Comparison of genetic algorithm neural network with logistic regression for predicting outcome after surgery for patients with nonsmall cell lung carcinoma

Abstract

BACKGROUND. Neural networks have been used to predict outcome in cancer patients. Their accuracy compared with standard statistical methods has not been fully assessed. METHODS. In this study, the authors examined the ability to predict the outcome of surgery in 620 patients with nonsmall cell lung carcinoma (NSCLC) by a genetic algorithm neural network (GANN) using Bayes' theorem compared with logistic regression, and the predictive value of tumor volume measures in addition to standard indices such as histologic type and stage. Predictive methods were compared by examining accuracy of classifying target outcome of patients living or dead at 6, 12, 18, and 24 months after surgery. RESULTS. GANN was a significantly better predictor of outcome than logistic regression at all time points (McNemar, P < 0.01). Measures of tumor volume produced significant improvement in the prediction of 12-, 18-, and 24-month time points with GANN, and at 18- and 24-month time points with logistic regression (Wilcoxon matched pairs signed rank test, P < 0.02). CONCLUSIONS. In this study of surgically treated NSCLC patients, outcome predictions were significantly improved by including measures of tumor volume. For predicting individual patient outcome, GANN was found to be highly accurate and significantly better than logistic regression.