Statistical superiority sometimes makes little clinical sense

Abstract

A study by Rumaihi et al. [1] reported in this issue set out to investigate how an abnormal PSA level and DRE findings were used to detect prostate cancer in a group of 651 men referred to their centre and who had a 12-core TRUS-guided biopsy of the prostate (TRUSBP), based on an abnormal PSA level (>4 ng/mL) or an abnormal DRE. The biopsy findings were cancer in 27.6%, BPH in 42.2% and prostatitis in 36.4%. The investigators then set out to study the capability of total PSA level, a DRE and combinations of both for detecting prostate cancer, and calculating the sensitivity, specificity, positive predictive value, negative predictive value and overall accuracy for each measurement. They also analysed the receiver operator characteristic curves of different PSA levels, finding that a value of 7.9 ng/mL should prompt a biopsy, rather than the ‘normal’ level of 4 ng/mL, and gave the ‘best results’ from a statistical perspective view. Indeed, using the higher rather than the lower value to justify TRUSBP resulted in almost 50% accuracy compared to 32.2% with a PSA level of 4 ng/mL. Combining an abnormal DRE with the two PSA thresholds gave an accuracy of 63.4% for a PSA level of 4 ng/mL, which could be improved to 76.9% using the higher value. The higher threshold was initially accepted as statistically superior; consideration that it might be related to higher levels of prostatitis found in men referred to them was discussed and then they became aware of failing to detect prostate cancers by using the higher PSA value to prompt TRUSBP. Certainly, 78 such cases (43% of all their detected cancers) would have been missed by using a threshold PSA level of 7.9 ng/mL to justify TRUSBP! When these cases were analysed by Gleason score, it was found that 27% had scores of ⩾7. Thus, despite the statistical superiority of using the higher PSA value to initiate TRUSBP, the clinical effect made little sense. Why would you want to not detect 43% of the cancers, particularly when a quarter of those missed were aggressive cases? The authors concluded appropriately that their ‘study supports the need for other diagnostic tests to help in solving this dilemma’. This has always been the case with the use of PSA and PSA derivatives for detecting prostate cancer.