Predicting prostate biopsy outcome: artificial neural networks and polychotomous regression are equivalent models

Abstract

Complex statistical models utilizing multiple inputs to derive a risk assessment may benefit prostate cancer (PC) detection where focus has been on prostate-specific antigen (PSA). This study develops a polychotomous logistic regression (PR) model and an artificial neural network (ANN) for predicting biopsy results, particularly for clinically significant PC. There were 3,025 men undergoing TRUS-guided biopsy (BX) with PSA <10ng/ml selected. BX outcome classified as benign, atypical small acinar proliferation or high-grade prostatic intraepithelial neoplasia (ASAP/PIN), non-significant (NSPC) or clinically significant PC (CSPC). PR and ANN models were developed to distinguish between BX categories. Predictors were age, PSA, abnormal digital rectal examination (DRE), positive transrectal ultrasound (TRUS) and prostate volume. Among the BXs, 44% were benign, 14% ASAP/PIN, 16% NSPC and 25% CSPC. Median age, PSA and volume were 64years, 5.7ng/ml and 50cc. TRUS lesion was present in 47%, and DRE was abnormal in 39%. PR and ANN models did not differ on percentage BX outcomes correctly predicted (55, 57%, respectively) and were equally poor for both ASAP/PIN (0%) and NSPC (2%). For PR and ANN, 74-78% ASAP/PIN predicted benign, 2% NSPC and 20-24% CSPC. For NSPC, 69-71% predicted benign, 27-29% CSPC. Benign outcomes were well identified (86-88%), although 12-13% classified CSPC. CSPC was correctly identified in 65-66% with misclassifications largely benign (33% for PR and ANN). Neither PR nor ANN was able to distinguish between the four biopsy outcomes: ASAP/PIN and NSPC were not distinguished from benign or CSPC. ANN did not perform better than PR. Inclusion of additional predictors may increase the performance of statistical models in predicting BX outcome.