ECG Signal De-noising And T Wave Detection Based On Deep Learning

Sudden cardiac death(SCD)has become a highest rank in the field of cardiovascular problems with high morbidity,high mortality rate and sudden onset.In the background of telemedicine,a portable ambulatory electrocardiogram(ECG)monitor has become an effective method for prediction of sudden cardiac death.In the dynamic ECG,Q-T interval and the shape is often accompanied by abnormal changes,so the identification and analysis of the Q-T are crucial for the early prediction of sudden cardiac death.However,T wave signal is weak,varied morphology and is susceptible to noise interference,ECG denoising and the T wave detection have become the greatest problems in portable ambulatory ECG monitor.In this paper,considering the differences of human,the amount of the noise,and the advantages of big data characteristics of ECG signal in the telemedicine,we introduce the deep learning to denoise the ECG and achieve the T wave automatic detection.The main works are as follows:(1)The ECG signal can be overlapped by a part of noise in the frequency domain.That makes the frequency-based methods nearly incapable of noise reduction.To solve this problem,the paper proposed an ECG denoising algorithm based on deep neural network constructed by auto-encoder.Using stacked denoising auto-encoder can extract sophisticated characteristics of input signal.Using the capacity of denoising auto-encoder that it can extract the signal robustness characteristics,and finish the task to reconstruct the original signal from the noised-signal.Building training data based on the similarity of the heart beats and adjusting the network parameters make it easier to structure the neural network achieving ECG signal noise reduction.(2)For the situation of some residual noise with a jagged morphology in a part of denoised signals,the paper proposed wavelet and contractive denoising auto-encoder to optimize noise reduction model.The contractive denoising auto-encoder can punish and inhibit larger change in the hidden layer by increasing the Frobenius norm of the Jacobian.The adaptive thresholding wavelet transform can filter out the noise that spectrum distribution is known,thus reduces the network layer nodes and simplify the computational complexity of the algorithm.(3)In the existing algorithm,T wave morphology detection and feature points detection interact with each other.Under the premise of T wave morphology is known,it will improve the accuracy of T wave feature points detection.However,it could not determine the T wave morphology,if the information of T wave feature points is unknown.In order to solve the contradiction,the automatic detection algorithm of T wave is proposed based on morphology guidance.The morphological characters of T wave can be extracted by sparse auto-encoder,such as single upright peak,single inverted peak,bidirectional peaks(negative,positive)and bidirectional peaks(positive,negative).Distorted gaussian function is used for mathematical modeling based on morphological characters of T wave.The feature points of T wave can be detected by the analysis of template and the correlation of T wave.In this paper,the proposed ECG denoising method and T wave detection method have been used in the remote intelligent ECG monitoring platform which is carried out by our project team.ECG signals collected form ECG monitoring platform have been denoised by the proposed method.The experiment results show that the denoising algorithm in this paper can filter complex noise out,meanwhile keep the main feature wave of ECG signals.Moreover,the feature points of T wave that collected form ECG monitoring platform can be marked automatically based on the automatic detection algorithm of T wave.The research has greatly improved the intelligence of ECG monitoring system under the remote medical environment.