Coronary surgery mortality prediction using artificial neural networks.

Abstract

This thesis demonstrates the application of a feedforward backpropagation-trained artificial neural network using the weight-elimination cost function to the estimation of in-hospital mortality for coronary artery bypass grafting patients from the San Francisco Heart Institute in Daly City, California, USA. The highly-skewed a priori statistics due to the low mortality rate present difficulties for modelling this data. Artificial training and test datasets with higher mortality rates were developed to improve the classification performance of the artificial neural networks. Sensitivity was considered the most important measure of performance for this work. Given that current mortality risk models are unable to accurately identify high-risk patients (those who do not survive the surgery), focussing on increasing the sensitivity rate will indicate when more of the patients who are difficult to classify are correctly identified. The final result was an increase in sensitivity when training with a dataset with a higher mortality rate than the test set. This dataset modification approach resulted in only small changes for other performance measures (specificity, predictive positive value, predictive negative value, and correct classification rate), and thus helped to achieve the main goal of the study.