Hypothesis testing with classifier systems for rule-based risk prediction

Abstract

Analysis of medical datasets has some specific requirements not always fulfilled by standard Machine Learning methods. In particular, heterogeneous and missing data must be tolerated, the results should be easily interpretable. Moreover, with genetic data, often the combination of two or more attributes leads to non-linear effects not detectable for each attribute on its own. We present a new ML algorithm, HCS, taking inspiration from learning classifier systems, decision trees and statistical hypothesis testing. We show the results of applying this algorithm to a well-known benchmark dataset, and to HNSCC, a dataset studying the connection between smoke and genetic patterns to the development of oral cancer. © Springer-Verlag Berlin Heidelberg 2007.