Year: 2004
Type: article
Abstract: We thank Dr. Stephan and colleagues for their interest in our recent report.1 The ideal strategy for determining a patient's risk for a positive prostate biopsy is unknown. Stephan and colleagues compared the results of studies conducted by his group and others using artificial neural networks (ANN) with the results of our logistic regression (LR) model. An important consideration before comparing these studies is whether the patient groups are evenly matched, as data sets that contain increased numbers of high-risk patients will overestimate model's sensitivity and specificity. Our study had a positive biopsy rate of 24.3%, which is similar to detection rates in population-based screening trials. We included only patients with prostate-specific antigen (PSA) levels ≤ 10 ng/mL who presented before their initial prostate biopsy was performed. In contrast, the study conducted by Stephan et al. included men with PSA values > 10 ng/mL and men with a previous diagnosis of malignant disease in model development, which accounted for the high incidence of prostate carcinoma in their study (60.7%). Inclusion of patients who are known to have malignant disease introduces the potential for selection and examiner biases, which were not a factor in our study, as all data variables were recorded before cancer status was known. For model assessment and selection, ANN ideally requires three independent data sets—namely, training, validation, and test sets.2 A training set is used to fit the model, whereas the validation set is used to estimate prediction error and prevent overfitting, a well recognized problem in ANN. The test set then is used to evaluate model performance. Finne et al.3 used the leave-one-out cross-validation method to avoid overfitting; although this approach addresses the problem of overfitting, it is likely to overestimate the model performance, due to a lack of a completely independent test data set such as the one used in our study. In a recent metaanalysis performed by Sargent et al.,4 LR outperformed ANN in large clinical studies. These investigators concluded that ANN should not take the place of standard statistical methods. Few studies have compared ANN with LR for the prediction of prostate carcinoma detection. Stephan et al.5 previously showed that LR performed as effectively as ANN did in this setting. We remain unconvinced that ANN offers significant advantages over LR-based methods, but this hypothesis would be best tested by applying both methodologies to independent data sets. We agree with Stephan and colleagues that inclusion of the ratio of free PSA to total PSA (% fPSA) or other molecular forms could improve our model. In our study, however, a high PSAD was the strongest predictor of prostate carcinoma (relative risk, 11.4), and PSAD has been shown by others to be equivalent or superior to % fPSA.6 Therefore, any model improvement generated by the inclusion of % fPSA probably would be incremental.
Source: Cancer
DOI: 10.1002/cncr.20197
Authors: Mark Garzotto, Tomasz M. Beer, Motomi Mori
Language: English
Subfield: Pulmonary and Respiratory Medicine
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4251653473