Power-Line Interference Suppression in Electrocardiogram Using Recurrent Neural Networks

Suppression of the Power-Line Interference (PLI) in electrocardiogram (ECG) has been the key to accurately diagnosing the heart condition. A novel method to suppress PLI is presented to preprocess ECG signals. The extraction of PLI is accomplished by a trained model based on Recurrent Neural Networks (RNN), with Gated Recurrent Unit (GRU). The RNN-based model is trained to adapt the amplitude and phase of a 50-Hz sinusoidal signal to approximate the PLI in the ECG signals. The ECG signals are then filtered by subtracting the extracted PLI signals. The proposed method could significantly relieve the distortion of the ECG signals, especially in the scope of QRS complex. Detection of the existence of PLI is unnecessary because the extracted signals from PLI-free ECG signals by the trained model are negligible. Experiments are conducted on both synthetic and real ECG signals. The proposed method was evaluated by comparing with a state-of-the-art filtering method based on Kalman smoother and a traditional IIR notch filter. The results show that our proposed method improves the output SNR value by 10.5% in comparison with the referenced Kalman smoother method, and more than 20.9% and 16.3% reduction of the settling time of step increase and decrease response has been achieved, respectively. The results show that the model has enough generalization ability for unseen signals without retraining. The sequentially input and output architecture of the model makes it suitable for real-time applications and hardware implementation.