TABLE 3.

Performance comparison between the proposed deep learning model in this study and previously tested algorithms on the PhysioNet Challenge 2017 dataset (Goldberger et al., 2000).

Rank	Year	Authors	Algorithm	F1 score (%)
=1	2020	This study	CNN and BiLSTM	82.6
=1	2017	Teijeiro et al. (2017)	Feature engineering and LSTM	83.1
=1	2017	Datta et al. (2017)	Feature engineering and AdaBoost	82.9
=1	2017	Zabihi et al. (2017)	Feature engineering and Random Forest	82.6
=1	2017	Hong et al. (2017)	Feature engineering and XGBoost	82.5

The database contains 8,527 single-lead electrocardiogram (ECG) segments with four categories, which are normal sinus (5,154), atrial fibrillation (AF; 771), other rhythms (2,557), and noise (46). Note that all algorithms are ranked #1 according to the Challenge evaluation measure, as the rounding value of all F₁ scores is 83%. The PhysioNet Challenge 2017 website provided the ranking. Note that the F₁ score of this study was achieved by the 10-fold cross-validation method. BiLSTM, bidirectional long short-term memory; CNN, convolution neural network; LSTM, long short-term memory. Note that the bold font highlights the performance of the proposed method (Method II).