An application of belief merging for the diagnosis of oral cancer

Abstract

Machine learning employs a variety of statistical, probabilistic, fuzzy and optimization techniques that allow computers to “learn” from examples and to detect hard-to-discern patterns from large, noisy or complex datasets. This capability is well-suited to medical applications, and machine learning techniques have been frequently used in cancer diagnosis and prognosis. In general, machine learning techniques usually work in two phases: training and testing. Some parameters, with regards to the underlying machine learning technique, must be tuned in the training phase in order to best “learn” from the dataset. On the other hand, belief merging operators integrate inconsistent information, which may come from different sources, into a unique consistent belief set (base). Implementations of merging operators do not require tuning any parameters apart from the number of sources and the number of topics to be merged. This research introduces a new manner to “learn” from past examples using a non parametrised technique: belief merging. The proposed method has been used for oral cancer diagnosis using a real-world medical dataset. The results allow us to affirm the possibility of training (merging) a dataset without having to tune the parameters. The best results give an accuracy of greater than 75%.