Early Detection Research Network

Prediction of survival in patients with esophageal carcinoma using artificial neural networks.

Accurate estimation of outcome in patients with malignant disease is an important component of the clinical decision-making process. To create a comprehensive prognostic model for esophageal carcinoma, artificial neural networks (ANNs) were applied to the analysis of a range of patient-related and tumor-related variables.

Clinical and pathologic data were collected from 418 patients with esophageal carcinoma who underwent resection with curative intent. A data base that included 199 variables was constructed. Using ANN-based sensitivity analysis, the optimal combination of variables was determined to allow creation of a survival prediction model. The accuracy (area under the receiver operator characteristic curve [AUR]) of this ANN model subsequently was compared with the accuracy of the conventional statistical technique: linear discriminant analysis (LDA).

The optimal ANN models for predicting outcomes at 1 year and 5 years consisted of 65 variables (AUR = 0.883) and 60 variables (AUR = 0.884), respectively. These filtered, optimal data sets were significantly more accurate (P < 0.0001) than the original data set of 199 variables. The majority of ANN models demonstrated improved accuracy compared with corresponding LDA models for 1-year and 5-year survival predictions. Furthermore, ANN models based on the optimal data set were superior predictors of survival compared with a model based solely on TNM staging criteria (P < 0.0001).

ANNs can be used to construct a highly accurate prognostic model for patients with esophageal carcinoma. Sensitivity analysis based on ANNs is a powerful tool for seeking optimal data sets.

Imamura M, Maeda M, Meltzer SJ, Mori Y, Sato F, Selaru FM, Shibata D, Shimada Y, Stass SA, Watanabe G


Cancer, 2005, 103 (8)