Development, validation, and head-to-head comparison of logistic regression-based nomograms and artificial neural network models predicting prostate cancer on initial extended biopsy.

OBJECTIVES: Using cohorts examined by extended biopsy, we developed and validated multivariate models predicting prostate cancer on initial biopsy and examined whether these extended biopsy-based models outperform previously established models.
METHODS: Initial extended biopsy (median 22 cores) was performed in 1509 Japanese men including 1083 at Tokyo Medical and Dental University Hospital (TMDU) and 426 at Cancer Institute Hospital (CIH). Logistic regression-based nomograms 1 and artificial neural network (ANN) 1 incorporating age, digital rectal examination, and prostate-specific antigen (PSA) and free PSA, and nomogram 2 and ANN2 further incorporating transrectal ultrasound (TRUS) findings and prostate volume were constructed on the TMDU data. These and previously established models were externally validated on the CIH data set and predictive accuracy was compared directly.
RESULTS: Without TRUS-derived information, nomogram 1 outperformed the ANN1. With TRUS-derived information, nomogram 2 was more accurate than ANN2. External validation revealed applicability of the Western models to Japanese population, superiority of the nomograms over ANN models, and better predictive accuracy of our extended biopsy-based nomograms than the previous 6-10-core biopsy-based models. Using nomograms 1 and 2, 16% and 19% unnecessary biopsies would be saved at 95% sensitivity.
CONCLUSIONS: We developed new nomograms predicting prostate cancer on initial biopsy in men with PSA <20ng/ml. Predictive accuracy of these extended biopsy-based nomograms is better than those of previously established models based on 6-10-core biopsies. Our models might help clinicians to decide if a patient requires biopsy and to avoid unnecessary biopsies.