OBJECTIVES: To develop a model capable of predicting prostate biopsy outcomes in a large screening population, with independent validation in the referral setting. METHODS: Data from 3814 men participating in the Tyrol screening project were used to develop the model. Prospectively collected data from two independent sites in the United States (Virginia Mason Clinic, Seattle, Wash and Stanford University, Stanford, Calif) were used to validate the model independently. The Tyrol data was split randomly into three cross-validation sets, and a feed-forward, back error-propagation artificial neural network (ANN) was alternately trained on a combination of two of these data sets and validated on the remaining data set. Similarly, three logistic regression (LR) models were produced and validated using identical cross-validation data sets. The Tyrol model with the median area under receiver operating characteristic curve (AUROC) was then validated against the Virginia Mason (n = 491) and Stanford University (n = 483) data sets. RESULTS: The AUROCs for the three cross-validations were 0.74, 0.76, and 0.75 for the ANN and 0.75, 0.76, and 0.75 for the LR models. The mean AUROC for both ANN and LR was 0.75 with a standard deviation of 0.009 for ANN and 0.006 for LR. The AUROCs for the Virginia Mason and Stanford University data were 0.74 (both ANN and LR) and 0.73 (ANN) and 0.72 (LR), respectively. CONCLUSIONS: This model, designed to predict the prostate biopsy outcome, performed accurately and consistently when validated with data from two independent referral centers in the United States, suggesting that it generalizes well and may be of clinical utility to a broad range of patients.
OBJECTIVES: To develop a model capable of predicting prostate biopsy outcomes in a large screening population, with independent validation in the referral setting. METHODS: Data from 3814 men participating in the Tyrol screening project were used to develop the model. Prospectively collected data from two independent sites in the United States (Virginia Mason Clinic, Seattle, Wash and Stanford University, Stanford, Calif) were used to validate the model independently. The Tyrol data was split randomly into three cross-validation sets, and a feed-forward, back error-propagation artificial neural network (ANN) was alternately trained on a combination of two of these data sets and validated on the remaining data set. Similarly, three logistic regression (LR) models were produced and validated using identical cross-validation data sets. The Tyrol model with the median area under receiver operating characteristic curve (AUROC) was then validated against the Virginia Mason (n = 491) and Stanford University (n = 483) data sets. RESULTS: The AUROCs for the three cross-validations were 0.74, 0.76, and 0.75 for the ANN and 0.75, 0.76, and 0.75 for the LR models. The mean AUROC for both ANN and LR was 0.75 with a standard deviation of 0.009 for ANN and 0.006 for LR. The AUROCs for the Virginia Mason and Stanford University data were 0.74 (both ANN and LR) and 0.73 (ANN) and 0.72 (LR), respectively. CONCLUSIONS: This model, designed to predict the prostate biopsy outcome, performed accurately and consistently when validated with data from two independent referral centers in the United States, suggesting that it generalizes well and may be of clinical utility to a broad range of patients.
Authors: Shahrokh F Shariat; Michael W Kattan; Andrew J Vickers; Pierre I Karakiewicz; Peter T Scardino Journal: Future Oncol Date: 2009-12 Impact factor: 3.404
Authors: Stephen J Eyre; Donna P Ankerst; John T Wei; Prakash V Nair; Meredith M Regan; Gerrardina Bueti; Jeffrey Tang; Mark A Rubin; Michael Kearney; Ian M Thompson; Martin G Sanda Journal: J Urol Date: 2009-12 Impact factor: 7.450