Statistical Considerations for Patient Selection and Triggers for Intervention in Active Surveillance

Abstract

The decision whether to refer a prostate cancer patient for curative therapy, or consider active surveillance, depends on the answer to the question: if I do not treat this patient right now, what is the chance that, next time I evaluate him, his cancer will have progressed to become incurable? This is pretty much exclusively a statistical prediction problem. In the case of active surveillance, the endpoint for prediction research is highly problematic. From a research standpoint, the ideal design would be to follow patients without treatment and then see which patients die from disease. Yet to do so would obviate the rationale for active surveillance, which is to follow patients carefully and to cross them to active treatment at the first sign of disease progression. Accordingly, it is disease progression that is typically the endpoint in active surveillance studies. But not only is progression a surrogate endpoint, but predictors for progression can also become criteria for progression, leading to circular reasoning. The best example is PSA velocity: if patients are defined as progressing if their PSA rises above 10 ng/ml, then patients with a greater PSA velocity at baseline are more likely to progress. It would be unsound on this basis to suggest that men with high PSA velocity are not eligible for active surveillance. Criteria for active surveillance can be based on long-term cohort studies of prostate cancer patients followed to death. In particular, it need not matter whether these patients receive treatment (such as radical prostatectomy) or are managed conservatively (even if this is without an active surveillance approach). This is because risk factors are unlikely to change dramatically depending on treatment. Indeed, the results from such studies are consistent. Grade is strongly associated with the risk of prostate cancer death, stage somewhat less so; age moderates this association due to competing risks of death; PSA is only predictive if it is very high; and PSA kinetics have little or no important predictive contribution. That said, it is clear that grade, stage, and age are less than perfect predictors of prostate cancer outcome. Future research on active surveillance should follow some simple principles: separate the predictor and the endpoint, choose clear decision points, specificity may be more of a problem than sensitivity, collaborate and share data, and focus on clinical rather than statistical significance.