A novel approach for congestive heart failure and arrhythmia classification using OD- 1D-LBP with LSTM and 1D-CNN

Abstract

Congestive heart failure (CHF) and heart rhythm disorders (ARR) are known to be the most important heart diseases of the last decades, which have had negative effects on human health directly or indirectly. The early diagnosis of these types of heart diseases and rapid and practical medical intervention is very vital. For the diagnosis of such diseases, it is so difficult for doctors to analyze long ECG signals quickly and detect instantaneous parameter changes in the ECG signal with manual observation technique. Therefore, it is necessary to develop efficient Computer Aided Diagnosis (CAD) systems for the early diagnosis of diseases such as ARR and CHF. For this purpose, within the scope of digital signal processing, various statistical and mathematical algorithms have been developed to extract features from signals and various machine learning methods are applied for classification. In this study, a new method is proposed to analyze and classify the ECG signals of CHF, ARR ,and NSR (Normal Sinus Rhythm). This method is based on the local-binary-pattern (LBP) algorithm and is called Orthogonal Difference One Dimensional Local Binary Pattern (OD-1D-LBP). This method is an approach that uses binary information obtained by comparing each point on the signal with its neighbors. Histograms of new signals that are obtained from the proposed method (OD-1D-LBP) are given as input to Long Short-Term Memory (LSTM) and one-dimensional Convolutional Neural Networks (1D-CNN) for classification. If 70% of the input data is applied as training data and 30% as test data, the accuracy rates are obtained at 98.63% for LSTM and 98.86% for 1D-CNN. If only ACF and ARR data are classified; the accuracy rate is obtained at 98.94% for LSTM; and 97.40% for 1D-CNN. Similarly, when ACF and NSR ddataare classified; the accuracy rate is obtained at 99.4% for LSTM; and 98.9% for 1D-CNN. In case ARR and NSR data are classified; the accuracy rates are obtained at 99.2% for LSTM; and 95.2% for 1D-CNN.