Inter-patient detection of atrial fibrillation in short ECG segments based on LSTM network with multiple input layers

Atrial fibrillation (AF) is the most common type of cardiac arrhythmia. The electrocardiogram (EGG) signal is commonly used for AF diagnosis. AF detection methods based on ECG accelerate the work of the cardiologist when annotating an examination. This paper proposes a deep network based on the long short-term memory (LSTM) architecture using single and multiple inputs for AF detection from ECG signals. The single-input network uses a 10-s pre-processed ECG signal segment at the input, whereas the multiple-input network in a deep layer has an additional vector with four features computed from the RR intervals. We employed the inter-patient paradigm in our experiments, and various databases were used to train and test our model, namely, the Long-Term AF Database (LTAFDB), MIT-BIH AF Database (AFDB), and MIT-BIH Normal Sinus Rhythm Database (NSRDB). The multiple-input network obtained the best results in all datasets due to the addition of an original interval RR-based feature vector. The performance in terms of sensitivity, specificity, and accuracy with a 95% confidence interval was 94.62 ± 2.01%, 95.19 ± 1.18%, and 94.83 ± 1.92%, respectively, on the LTAFDB, and 91.53 ± 5.41%, 91.08 ± 4.68%, and 90.87 ± 5.10%, respectively, on the AFDB. In the NSRDB, the specificity was 98.52 ± 0.98%. The use of multiple inputs based on the ECG signal and a feature vector extracted from the RR intervals outperformed the single-input network in the detection of AF episodes. The combination of the original input data and network architecture was demonstrated to be a valuable tool for AF detection in short ECG segments.