On the overall ROC of multistage systems

The receiver operating characteristic (ROC) curve is a useful tool to evaluate the performance of classifiers, and is widely used in signal detection, pattern recognition and machine learning. For complex object classification, multiple single classifiers are often used and they are concatenated into a multistage classification system. Thus, it is necessary to obtain the overal ROC curve, because the ROC curves of the individual classifiers are not useful for the overall system since it has multi-level decision thresholds. In this paper, a systematic approach was introduced for measuring the performance of multistage systems via estimating the overall ROC curve. Two new ROC models sharing the same properties of classical ROC curves were proposed, inspired by the Gaussian and logistic distributions. The models were then experimented on a recently introduced multistage system for epileptic spike classification from electroencephalogram data. Experimental results indicated that the proposed ROC models can be used for multistage classification systems.