Ecg Delineation for Qt Interval Analysis Using an Unsupervised Learning Method

Abstract

This paper presents a novel approach for automatic ECG delineation with focus on QT interval estimation, using an unsupervised learning algorithm. A three-dimensional feature space is created which uses the characteristics of ECG waveform at its inflection points. To this end, three features are introduced, including the Truncated Energy, which makes our method robust to baseline wandering and noise. Using the fact that the logarithm of features exhibits a mixture of four Gaussian distributions, each for one of the P wave, QRS complex, T wave and baseline, an unsupervised clustering algorithm based on Expectation Maximization is applied. The experimental results reveal that the proposed algorithm extracts the ECG waves accurately, even if they have a very low energy and amplitude. No pre-processing and windowing approach is required in the proposed method resulting in a significantly higher resolution and lower computational complexity in estimating the onset and offset of ECG waves. Furthermore, the proposed algorithm is robust to noise and baseline wandering, thanks to the Laplacian of Gaussian filter employed for inflection point detection.