Evolution of Artificial Neural Network Architecture: Prediction of Depression after Mania

Abstract

Artificial neural networks (ANNs) are compared to standard statistical methods for outcome prediction in biomedical problems. A general method for using genetic algorithms to "evolve" ANN architecture (EANN) is presented. Accuracy of logistic regression, a fully interconnected ANN, and an EANN for predicting depression after mania are examined. All methods showed very good agreement (training set accuracy, chi-square all p < 0.01). However, significant differences were found for stability (test set accuracy); logistic regression being the most unstable and EANN being significantly more stable than a fully interconnected ANN (McNemar p < 0.01). We conclude that the EANN method enhances ANN stability. This approach may have particular relevance for biomedical prediction problems, such as predicting depression after mania.