PSA is often seen as a diagnostic test for prostate cancer in the same way that, say, a throat culture is diagnostic for streptococcal pharyngitis: a culture could be taken every day and would show nothing, until the patient suddenly develops an infection and the test turns positive. Anecdotally, patients with a PSA of 3.8 ng / ml, only a fraction below a typical biopsy threshold, are commonly told that their test is “normal” or “negative” and to come back in a few years’ time. Exactly the same advice is given to men with a PSA of 0.4 ng / ml. It seems reasonable to suppose that a man with a PSA of 3.8 ng / ml is at greater risk for a future diagnosis of prostate cancer than a man with a PSA of 0.4 ng / ml. There is solid empirical evidence that this is indeed the case.
Four classic papers from the mid-1990’s clearly demonstrated that PSA is prognostic as well as diagnostic for prostate cancer. Each of these papers used a nested, case-control design, in which archived bloods taken before the discovery of PSA were retrieved from storage and subject to PSA assay. The pioneering study, albeit small, with 18 cases and 36 controls, was that of Carter et al., who used data from the Baltimore Longitudinal Study of Aging to show that PSA was associated with prostate cancer risk up to 25 years before diagnosis[4
]. Carter et al. focused on changes in PSA but did not directly compare the predictive value of PSA velocity or doubling time with absolute PSA level. Such an analysis was conducted by Whittemore et al., who used blood samples taken in 1964 – 1971 as part of a Kaiser Permanente screening study to predict cancers diagnosed to the end of 1987, before PSA testing became available in the clinical practice. Absolute level of PSA was a very strong predictor of subsequent prostate cancer diagnosed within 7 years with an area-under-the-curve (AUC) of 0.92 and outperformed PSA velocity. Of critical importance to the interpretation of the Malmö studies (see below), which included almost exclusively Caucasian patients, is that “there were no important or statistically significant differences in the performance of [PSA] by race”[5
], suggesting that data from European populations can be applied to African Americans. The association between PSA and subsequent prostate cancer was then replicated by Stenman et al., using a Finnish cohort with similarly long follow-up: blood samples taken in 1968 – 1973, with follow-up to 1980. The authors reported high sensitivity and specificity of PSA for the detection of cancer within 5 years for men aged <65[6
]. The fourth early paper reporting long-term prognostic value to PSA was based on the Physicians’ Health Study, with 366 cases diagnosed within 10 years of the initial blood sample. Using a 4 ng/ml cut-off, specificity was 91% and sensitivity 43%, although the latter increased markedly for more aggressive cancers diagnosed within 4 years (sensitivity of 87%)[7
The four papers described above clearly demonstrate proof-of-principle for PSA as a prognostic factor for prostate cancer. However, it is highly questionable whether they are of practical value for risk stratification. First, the cohorts were heterogeneous with respect to age. The Physicians’ Health Study, for example, included men aged 40 – 84; the Kaiser Permanente cohort included those aged 44 – 81. This is problematic not only because PSA changes with age – making cut-offs derived from these studies somewhat meaningless – but because decisions about screening are highly age dependent. For example, we might consider telling a 70 year-old man with a low PSA that further screening would not be of benefit; it is unlikely that a 40 year-old could be similarly reassured that screening could be avoided. Second, the follow-up in some studies was relatively short, such as 10 years in the Physicians’ health study and 12 in the Finnish cohort. Given that prostate cancer is detectable by PSA several years before clinical diagnosis, very long-term follow-up is required in order to make decisions about screening. Third, measuring PSA in stored blood is not straightforward[8
] and questions can be raised about the assays used. Fourth, several studies had relatively few cases, such as 18 in the Baltimore cohort and 44 in Finland. Fifth, in many of the studies, it is not fully clear the degree to which the study cohort is representative of the target population. Finally, the authors of the studies did not choose to present clinically relevant statistics. Sensitivity and specificity have their uses, but cannot help determine which men should be screened and which should be advised against further screening.