A comparison of Cox Regression and neural networks for risk stratification in cases of acute lymphoblastic leukaemia in children

For most diseases there is considerable interest in the problem of classification, both in relation to medical diagnosis and for prognosis. Multivariate statistical methods are conventionally used as an aid to clinical decision making. Neural Networks (NNs) offer an alternative approach to this type of classification problem. Exploiting 1271 cases from the United Kingdom Medical Research Council UKALL X trial for childhood Acute Lymphoblastic Leukaemia (ALL), cases were stratified as 'high risk' or 'standard risk' using both the survival analysis technique of Cox Regression and trained neural networks. Based on 10 random trials with a further 300 cases, and predicting overall Jive year survival from age, sex and white cell count only, there was no significant difference between the two approaches in terms of mean Receiver Operating Characteristic area, though the regression model was slightly superior to a single neural network at high sensitivity (Wilcoxon signed rank test; p = 0.033). A composite of two networks, one of which included additional prognostic factors, restored the position of no significant difference. It was concluded that in the UKALL X dataset, factors predictive of outcome are fully described by a Cox regression analysis, and that a neural network-based analysis identified no additional prognostic features. The value of the network analysis lay in suggesting that the maximum amount of prognostic information has been extracted from the database.