Noise Mitigation in ECG Signal Using Unidirectional and Bidirectional Long Short-Term Memory (LSTM) Neural Network

Abstract

Effective and proper analysis of Electrocardiogram (ECG) signal is important for accurate diagnoses of patient ECG. However,
ECG signal is frequently corrupted with different artifacts such as 50Hz powerline noise, baseline wander noise, and motion
artifact, etc. Consequently, these interferences that affect the ECG signal may leads to inaccurate diagnoses of patient ECG.
Several researches have been done to reduce these interferences using recurrent neural network and unidirectional long short term
memory (LSTM) neural network. However, recurrent neural network suffers from vanishing gradient problem and fails to learn
the long term dependencies of the sequence data. Although, the unidirectional LSTM neural network was able to solve the
vanishing gradient problem but then it fails to learn future data which is needed to improve the network performance. Hence, this
work presents a novel approach to de-noise ECG signals, utilizing unidirectional and bidirectional long-short term memory
(UBLSTM) neural network. This technique allows the networks to learn both past and future data. The network is trained using
ECG data collected from physionet challenge 2017. The validation results showed clearly that the proposed model has
successfully filtered the ECG signal corrupted with powerline noise and baseline wander noise with an improved signal to noise
ratio (SNR) of 22.8dB, 22.1dB and 20.2dB for all the three records ECG signals considered in this work.