Abstract 213: Automatic Prediction of Arrhythmia Severity Using Time and Frequency Domain Features
Introduction: Early arrhythmia prediction may help in resuscitation. Although some arrhythmias have minimal consequences, others may eventually lead to stroke or sudden cardiac death. ECG is the primary tool for arrhythmia detection and is often visually analyzed for arrhythmias. However, a reliable, objective, real-time, and automated method to process and classify arrhythmias from the hidden information in ECG patterns such as the intervals and amplitudes of ECG waves could be of considerable importance, in particular for applications such as resuscitation to warn clinicians of early deterioration. We hypothesize that ECG beats' classification can predict the severity of arrhythmia.
Methods: Discrete wavelet transform (DWT) was applied to process ECG signal for deflection points detection (P-QRS-T). After detection, beat-by-beat features from time domain, including QT-interval, ST-segment, QRS width, and amplitudes of P and T, as well as frequency domain including high frequency coefficients, mean, median and the standard deviation of the coefficients by applying DWT using Daubechies 4 at level 2, are extracted. The features are then used as input attributes to train a support vector machine with 10-fold cross validation to classify every beat into one of three functional classes: normal (N), premature ventricular contraction (V) or premature atrial contraction (A). The output is a vector of classified beats. To find severe abnormalities, a window of 30 last predicted output classes is formed. If the vector is found to contain three V beats in a row, or six V or A beats in (not necessarily consecutive), the system reports a severe abnormality to the user. The MIT-BIH arrhythmia database was used to test the algorithm.
Results: The database contains 48 excerpts, each lasting 30-min with a total of more than 109,000 annotated beats. The overall sensitivity and specificity of the new approach was found to be 95.4% and 99.2%, respectively. A severe abnormality was reported for 17 subjects.
Conclusions: The proposed method analyzes, detects and predicts the severity of three types of arrhythmias. The method may be useful to detect ectopic beats and other patterns that may precede development of life-threatening arrhythmias such as ventricular fibrillation.
- © 2010 by American Heart Association, Inc.