Combination Hand-Crafted Features and Semi-Supervised Features Selection From Deep Features for Atrial Fibrillation Detection

Atrial fibrillation (AF) is the one of most prevalent type of heart disorder, and early diagnosis plays a crucial role in reducing the risk of cardiac failure and stroke. However, detecting AF is often problematic as it is frequently asymptomatic or presents occasional symptoms. Currently, there are hand-held Electroencephalogram (ECG) devices, such as smartwatches, that could be helpful in detecting AF. Single-lead ECG devices are being widely recognized for their potential in monitoring patients’ heart health globally. Studies have shown that regular monitoring of standardized ECG signals can considerably reduce mortality rates associated with severe cardiovascular diseases. The researchers aimed to utilize a machine learning approach for detecting AF from short ECG signals. However, the current automated detection method for AF has scope for enhancement. This study proposed a novel feature set that combines deep features extracted from cyclic correlation of ECG signals with manually designed features for predicting ECG rhythms. The effectiveness of this method was evaluated using the PhysioNet Challenge 2017 dataset, which consists of numerous ECG recordings. The performance of the proposed method was superior to state-of- art techniques employed on the same dataset and assessed using the same metric. The proposed method classified ECG recordings into four different classes and obtained an average F-score of 90.21% based on 10fold cross-validations.