Subject-optimized feature selection for accurate classification of cardiac beats

This paper presents a new method of subject-dependent optimization of ECG feature set for heart beats classification. The algorithm learns from randomly selected strips of the signal and preliminarily classifies the example beats. After an approval from a human operator, these classified beats are distributed to a learning and testing sets, and a genetic algorithm with aggressive mutation is used to select features with best discriminative power. Thanks to the integer representation of features in few genes only, the initial feature space of 57 elements is limited by the algorithm to 3 - 5 features optimized for a particular subject. The proposed method was tested on MIT-BIH Arrhythmia database providing reference beat types for each record. We implemented algorithms calculating 57 different parameters of the beat (mainly focused on the QRS), based on shape, acceleration, area, length etc. and used kNN and SVM as classification methods. We built the learning set out of 15 strips of 1 Os length and assume that the feature set contains maximum of 5 elements. Comparing the results to the classification based on reference minimum correlation-based selection of features we observed significant reduction of misclassified beats ratio (for SVM and 3 features from 2.7% to 0.7% in average).