Gene expression classifiers and out-of-class samples detection

The proper application of statistics, machine learning, and data-mining techniques in routine clinical diagnostics to classify diseases using their genetic expression profile is still a challenge. One critical issue is the overall inability of most state-of-the-art classifiers to identify out-of-class samples, i.e., samples that do not belong to any of the available classes. This paper shows a possible explanation for this problem and suggests how, by analyzing the distribution of the class probability estimates generated by a classifier, it is possible to build decision rules able to significantly improve its performances.