Hybrid Feature Selection Method Based on Neural Networks and Cross-Validation for Liver Cancer With Microarray

Abstract

This paper proposes a method that extracts a feature set for accurate disease diagnosis from a feature (aptamer) array. Our method uses an artificial intelligence of the neural network and 10-fold crossvalidations and is verified by the p-value of the aptamer array response to specimens of 80 liver cancer patients and 310 healthy people. The proposed method is compared with the one-way ANOVA method in terms of accuracy, the number of features, and computing time to determine the feature set required to achieve the same accuracy. An increase in the number of features dramatically improves the diagnosis accuracy of the two methods for 2-10 features. The accuracies with 10 features are 93.5% and 87.5%, and the increases in accuracy per additional feature are 3.39% and 2.65% for our method and the one-way ANOVA, respectively. For the same accuracy, our method needs only 1/2-1/3 number of features of the ANOVA. An interesting statistical characteristic of cross-validation is that diagnostic accuracy saturates after 10000 cross-validations.

Keywords

• ANOVA
• artificial intelligence
• disease diagnosis
• feature selection
• microarray
• neural network