Efficient ECG Analysis with High F1 Score and Low Computation Complexity

Electrocardiograms (ECG) are important signals of the human body to analyze its state, monitor health, and even save a life when early deviations appear. Official statistics reveal that one of the major death causes is heart disease and its consequences. In this paper, the new approach to the detection of Atrial Fibrillation (AF) Arrhythmia in real-time is developed and presented. This type of cardiac arrhythmia affects many people. If it remains undetected, it could develop into chronic disability or even could result in early mortality. Also, it is hard to diagnose this disease because of its non-regular nature, especially during the early stages. Early detection is an important issue because it can help to save and extend human life. Real-time non-invasive and effective approaches to detect AF are being developed, but their properties still could be improved. The performance of such models and methods depends on the data quality and feature engineering largely. ECG as a heart electrical activity record is used to monitor heart activity and to detect a broad range of heart disabilities. Our research uses the dataset of PhysioNet Computing in Cardiology Challenge 2017. It consists of 8528 single-lead ECG records of heart rhythms. The proposed method and appropriate model outperform the best known State-of-the-Art results in terms of
$F$
1
score, which is the main metric of the Challenge results mentioned above. Also, the presented approach is more computationally effective and less energy-consuming. The latter allows an embedded implementation of this model.