Accurate arrhythmia classification using auto-associative neural network.

Currently about one in eighteen of the American population suffer from cardiac Arrhythmias that lead to Coronary Heart Diseases and this rate is steadily increasing. An early monitoring and diagnosis of Arrhythmia based on Electrocardiogram signals can help in reducing mortality. This paper primarily focuses on the application of Auto Associative Neural Network as a new classification approach, which does not require feature extraction task. The weights of a trained Neural Network are stored as class representative models that results in high compression gain with respect to the size of training data. The evaluation of the proposed technique is tested on segmented ECG beats of four different classes of Arrhythmia excluding normal pattern. These beats have been extracted from the MIT/BIH Arrhythmia database and compared against the state-of-the art template matching technique such as Dynamic Time Warping. The proposed technique yields an average accuracy of more than 97% and a relative compression gain of above 90%.