On Knowledge Discovery Experimented with Otoneurological Data

Abstract

Diagnosis of otoneurological diseases can be challenging due to similar kind of and overlapping symptoms that can also vary over time. Thus, systems to support and aid diagnosis of vertiginous patients are considered beneficial. This study continues refinement of an otoneurological decision support system ONE and its knowledge base. The aim of the study is to improve the classification accuracy of nine otoneurological diseases in real world situations by applying machine learning methods to knowledge discovery in the otoneurological domain.

The phases of the dissertation is divided into three parts: fitness value formation for attribute values, attribute weighting and classification task redefinition. The first phase concentrates on the knowledge update of the ONE with the domain experts and on the knowledge discovery method that forms the fitness values for the values of the attributes. The knowledge base of the ONE needed update due to changes made to data collection questionnaire. The effect of machine learnt fitness values on classification are examined and classification results are compared to the knowledge set by the experts and their combinations. Classification performance of nearest pattern method of the ONE is compared to k-nearest neighbour method (k-NN) and Naïve Bayes (NB). The second phase concentrates on the attribute weighting. Scatter method and instance-based learning algorithms IB4 and IB1w are applied in the attribute weighting. These machine learnt attribute weights in addition to the weights defined by the domain experts and equal weighting are tested with the classification method of the ONE and attribute weighted k-NN with One-vs-All classifiers (wk-NN OVA). Genetic algorithm (GA) approach is examined in the attribute weighting. The machine learnt weight sets are utilized as a starting point with the GA. Populations (the weight sets) are evaluated with the classification method of the ONE, the wk-NN OVA and attribute weighted k-NN using neighbour’s class-based attribute weighting (cwk-NN). In the third phase, the effect of the classification task redefinition is examined. The multi-class classification task is separated into several binary classification tasks. The binary classification is studied without attribute weighting with the k-NN and support vector machines (SVM).