Nonlinear Characterization of ECG Signals for Automatic Arrhythmia Detection

The objective of this report is to design a system to distinguish healthy and pathological ECG signals through the study of measurements based on Chaos Theory and the use of Artificial Neural Networks. A new database is created from Arrhythmia Database. Afterward, the study and feature extraction are carried out, specifically by using: Shannon entropy, Maximum Exponent of Lyapunov, Correlation Dimension, Correlation entropy, Lempel-Ziv Complexity and Hurst Exponent parameters. Regarding the classifier, a system based on Artificial Neural Networks is chosen and it is defined two classifications: between-different classes (diseases) and normal heartbeat and healthy-pathologic kind. Different studies are conducted and they are performed some parameter experimentations such as the optimum frame size of the database registers or the number of neurons. The utility of the use of these nonlinear parameters and the performance of the detection system are also assessed. The optimum frame size is estimated at 20s. The Hurst Exponent, Maximum Exponent of Lyapunov (Rosenstein) and Lempel-Ziv Complexity are the parameters reporting the better success rates. Regarding the number of neurons, from 10 onwards, a significative difference between percentages of success obtained does not exist. Into account in the overall assessment, the system offers 95% accuracy for the healthy-pathologic classification. This automated system, able to detect cardiac pathologies, helps to resolve subjective problems in heart disease diagnosing and also to facilitate doctor’s work.