Evaluation of the logarithmic-sensitivity index as a neural network stopping criterion for rare outcomes

Rare outcomes are often difficult to classify using an automated neural network. The logarithmic-sensitivity index was introduced to optimize both sensitivity and specificity at the same time while slightly favouring sensitivity. This index succeeded in identifying the optimal stopping point when training an automated network to classify rare outcomes. The automated networks achieved equal or better classification performance than the manually-optimized networks. When the classification performance of the maximum log-sensitivity index is compared with networks whose stopping criteria are based on the highest correct classification rate or the lowest mean squared error, the log-sensitivity networks better classified the rare outcomes (higher sensitivity) while maintaining a sufficiently high specificity rate. This means that the log-sensitivity index is a valuable time-saving tool, because the networks can be run automatically without user supervision and classification performance is not compromised.