Introducing ROC Curves as Error Measure Functions: A New Approach to Train ANN-Based Biomedical Data Classifiers

摘要

This paper explores the usage of the area (Az) under the Receiver Operating Characteristic (ROC) curve as error measure to guide the training process to build machine learning ANN-based classifiers for biomedical data analysis. Error measures (like root mean square error, RMS) are used to guide training algorithms measuring how far solutions are from the ideal classification, whereas it is well known that optimal classification rates do not necessarily yield to optimal Az's. Our hypothesis is that Az error measures can guide existing training algorithms to obtain better Az's than other error measures. This was tested after training 280 different configurations of ANN-based classifiers, with simulated annealing, using five biomedical binary datasets from the UCI machine learning repository with different test/train data splits. Each ANN configuration was trained both using the Az and RMS based error measures. In average Az was improved in 7.98% in testing data (9.32% for training data) when using 70% of the datasets elements for training. Further analysis reveals interesting patterns (Az improvement is greater when Az are lower). These results encourage us to further explore the usage of Az based error measures in training methods for classifiers in a more generalized manner.