Logarithmic-sensitivity index as a stopping criterion for automated neural networks

Previous work on mortality prediction for coronary surgery patients with neural networks has been hampered by its low rate of occurrence (although this is certainly good from a medical point of view). The mortality rate in the coronary artery bypass grafting database is typically less than 5%, thus the identification of non-survivors is difficult. One possible remedy to this dilemma is to artificially increase the mortality rate in the training set presented to the network during its learning phase. This method increased the likelihood that the network will recognize patterns attributable to non-survivors and improved results. To optimize the sensitivity and specificity performance of a neural network at the same time, a new performance index - the logarithmic-sensitivity index - was introduced. Its ability to identify the optimal stopping point when training with an automated network was compared with the results found when the network was optimized manually. Results show that the log-sensitivity index succeeded in finding a good balance between sensitivity and specificity of the test set and the automated results had a higher mean sensitivity although it was within the error bounds of the manual results. This means that the log-sensitivity index is a valuable timesaving tool, because the networks can be run automatically without user supervision.