Comparison of two different artificial neural networks for prostate biopsy indication in two different patient populations

Abstract

Objectives: Different artificial neural networks (ANNs) using total prostate-specific antigen (PSA) and percentage of free PSA (%fPSA) have been introduced to enhance the specificity of prostate cancer detection. The applicability of independently trained ANN and logistic regression (LR) models to different populations regarding the composition (screening versus referred) and different PSA assays has not yet been tested.

Methods: Two ANN and LR models using PSA (range 4 to 10 ng/mL), %fPSA, prostate volume, digital rectal examination findings, and patient age were tested. A multilayer perceptron network (MLP) was trained on 656 screening participants (Prostatus PSA assay) and another ANN (Immulite-based ANN [iANN]) was constructed on 606 multicentric urologically referred men. These and other assay-adapted ANN models, including one new iANN-based ANN, were used.

Results: The areas under the curve for the iANN (0.736) and MLP (0.745) were equal but showed no differences to %fPSA (0.725) in the Finnish group. Only the new iANN-based ANN reached a significant larger area under the curve (0.77). At 95% sensitivity, the specificities of MLP (33%) and the new iANN-based ANN (34%) were significantly better than the iANN (23%) and %fPSA (19%). Reverse methodology using the MLP model on the referred patients revealed, in contrast, a significant improvement in the areas under the curve for iANN and MLP (each 0.83) compared with %fPSA (0.70). At 90% and 95% sensitivity, the specificities of all LR and ANN models were significantly greater than those for %fPSA.

Conclusions: The ANNs based on different PSA assays and populations were mostly comparable, but the clearly different patient composition also allowed with assay adaptation no unbiased ANN application to the other cohort. Thus, the use of ANNs in other populations than originally built is possible, but has limitations.