Clinical Examples as Non-uniform Learning and Testing Sets

Clinical examples are widely used as learning and testing sets for newly proposed artificial intelligence-based classifiers of signals and images in medicine. The results obtained from testing are usually taken as an estimate of the behavior of automatic recognition system in presence of unknown input in the future. This paper investigates and discusses the consequences of the non-uniform representation of the medical knowledge in such clinically-derived experimental sets. Additional challenges come from the nonlinear representation of the patient status in particular parameters' domain and from the uncertainty of the reference provided usually by human experts. The presented solution consists of representation of all available cases in multidimensional diagnostic parameters or patient status spaces. This provides the option for independent linearization of selected dimensions. The recruitment to the learning set is then based on the case-to-case distance as selection criterion. In result, the classifier may be trained and tested in a more suitable way to cope with unpredicted patterns.