Clinical utility of human glandular kallikrein 2 within a neural network for prostate cancer detection

OBJECTIVE

To assess, using artificial neural networks (ANNs), human glandular kallikrein 2 (hK2), prostate‐specific antigen (PSA), and percentage free/total PSA (f/tPSA), for discriminating between prostate cancer and benign prostatic hyperplasia (BPH).

MATERIAL AND METHODS

Serum samples from 475 patients with prostate cancer (n = 347) or BPH (n = 128) within the PSA range of 1–20 ng/mL were analysed for tPSA, fPSA and hK2 (research assay, Toronto, Canada). Data were analysed in the ranges of 1–4, 2–4, 4–10, and 2–20 ng/mL tPSA. Back‐propagation ANN models with the variables PSA, f/tPSA, and hK2, hK2/fPSA and hK2/(f/tPSA) were constructed. The diagnostic validity was evaluated by receiver‐operating characteristic (ROC) curve analysis.

RESULTS

Whereas the median concentration of hK2 was not significantly different between patients with BPH or prostate cancer in any of the tPSA ranges, the f/tPSA, hK2/fPSA and hK2/(f/tPSA), and the hK2‐based ANN outputs were always significantly different between patients with prostate cancer or BPH. Using ROC curve comparison, all variables were significantly better than hK2 in all ranges. The hK2‐based ANN performed better than f/tPSA except in the 4–10 ng/mL tPSA range. At 90% and 95% sensitivity, the hK2‐based ANN was also significantly better than f/tPSA in the 1–4 ng/mL tPSA range. hK2/(f/tPSA) achieved equal results to the hK2‐based ANN except in the range 2–20 ng/mL tPSA.

CONCLUSIONS

The hK2‐based ANN improves the outcome of f/tPSA but not hK2/(f/tPSA) in almost all analysed subgroups. When comparing the results at 90% and 95% sensitivity the hK2‐based ANN only performed significantly better than f/tPSA in the lowest tPSA range. Only in lower tPSA ranges do hK2‐based ANNs show an advantage for further improving prostate cancer detection.