A (-5, -7) proPSA based artificial neural network to detect prostate cancer

OBJECTIVE: The pro-forms of prostate specific antigen (-2,-5,-7 proPSA) and also %free PSA based artificial neural networks (ANN) have been suggested to enhance the discrimination between prostate cancer (PCa) and no evidence of malignancy (NEM). This study reports on the combined use of proPSA within a %free PSA based ANN to enhance specificity of PCa. METHODS: Serum samples from 898 patients with PCa (n=384) or NEM (n=514) within the PSA range 1-10 microg/l were analyzed for PSA, free PSA and (-5,-7) proPSA (Roche assays). Patient data from two centers - taken first from the Swiss site of the ERSPC (Aarau) and from a referral population (Berlin) have been analyzed. Leave-one-out ANN models with the variables PSA, %fPSA, proPSA, prostate volume and status of digital rectal examination (DRE) were constructed and compared by receiver-operating characteristic (ROC) curve analysis. RESULTS: (-5,-7) proPSA was only significantly different between NEM and PCa in the PSA range 4-10 microg/l. Within the PSA range 4-10 microg/l (Berlin group) the ANN including only the two variables %fPSA and proPSA could reach the same performance like the conventional ANN with PSA, %fPSA, age, prostate volume and DRE (both AUCs: 0.84) However, at 95% sensitivity all ANN could not improve specificity compared to %fPSA. CONCLUSIONS: ProPSA as single parameter did not improve specificity over %fPSA whereas proPSA and %fPSA within an ANN in the PSA range 4-10 microg/l substituted prostate volume and DRE. At 95% sensitivity only ANN with prostate volume and DRE perform significantly better than %fPSA.