Author + information
- 1National Key Laboratory of Complex System Intelligent Control and Decision, School of Automation, Beijing Institute of Technology, Beijing
- 2Department of Cardiology Internal Medicine, Nanlou Branch of Chinese PLA General Hospital, Beijing
- 3Mathematical Institute, University of Oxford, Andrew Wiles Building, Radcliffe Observatory Quarter, Woodstock Road, Oxford, UK
The current expert system for the classification of electrocardiogram (ECG) doesn't have the ability of learning. Classification of ECG signals based on neural network cannot achieve accurate detection of ECG signals due to slow converge and uneasy approach of global optimal solution. In this study, a method of ECG signal classification based on fuzzy characteristics extraction and adaptive particle swarm optimization radial basis function (PSO-RBF) algorithm was proposed to solve the above two problems.
The ECG data of 1200 patients' records, which had 6 categories, such as normal sinus arrhythmia, sinus bradycardia, sinus tachycardia, sudden cardiac death, atrial fibrillation and congestive heart failure, were got from the PhysioBank database. Least mean square (LMS) adaptive filter was used to filter the noises of the ECG signals, such as power frequency interference, electromyography noise, electrode contact noise and respiratory amplitude interference. Then the wavelet multi-scale analysis was used to detect the P, QRS and T wave groups of the ECG signals and geometric parameters were obtained, such as pStart (P wave start point value), pEnd(P wave end point value), pTop(The highest value of P),etc. The geometric parameters for each of the wave groups were normalized and fuzzed as notched P wave high P wave, low P wave, and normal P wave, etc. The fuzzed characteristics of the every patient's ECG signals was divided into training group and validation group. After the training group respectively carried out network training by the traditional RBF algorithm and adaptive PSO-RBF algorithm, the trained network was used to classify the ECG signals of the validation group into 6 categories mentioned above. At last, the accuracy of each categories was verified.
The results of adaptive PSO-RBF algorithm showed that the discrimination rate of several typical cardiac rhythms were more than 90%, but the traditional RBF algorithm was less than 90%. For instance, the discrimination rate of normal sinus arrhythmia was 94.3%, sinus bradycardia 95.2%, sinus tachycardia was 95.7%, the discrimination rate of sudden cardiac death was 91.1%, the discrimination rate of atrial fibrillation was 92.7%, and the discrimination rate of congestive heart failure was 93.6%.
The use of fuzzy characteristics extraction method improves the robustness of the system. Compared with the traditional RBF algorithm, adaptive PSO-RBF neural network can be more accurate on the classification of ECG, and it's far superior to general expert systems which base on logical reasoning and traditional RBF neural network.