which leads to selection bias (6). The allocation ratio, i.e., proportion of subjects assigned to the trial arms does not need to be equal, but it must be identical across the participants (overall in simple randomisation and within the group in stratified randomisation). The current requirement is that computer-generated random numbers are used, preferably assigned by a separate party not responsible for recruitment and only after registering a participant to the trial so that potential for tampering the random allocation is minimised.

Randomisation can, however, generate comparability only at baseline, at the time of allocation. Selective loss during a study can lead to subsequent distortion of the initial balance (7,8). Use of passive data collection through e.g., registers can avoid this, but maintaining validity also requires that researchers are allowed to compile data on subjects who have withdrawn from the study. This is clearly an ethical issue, as there is a conflict between autonomy and common good that can be obtained from valid knowledge available only from unbiased research.

An adequate sample size is a prerequisite for achieving a balanced distribution of the characteristics between the trial arms through randomisation. Further, realistic power calculations with reasonable expected event rates and allowance for non-compliance are needed to define the target sample size (9).

The research question of a trial should fulfil the requirement for equipoise, i.e. there should be a balance of uncertainty regarding the study hypothesis (10). This means that there is sufficient preliminary evidence to support a hypothesis (proof of principle) and safety data (often from phase 1–2) studies, but also genuine uncertainty about the effects of the experimental intervention. For this purpose, a systematic literature review should be incorporated as part of the proposal. Too early randomized controlled trial (RCT) (‘a long shot’) can lead to low chance of achieving a benefit and a poorly characterised risk profile. Besides the intervention studied, the comparator must also be chosen equally carefully. As a general principle, it should represent normal health care, i.e., the approach that would be provided in the absence a trial. The outcome should be chosen so that it quantifies the real goal of the intervention to the patient, i.e., benefit to the subject and it should not be only an indirect indicator such radiological or biochemical marker of an early stage of the disease process (11).

Blinding (efforts at keeping either participants or investigators unaware of the allocation) can improve comparability of end-point data and hence reduce information bias (11). It is, however, not possible for interventions that cannot be mimicked by sham procedures. Further, blinding is important primarily for outcomes involving a degree of judgment. Subjective outcome measures can be easily influenced by beliefs about interventions and their effects (placebo effect). Therefore, it is important to blind the evaluators of any ‘soft’ end-points. Objective outcomes, on the other hand, are not prone to placebo effects.

In order to maintain the benefits of randomisation, the analysis should follow the intention-to-treat (or screen) principle (12). This means that the groups compared are fully in compliance with random allocation. Only predefined exclusions can be justified and even they only in case that it had later turned out that subjects had been randomised even if they were not eligible. It is crucial that similar criteria are applied consistently across the arms and information is obtained in a similar fashion for the trial arms (to avoid asymmetrical exclusions of ineligible subjects) (13,14). ‘Treatment received’ analysis compares the participants with and without the intervention (either just excluding non-compliers in the intervention arm, or also combining them with ‘contaminators’ in the control arm), but such an approach loses entirely the advantage of randomisation and treats the data as a cohort study. Sometimes even the data analysis is performed in a blinded fashion, before opening the code revealing which arm is which. This can avoid subtle biases in the analysis phase of the study.

A randomised trial usually requires monitoring, frequently conducted by external, independent experts. Trials can be terminated early, if the results show convincing evidence of either benefit or futility of the intervention, or unacceptable risks (15,16). This required for ethical conduct and good research practice, but it can also lead to premature discontinuation of a study, with inconclusive results, if careful consideration is not applied. A pre-specified monitoring plan should be employed and alpha spending considered to avoid type one error. For instance, a Finnish population-based colorectal cancer screening trial was discontinued when the 5-year results showed no benefit (17). The follow-up may have been too short, especially given evidence of benefit from previous trials. An Estonian trial of menopausal hormone treatment was also terminated early, mainly based on results from other trials that were far from identical (18).

Besides quantifying the impact on the primary goals of interventions (providing the benefit to the patient, avoiding adverse effect of disease), they can also provide best estimates of the cost-effectiveness and quality of life impacts.

Randomised trials of prostate cancer screening

Several small prostate cancer screening trials have been carried out, while only the PLCO and ERSPC have sufficient sample size for assessing the effect of PSA screening on prostate cancer mortality (). Some other studies have also been conducted such as Stockhom-3 study (19) and Tyrol study (20), which are not trials, as they involved no randomisation. Also, several case-control studies have been conducted (21-23), but they are not covered here.

Table 2

Prostate cancer mortality in prostate cancer screening trials

Trial	Size of study population (screening + control)	Target age group (years)	Screening tests	Participation (%)	Follow-up (years)	Number of prostate cancer deaths (screening + control)	RR for prostate cancer mortality
Norrköping	1,494+7,532	50–69	DRE, PSA	70–78	20	30+130	1.2 (0.8–1.7)
Stockholm	2,400+24,772	55–70	DRE, TRUS, PSA	74	13	53+506	1.1 (0.8–1.5)
Quebec	31,133+15,353	45–80	PSA, DRE	24	11	153+75	1.0 (0.8–1.3)
PLCO	38,340+38,343	55–74	DRE, PSA	85–89	15	255+244	1.0 (0.9–1.2)
ERSPC	72,891+89,352	55–69	PSA	83	13	355+545	0.8 (0.7–0.9)

RR, risk ratio.

The Norrköping trial (24) was started in 1987. It did not involve real randomisation, but every sixth man was allocated to screening from a list of birth dates. Four screening rounds were used with a 3-year interval. Hence, the control group was 5 times larger than the intervention group (). Initially only DRE was used, but in the two last rounds also PSA (in the last round men older than 69 years, 46% of the screening group, were no longer invited). Participation ranged 70–78% by round. There were 43 screen-detected and 42 interval cases in the screening groups and 292 prostate cancers in the control group (cumulative incidence 5.7% vs. 3.9%). Cumulative prostate cancer mortality at 20 years was 2% (30/1,494) in the screening group and 1.7% (130/7,532) in the control group, RR =1.2, 95% CI: 0.8–1.7 (though the researchers report deaths among prostate cancer cases as their main results). The study was underpowered for assessing a mortality effect.

The Stockholm trial (25) had a very small screening group of 2,400 men. It had only a single screening round in 1988 using DRE, TRUS and PSA. The PSA cut-off for biopsy was 10 µg/L. Of the invited men, 74% participated and 65 cancers were detected (3.6%). Overall, prostate cancer incidence was 4.0 per 100 in the screening group and 5.2 per 1,000 in the control group (). There were 53 deaths from prostate cancer in the screening arm and 506 in the control arm, corresponding to cumulative mortality of 2.2% vs. 2.0%, with RR =1.1, 95% CI: 0.8–1.5.

The Quebec trial was started in 1988 and used PSA as the screening test (cut-off 3 µg/L), at the initial screen also DRE was used (26). Unlike the Swedish studies, it allocated more men to the screening than control arm (2:1). The target age group was very broad, 40–79 years. Screening was offered on annually, but compliance was low (24%) and 7% of the men in the control group sought screening on their own at Laval University. The results demonstrated no mortality reduction between the trial arms with RR =1.0, 95% CI: 0.8–1.3 () (though the investigators presented an analysis of screened vs. non-screened men and claimed that the results showed a screening benefit).

The PLCO trial recruited volunteers aged 55–74 years from 10 centers. Both PSA and DRE were used in the three first annual screens and PSA only (cut-off 4 µg/L) in the last two rounds. The incidence of prostate cancer at 13 years was 108 per 10,000 person-years in the screening arm and 97 in the control arm (RR =1.12, 95% CI: 1.02–1.17) (27). There were 255 deaths from prostate cancer in the screening and 244 in the control arm by 15 years, with an RR =1.04, 95% CI: 0.87–1.24 (28). The study population was large, but prostate cancer mortality was lower than in the US population at large. It had weaknesses in diagnostic evaluation, as only a third of the screen-positive men underwent a prostate biopsy. Contamination was substantial, as 45% of the men in the control arm had been screened within 3 years prior to baseline, and extensive PSA testing continued in the control arm during the intervention phase (29). A recent modelling study suggested that if the conduct of the PLCO trial had been similar to the ERSPC, it would have shown a mortality benefit (and vice versa for ERSPC) (30). Similar conclusions have also been reached by other investigators (31-33).

ERSPC trial is the largest RCT on prostate cancer screening, with eight centres, albeit the French data have not been included in the mortality analyses. A 4-year screening interval was used, with a PSA cut-off of 3 µg/L. Of the men assigned to screening, 83% were screened at least once and there were on average 2.3 tests per man. Nearly 5,000 prostate cancers were detected through 140,000 screening test (detection 3.5%). Cumulative incidence of prostate cancer was 10.2% in the screening arm and 6.2% in the control arm. A 20% reduction in prostate cancer mortality was demonstrated already at nine years, and it has remained similar at 11 and 13 years of follow-up (34-36) (). A recent analysis of cause of death attribution demonstrated that there is practically no impact of any bias in the adjudication (37). The ERSPC trial has also been criticised for treatment imbalance between the arms. Some of the criticisms are however misguided, as the requirement for comparability is that similar men with similar disease should be treated equally across the arms. A shift in disease characteristics should be reflected in treatment distributions and it does not bias the results. Yet, some imbalance has been shown even after stratification for major prognostic factors (stage or risk group) (38). It appears, however, that the difference is small that it could account for only a small fraction of the screening effect. A full analysis of the issue is being conducted within the ERSPC. Substantial differences between the trial centres have emerged, with large mortality reductions in Sweden and the Netherlands, but very little screening impact in Finland (36,39). These remain to be explained in full but may be influenced by a diluting effect of contamination (40,41), as well as larger effect of continued screening past there round or 8 years.

Conclusions

Currently, randomised screening trials, ERSPC above others, have shown that it is possible to reduce prostate cancer mortality.

A randomised trial is a blunt instrument, very much like an epistemic sledgehammer. This means that it can provide a very reliable answer to a single research question (notwithstanding a parallel design with several interventions, or with adequate sample size and pre-specified protocol for a sub-group analysis). Hence it is crucial to design the aim carefully.

A screening trial needs a well-designed intervention and cannot assess the impact of other interventions. Hence, the optimal screening test can be defined by a randomised trial only if one comparing several alternative approaches side by side could be conducted, but it does not appear practicable.

Are there high-risk groups where balance of benefits and harms of PSA-based screening is more favourable? It is often assumed that targeting a high-risk population will increase the benefits of screening. This requires, however, empirical assessment and no material differences have been shown between men with a positive family history and those without it (42,43). The target group with the largest screening benefit can be evaluated within a trial population. Statistical power is, however, often inadequate for small subgroups, such as high-risk groups defined by family history or specific genetic alterations. A polygenic risk score is potential method for stratifying men by probability of aggressive prostate cancer and prostate cancer death. However, empirical evidence remains scarce (44,45).

Any cancer screening trial must have a long follow-up to show a mortality reduction, as prostate cancer deaths occur mainly after 75 years of age, long after the window of opportunity for early detection with curative treatment. Hence, novel approaches will have always been developed by the time the final results are available. In this respect, evidence from the trials is always outdated. This can also apply to the potential effect of introduction of new treatment approaches on the screening effect.

Even if prostate cancer is among the most frequent causes of cancer death, it accounts only for a small percentage of all deaths. This means that it will not be possible to show an effect in overall mortality.

The ERSPC trial has shown that PSA-based screening can reduce prostate cancer mortality. However, the substantial excess detection weighs heavily against the benefit. How could overdiagnosis be avoided? PSA alone does not appear to have sufficient specificity for clinically important cancers and those screening trials that have not shown clearly higher risks of prostate cancer in the screening arm have not shown mortality benefit either. Also within the ERSPC, the centers with substantial mortality reduction have also markedly elevated prostate cancer risk in the screening arm (39). Kallikrein panels such as Prostate Health Index (PHI) and 4Kscore and magnetic resonance imaging (MRI) hold promise for improving this (46,47), but their impact remains to be evaluated in randomised trials. One is on-going in Sweden and another being launched in Finland (48).

More general clinical issues not limited to screening but with substantial impact on screening outcomes include improved prognostic stratification to guide which men should be curatively treated (or how to identify patients at high risk of disease progression), and tailoring therapeutic approaches to maximise benefits and minimise harms by e.g., early endocrine treatment or chemotherapy.