Predicting preterm birth using artificial neural networks

Abstract

This paper has three contributions: 1) to evaluate how changing the a priori distribution of the training set affects the performance of a back-propagation feed-forward artificial neural network (ANN) in predicting PreTerm Birth (PTB) for obstetrical patients, 2) to assess the effectiveness of the weight elimination cost function in improving the ANN's classification of PTB and in identifying a new minimal dataset, and (3) to determine if PTB can be predicted outside of clinical trial situations using data readily available to the physician during obstetrical care. The ANN was trained and tested on cases with 8 input variables describing the patient's obstetrical history; the output variable was PTB before 37 weeks gestation. To observe the impact of training with a higher-than-normal prevalence, an artificial training set with a PTB rate of 23% was created. Networks trained on higher-than-normal prevalence achieved higher sensitivity rates and greater C-index values, at the cost of slightly lower specificity and correct classification rates.