Detection of Atrial Fibrillation from ECG Signal Using Efficient Feature Selection and Classification

Abstract

Atrial fibrillation (AF) is a life-threatening cardiac condition caused by inadequate blood flow, resulting in abnormal ECG records, blood clotting, and cardioembolic strokes. In recent years, physicians have been particularly concerned with early detection and diagnosis to overcome cardiogenic stroke. AF can be easily identified at the initial stages due to the development in computer-aided diagnosis. The performance of this method is affected by noise and the variations in pattern of the ECG, which leads to false diagnosis. Current signal processing and shallow machine learning (ML) approaches are severely limited in their ability to detect this condition accurately. Deep neural networks have been shown to be extremely effective at learning nonlinear patterns in a wide variety of problems, which include computer vision tasks. Deep learning models are computationally costly, non-explainable, and require a large quantity of data to discover characteristics. In contrast, ML approaches are explainable and require good feature extraction. In this manuscript, ML based supervised classification method is developed based on feature ensembling. ECG signals are preprocessed (mean subtraction followed by Butterworth filtering and computation of RR intervals) and subjected to feature extraction (by entropy-, wavelets-, & statistical-features). The variations due to AF are effectively captured and selective features are ensembled to perform classification by SVM and KNN. This method is experimented on five different databases (such as: PAF prediction Challenge, Long-Term AF, Intracardiac, AF termination Challenge, and MIT-BIH atrial fibrillation) and the classification performance is found to be the highest compared to the state of art. To evaluate the effectiveness of the proposed technique, AF-specific characteristics are retrieved from the ECG signal in the presence of artificially added noise and the features are fed to classifiers for classification. Performance of the proposed method is compared with the deep learning based approaches. The graphical abstract of the proposed atrial fibrillation detection method is presented. The overall accuracy of the proposed method was found to be 91.88\(\%\) and 91.99\(\%\) for wavelets-SVM and ensemble wavelet-SVM, respectively. This model attained 100\(\%\) accuracy for entropy and statistical features with SVM and KNN, respectively.