Generation of decision trees from EEG data

Tests and evaluates a machine learning methodology on data concerning brainstem auditory evoked potentials (BAEPs) of human patients aged 2-15 years described using the standard attributes. The used data set consists of data collected during the years 1993-1995 at the Faculty Hospital in Motol, Prague (the training set consisted of 250 cases). The data have been processed by a program inspired by the algorithm ID3, constructing a decision tree from a batch of training examples. This program is implemented in the Pascal programming language. It contains many functions supporting experiments with the generated decision tree, e.g. its pruning, classification of new examples and estimation of its predictive quality. Any decision tree identifies several distinguished points among the possible values of each considered attribute-the tree suggests a discretization of the set of its values. This discretization can be generated by various heuristic methods. We try to identify the discretization method most suitable for the considered BAEP domain, where the diagnostic conclusion seems to be closely related to the statistical properties of individual attributes. There will be compared the diagnostic results obtained from the decision trees induced through various approaches from pure BAEP data, from the BAEP dataset extended by a few derived attributes (e.g. differences between latencies of some waves) and the tree generated from a man-made set of rules. Special attention is devoted to the mis-matches and exceptions identified