Development of atrial fibrillation detection algorithm based on extracted ECG features and hybrid multilayer perceptron network

Heart is made up by bundles of cardiac muscle. It has a pacemaker that generate electrical signal to escalate muscle for contraction and relax. However, when electrical signal undergoes internal or external disturbance, it may lead to cardiac abnormality. Cardiac abnormality refers to the capricious electrical activity of the cardiac muscles. It sometimes does not exhibit any symptom as it's still in the early stage or phase one, but which may lead to sudden end due to the heart cease functioning. In order to reduce the sudden death episode, a lot of research has been done to give early warning to the patient. In this research, a new approach has been developed which capable to detect atrial fibrillation (AF) activity based on electrocardiogram (ECG) signal using the Hybrid Multilayer Perceptron (HMLP) neural network. An intelligent system is designed to solve the problem at the earliest stage. A dataset of ECG signal is taken from the MIT-BIH database used in the research to train and generalize the HMLP network, as well as to test the network performance. Continuous ECG signal dataset needs to be segmented into a complex contains P, QRS and T waves, since information of cardiac abnormality depends on those waves. In this research, the R to R peak interval (RRI) is used to segment the ECG signal in a complex form. At this stage, rectangular pulses are overlapping with a complex ECG then the interception points are taken as ficidual point of the complex. The ficidual point contains the amplitude and duration at each intercept point. Furthermore, the amplitudes and durations are considered as the input vector to HMLP network. The HMLP network is trained by a number of training algorithms which are Levenberg Marquardt (LM), Bayesian Regularization (BR), Back- Propagation (BP) and Resilient (R). Then, it will come out with the prediction performance, including accuracy, sensitivity and specificity performances. The performances are divided into training and testing before overall performance is taken as the average (after 5 or more iterations). In the research, the performance of the classifier is measured by the mean square error (MSE) and regression (Reg) at each iteration. The lowest MSE performance and most approaching to 1 of regression performance may improve accuracy of the network. The Graphical User Interface (GUI) is designed to represent a whole research project with prediction performance of 98.61% accuracy.