Predictive models for prostate cancer based on logistic regression and artificial neural network

This study aims to develop two predictive models for prostate cancer and to compare their diagnostic performance on the basis of logistic regression (LR) and artificial neural network (ANN). The data used in this study were collected from 586 men with histologically confirmed diagnoses. One third of those data were randomly retained for model validation. Bivariate correlate and logistic regression analyses were applied for predictor selection. The diagnostic abilities were evaluated and compared by performance parameters (including sensitivity, specificity, Youden index and accuracy) and receiver operating characteristic (ROC) analyses. The ultimate predictors for modeling included age, percent free prostate specific antigen (%fPSA), prostate volume, and PSA density (PSAD). The sensitivity, specificity, Youden index, and accuracy of LR model were 0.8623, 0.8616, 0.7239 and 0.8618 versus 0.8563, 0.8854, 0.7417 and 0.8771 for the ANN. The area under the curve (AUC) was 0.924 for LR model and 0.933 for the ANN, respectively. Comparison of ANN to LR model showed no statistically significant difference. The above results verified the high diagnostic validity of these two models. LR and ANN model can be incorporated into urologic decision making to assist the clinician making diagnosis and treatment and to reduce the unnecessary biopsies.