AF Detection by Exploiting the Spectral and Temporal Characteristics of ECG Signals with the LSTM Model

This research reinvestigates the detection of atrial fibrillation (AF) from a recurrent neural network (RNN) viewpoint. In particular, a long short-term memory (LSTM) model of RNN is designed to exploit the high-order spectral and temporal features of the multi-lead electrocardiogram (ECG) signals of patients with AF. To verify the proposed method, the LSTM model is tested with ECG data available from the PhysioNet and some normal ECG data collected in our labs. The results show that not only the deviation of the so-called RR intervals of ECG signals but also its temporal variations are critical to AF detection. The accuracy of AF detection can reach up to 98.3 %, with an LSTM model of using 30 hidden units. Considering more realistic applications, we further tested the model with subjects different from that of the training data. The accuracy is about 87% with high sensitivity. The experimental results show that the proposed model is able to effectively extract both the long-term and short-term characteristics of the spectral content of the AF ECG signals, making it a good candidate model for AF detection.