K H Gulkesen1, I T Koksal, U Bilge, O Saka. 1. Akdeniz University, Faculty of Medicine, Department of Biostatistics and Medical Informatics, Antalya, Turkey. hgulkesen@akdeniz.edu.tr
Abstract
PURPOSE: Several concepts to improve the diagnostic accuracy of prostate specific antigen (PSA) for prediction of prostate cancer have been studied. The aim of this study was to examine and compare the methods used for improving the diagnostic accuracy of PSA in a country with low incidence of prostate cancer. METHODS: 997 patients with prostate biopsy were included into study. Predictive models using PSA, PSA density (PSAD), free PSA/total PSA (f/tPSA), binary logistic regression (LR) analysis, artificial neural networks (ANNs), and decision trees (DTs) have been developed. For LR, ANNs and DTs, a validation group consisting of 241 cases was reserved. RESULTS: 193 (19%) biopsies out of 997 showed prostatic cancer. Median PSAD in patients with malignant and benign lesions were 0.21 and 0.16, respectively (p<0.001). According to 25% f/tPSA cut-off level, 18.4% of the patients with PSA<25% and 16.0% of the patients with PSA>25% had prostate cancer (p=0.423). Receiver operating characteristics (ROC) area under the curve (AUC) values for PSA, PSA density, f/tPSA, LR, ANNs, and DTs were 0.587, 0.625, 0.560, 0.678, 0.644, and 0.698, respectively. ROC AUCs in the validation group for LR, ANNs and DTs were 0.717, 0.516 and 0.629 respectively. CONCLUSIONS: For cases with f/tPSA<25%, no increased probability for prostatic carcinoma was observed. Multivariate models have higher AUCs than PSA, PSAD or f/tPSA. LR, DTs and ANNs showed similar results, however application of ANNs to the validation group produced a significantly lower AUC, limiting the value of ANNs in this situation.
PURPOSE: Several concepts to improve the diagnostic accuracy of prostate specific antigen (PSA) for prediction of prostate cancer have been studied. The aim of this study was to examine and compare the methods used for improving the diagnostic accuracy of PSA in a country with low incidence of prostate cancer. METHODS: 997 patients with prostate biopsy were included into study. Predictive models using PSA, PSA density (PSAD), free PSA/total PSA (f/tPSA), binary logistic regression (LR) analysis, artificial neural networks (ANNs), and decision trees (DTs) have been developed. For LR, ANNs and DTs, a validation group consisting of 241 cases was reserved. RESULTS: 193 (19%) biopsies out of 997 showed prostatic cancer. Median PSAD in patients with malignant and benign lesions were 0.21 and 0.16, respectively (p<0.001). According to 25% f/tPSA cut-off level, 18.4% of the patients with PSA<25% and 16.0% of the patients with PSA>25% had prostate cancer (p=0.423). Receiver operating characteristics (ROC) area under the curve (AUC) values for PSA, PSA density, f/tPSA, LR, ANNs, and DTs were 0.587, 0.625, 0.560, 0.678, 0.644, and 0.698, respectively. ROC AUCs in the validation group for LR, ANNs and DTs were 0.717, 0.516 and 0.629 respectively. CONCLUSIONS: For cases with f/tPSA<25%, no increased probability for prostatic carcinoma was observed. Multivariate models have higher AUCs than PSA, PSAD or f/tPSA. LR, DTs and ANNs showed similar results, however application of ANNs to the validation group produced a significantly lower AUC, limiting the value of ANNs in this situation.
Authors: Suman Kundu; Raluca Mihaescu; Catherina M C Meijer; Rachel Bakker; A Cecile J W Janssens Journal: Front Genet Date: 2014-06-13 Impact factor: 4.599
Authors: G Pourmand; R Ramezani; B Sabahgoulian; F Nadali; Ar Mehrsai; Mr Nikoobakht; F Allameh; Sh Hossieni; A Seraji; M Rezai; F Haidari; S Dehghani; R Razmandeh; B Pourmand Journal: Iran J Public Health Date: 2012-02-29 Impact factor: 1.429