Comparison of Artificial Neural Network with Logistic Regression as Classification Models for Variable Selection for Predict ion of Breast Cancer Patient Outcomes

The aim of this study was to compare multilayer p ercept ron n eu r al n etwor ks ( NNs) w ith s tandard log i stic re g ression (LR) toidentify ke y cov ar iates impacting on mor t alit y f rom cancer causes, disease-free sur v iv al (DFS), and disease recur rence usingArea Un der Receiver- Op er ating Char acter i stics ( AURO C) in breast cancer p atients. From 1996 to 2004, 2,535 patients diag n osedw i th pr imar y breast cancer entered into t he study at a sing le French cent re, w here the y re ceived standard t reat ment. For specificmor t alit y as well as DFS analysis, t he RO C cur ves were g reater w i th the N N m odels compare d to L R m odel w i th better sensitiv it yand s pecificit y. Four predictive factors were retained by b oth appro aches for mor t alit y : clinical size stage, Scarff Bloom Richard song r ade, nu mber of invaded n odes, and progesterone re ceptor. T he results enhanced the rele v ance of the u se of NN models inpredictive analysis in oncolog y, w hich appeared to be more ac cur a te in prediction in this French breast c ancer cohor t.