Pattern recognition with discrete and mixed data : theory and practice

This thesis is devoted to aspects related to the analysis of medical data bases in the context of pattern recognition. It contains both theoretical aspects and practical applications and its scope includes questions and problems that arise when applying pattern recognition methods and techniques to this type of data. The goal of the application of statistical pattern recognition techniques to medical records, is the classification of the (disease) patterns that may be present in such records in terms of the information they contain. Typically, a medical record contains a description of history, symptoms, results from laboratory tests, signals, etc., all related to a given patient, i.e. all the information normally required by a physician when making a diagnosis and/or a prognosis. Pattern recognition may be used in order to obtain procedures (computer implemented algorithms) to assign diagnostic or prognostic classes to a given patient, on the basis of information also used by a physician. These procedures are not intended to replace but to assist the physician in the decision making process. The procedures are called classifiers or discriminants and the symptoms, signals, etc., are called features. Each individual record is termed an object, and a collection of objects with qualitatively and/or quantitatively similar characteristics, as established by an expert, is called a class. It should be clear that pattern recognition can be applied to a wide variety of areas and problems, of which (computer-aided) medical decision making is just an example. In order to arrive at a classifier and restricting ourselves to what is called supervised learning, a set of objects known a-priori to belong to two or more classes (depending on the problem at hand) is needed. In this set, each object must be represented by a group of features and have a class assigned to it. The role of medical data bases is now clear: they are the set of objects required for supervised learning.