Session S34.1

Hyperbox Classifiers for ECG Beat Analysis

G Bortolan*, II Christov, W Pedrycz

ISIB-CNR
Padova, Italy

Hyperbox classifiers have been investigated for the detection and classification of different types of heartbeats in the electrocardiogram (ECG), which is of major importance in the diagnosis of cardiac dysfunctions. In particular the learning capacity and the classification ability for normal beats (N) and premature ventricular contractions (PVC) have been tested, with particular interest in the aspect of the interpretability of the results.
Hyperbox classifiers are a useful tool for an easily interpretability of the classification rules and a simple structure to envision the results. The learning process is developed with a hybrid architecture in two phases. First with the fuzzy clustering c-means, the clusters are localized and characterized by the corresponding prototypes. In the second step, hyperboxes will be defined around the prototypes, and optimized with the use of genetic algorithm, which are very suitable for problems with high dimensions. The MIT-BIH arrhythmia database has been used for testing and validating the proposed method. A total of 26 morphology features have been extracted from ECG and VCG signals and include amplitudes, areas, durations and combinations of them.
Results.
Two different experiments were performed: all the 26 features were considered with different number of hyperboxes (HB), and a limited number of features was investigated for the geometrical interpretability. The first test was performed using all the features and considering one, two or 7 HB. The mean sensitivity and specificity with one HB were 97.8% and 97.8% for N and 79.7% and 99.7% for PVC, whereas two HB produced 98.0% and 99.2% for N and 78.3% and 98.2% for PVC. The use of seven HB improved the results, with the drawback of limiting the interpretability and increasing the complexity of the classifier. The second experiment considered the use of few parameters, and it gave the visual possibility/ability to evaluate the geometrical interpretability of the classifier.

(Abstract Control Number: 50)