ECG feature relevance in a fuzzy arrhythmia classifier

A good fuzzy classifier for a three-class arrhythmia problem is analyzed “a posteriori”, in order to estimate the importance of the ECG features for the final system performance. A set of fuzzy rules is automatically built on forty MIT-BIH database´s files, using fourteen ECG measures to characterize each beat. The implemented fuzzy model correctly classifies 92% of normal beats, 80% of PVBs and 56% of SVPBs of the test set. The information contained in the fuzzy model is then quantified and the gain in information, derived from application of the fuzzy rules along a given input dimension, is measured. Only three ECG features (QRS width, PR segment and prematurity degree) seem to carry the main responsibility for the correct classification of the three arrhythmic classes. The uninformative character of ECG measures with lowest information gain is confirmed by the almost unaltered-when not improved-performance of the classifier when these features are removed from the input vector