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

Area of Science:

• Urology
• Oncology
• Biomarker Research

Background:

• Prostate-specific antigen (PSA) and its derivatives, like proPSA, are crucial in prostate cancer (PCa) diagnosis.
• Artificial neural networks (ANNs) incorporating %free PSA have shown promise in differentiating PCa from no evidence of malignancy (NEM).
• Enhancing diagnostic specificity for PCa remains a clinical challenge, particularly in the PSA grey zone (1-10 ng/mL).

Purpose of the Study:

• To evaluate the combined utility of proPSA and %free PSA within an ANN model for improving PCa specificity.
• To compare the performance of this combined ANN model against conventional diagnostic variables.
• To assess the diagnostic accuracy in the PSA range of 1-10 ng/mL.

Main Methods:

• Serum samples from 898 patients (384 PCa, 514 NEM) within the 1-10 ng/mL PSA range were analyzed.
• Measurements included PSA, free PSA, and (-5,-7) proPSA using Roche assays.
• Leave-one-out ANN models were constructed using variables like PSA, %fPSA, proPSA, prostate volume, and digital rectal examination (DRE) status, compared via ROC analysis.

Main Results:

• (-5,-7) proPSA showed significant differentiation between NEM and PCa within the 4-10 ng/mL PSA range.
• In the 4-10 ng/mL range, an ANN using only %fPSA and proPSA achieved performance comparable to conventional ANNs (AUC: 0.84).
• However, at 95% sensitivity, no ANN model significantly improved specificity compared to %fPSA alone.

Conclusions:

• ProPSA alone did not enhance specificity over %fPSA.
• In the 4-10 ng/mL PSA range, proPSA and %fPSA within an ANN effectively substituted for prostate volume and DRE.
• At 95% sensitivity, ANNs incorporating prostate volume and DRE demonstrated superior performance compared to %fPSA alone.