Robust detection of atrial fibrillation from short-term electrocardiogram using convolutional neural networks
Siti Nurmaini,Alexander Edo Tondas,Annisa Darmawahyuni,Muhammad Naufal Rachmatullah,Radiyati Umi Partan,Firdaus Firdaus,Bambang Tutuko,Ferlita Pratiwi,Andre Herviant Juliano,Rahmi Khoirani,Alexander Edo Tondas
DOI: https://doi.org/10.1016/j.future.2020.07.021
IF: 7.307
2020-12-01
Future Generation Computer Systems
Abstract:The most prevalent arrhythmia observed in clinical practice is atrial fibrillation (AF). AF is associated with an irregular heartbeat pattern and a lack of a distinct P-waves signal. A low-cost method for identifying this condition is the use of a single-lead electrocardiogram (ECG) as the gold standard for AF diagnosis, after annotation by experts. However, manual interpretation of these signals may be subjective and susceptible to inter-observer variabilities because many non-AF rhythms exhibit irregular RR-intervals and lack P-waves similar to AF. Furthermore, the acquired surface ECG signal is always contaminated by noise. Hence, highly accurate and robust detection of AF using short-term, single-lead ECG is valuable but challenging. To improve the existing model, this paper proposes a simple algorithm of a discrete wavelet transform (DWT) coupled with one-dimensional convolutional neural networks (1D-CNNs) to classify three classes: Normal Sinus Rhythm (NSR), AF and non-AF (NAF). The experiment was conducted with a combination of three public datasets and one dataset from an Indonesian hospital. The robustness of the proposed model was evaluated based on several validation data with an unseen pattern from 4 datasets. The results indicated that 1D-CNNs outperformed other approaches and achieved satisfactory performances with high generalization ability. The accuracy, sensitivity, specificity, precision, and F1-Score for two classes were 99.98%, 99.91%, 99.91%, 99.99%, and 99.95%, respectively. For the three classes, the accuracy, sensitivity, specificity, precision, and F1-Score was 99.17%, 98.90%, 99.17%, 96.74%, and 97.48%, respectively. Potentially, our approach can aid AF diagnosis in clinics and patient self-monitoring to improve early detection and effective treatment of AF.