Objectives: To show discriminative power between patients with prostate cancer (PCa) and those with "no evidence of malignancy" using "benign" prostate-specific antigen (bPSA) and the new automated Access benign prostatic hyperplasia-associated (BPHA) research assay within a percent free PSA (%fPSA)-based artificial neural network (ANN) model.
Methods: The sera from 287 patients with PCa and 254 patients with no evidence of malignancy were measured using the BPHA, total PSA (tPSA), and fPSA assays with Access immunoassay technology, with a 0-10 ng/mL tPSA range. Two ANN models with Bayesian regularization and leave-one-out validation using the 4 input parameters of tPSA, %fPSA, age, and prostate volume and 1 containing BPHA/tPSA were constructed and compared by receiver operating characteristic curve analysis.
Results: The BPHA/tPSA-based ANN reached the significant greatest area under the receiver operating characteristic curve (AUC 0.81; P = .0004 and P = .0024) and best specificity (53.9% and 44.5%) compared with the ANN without BPHA/tPSA (AUC 0.77; specificity 50% and 40.6%) and %fPSA (AUC 0.77; specificity 40.9% and 27.2%) at 90% and 95% sensitivity, respectively. The AUCs for tPSA (0.58), BPHA (0.55), BPHA/fPSA (0.51), prostate volume (0.69), and BPHA/tPSA (0.69) were significantly lower.
Conclusions: Although BPHA as single marker or ratio to tPSA did not improve the diagnostic performance of %fPSA or tPSA, the incorporation of BPHA/tPSA into an ANN model increased the specificity compared with %fPSA by 13% and 17% at 90% and 95% sensitivity, respectively. Thus, the automated BPHA research assay might improve PCa detection when incorporating this new marker into an ANN.