Snow and coworkers are among the first to report the application of ANNs to predictive modeling in prostate cancer. In 1994, Snow investigated the use of an ANN to predict positive biopsy in 1787 men from a screening population with PSA levels greater than 4ng/ml.23 The inputs included age, PSA level, digital rectal examination (DRE) findings and transrectal ultrasound (TRUS) findings. They reported a sensitivity of 84% and a specificity of 88% for predicting biopsy result at an ANN output cut-off of 0.35. The model's AUROC was not reported. The authors concluded that ANNs may be useful in reducing the number of unnecessary biopsies and detection of clinically insignificant tumors, and thus reducing costs and morbidity. However, the ability of the model by Snow and coworkers to generalize to new patient data may suffer because, according to the report, the test set used to select the best weight matrix was also used to validate the model. In a subsequent study,30 Snow and coworkers used data from another screening population to develop an ANN model for predicting positive biopsy in 1500 men who had an abnormal PSA or DRE examination. They reported a sensitivity of 72% and a specificity of 78%. Stamey and coworkers reported the results of a study31 using a commercially available ANN known as ProstAsure (Horus Therapeutics, Inc., Savananah, GA, USA). This model was developed using the following variables from 416 men: age, PSA level, prostatic acid phosphatase and creatine kinase isoenzymes (CK-MM, CK-MB, and CK-BB). The model's sensitivity was reported to be 81% and the specificity was reported to be 92%. In a separate report, Babaian reported a statistically significant advantage for the ProstAsure® Index compared to free PSA in detecting pro-
Was this article helpful?