Development of Hybrid Deep Learning Models for Robust ECG Signal Denoising under Real time Noise Conditions

Abstract

Increasing heart diseases are a major concern in the healthcare domain, especially in the cardiac department. To predict cardiovascular disease, electrocardiogram (ECG) signals are commonly used. The disease prediction is conducted using the pattern of ECG signals. However, the ECG signal output is biased with a noise signal produced from electrode motion, electromagnetic interference, muscular artifacts, and baseline wander. The noise signals adversely impact the ECG signal and lead to wrong predictions, which badly affect the patient's life. In this regard, ECG signal denoising techniques are important for better prediction of cardiovascular diseases. In this research, ECG signal denoising is performed using deep learning (DL) models. Also, developed the hybrid DL model for enhancing the effectiveness of ECG signal denoising. In this study, the MIT-BIH Atrial Fibrillation data set is used to generate real-time probable noise signals. The general types of noise signals produced in these circumstances are Gaussian Noise, baseline wander, electrode motion artefacts, muscle artefacts, and electromagnetic interference, which are created and tested with the hybrid deep learning models. There are four deep learning models, such as Denoising Autoencoder (DAE), Long Short-Term Memory (LSTM), hybrid DAE-LSTM and CNN-BiLSTM models, which have been developed. In order to enhance the denoising performance, the addition of temporal sequence learning and spatial feature extraction is inculcated. The developed DL models underwent detailed testing with a generated noise model and were evaluated for performance accordingly.  After the testing, it was found that the hybrid DAE-LSTM model outperforms other DL models in terms of the performance parameters such as signal-to-noise ratio, peak signal-to-noise ratio, and Pearson correlation coefficient. The results show that the signal-to-noise ratio of the hybrid DAE-LSTM model is 10.42 dB, PSNR of 22.54 dB, and PCC of 0.9408. As compared to all DL models, DAE shows the least value of MSE of 0.1099. These results indicate that ECG signals are completely recovering from the noise, and a better prediction of cardiovascular disease is possible.  Also, the hybrid DL model has a wide scope in the biomedical sector for the reliable and accurate prediction of cardiovascular diseases by referring to the ECG signal pattern.  The proposed approach is reliable and an Artificial Intelligence (AI) enabled health care approach. Apart from this, the proposed hybrid DL model can be used in smart wearable cardiac monitoring devices for improving the accuracy.