Performance of a neural network in detecting prostate cancer in the prostate-specific antigen reflex range of 2.5 to 4.0 ng/mL.

OBJECTIVES To explore the potential role of a neural network-derived algorithm in enhancing the specificity of prostate cancer detection compared with the determination of prostate-specific antigen (PSA) and free PSA (fPSA) while maintaining a 90% detection rate. Recent information suggests that the incidence of detectable prostate cancer is similar in men whose PSA values range from 2.5 to 4.0 ng/mL and from 4.0 to 10.0 ng/mL. If the PSA threshold triggering a prostate biopsy is lowered to 2.5 ng/mL, approximately 13% of men older than 50 would be added to the patient biopsy pool. METHODS One hundred fifty-one men were enrolled in a prospective, Institutional Review Board-approved protocol to evaluate the incidence of cancer in a population of men who participated in an early-detection program and whose PSA level was between 2.5 and 4.0 ng/mL. All the men underwent biopsy using an 11-core multisite-directed biopsy scheme, and all biopsy specimens were examined by one pathologist. All men had a second blood specimen drawn before the biopsy for a determination of serum PSA, creatinine kinase, prostatic acid phosphatase, and fPSA. A new neural network algorithm was developed with PSA, creatinine kinase, prostatic acid phosphatase, fPSA, and age as input variables to produce a single-valued prostate cancer detection index (PCD-I). This new algorithm was then prospectively tested in the 151 men. Performance parameters (including sensitivity, specificity, positive and negative predictive values, and biopsies saved) were calculated, and a comparative analysis was performed to evaluate the differences among the new algorithm, percent fPSA, PSA density, and PSA density-transition zone. RESULTS Cancer was histologically confirmed in 24.5% (37 of 151) of the men. The median age of the men was 62 years (range 43 to 74). At a sensitivity of 92%, the specificity for percent fPSA was 11%. The new algorithm (PCD-I) demonstrated an additional enhancement of specificity to 62% at 92% sensitivity. Clinically, the PCD-I would result in a savings of 49% (74 of 151) of all biopsies or 63.6% (71 of 114) of all unnecessary biopsies. CONCLUSIONS A new generation algorithm, derived from a neural network (PCD-I) incorporating the parameters of age, creatinine kinase, PSA, prostatic acid phosphatase, and fPSA can significantly enhance the specificity and reduce the number of biopsies while maintaining a 92% sensitivity rate.