Classification Of ST Segment In ECG Signals Based On Deep Learning Approach

Electrocardiogram(ECG)is a measure of the heart electrical signals.The analysis and diagnosis of ECG signal are of great significance in the detection of human heart health.In ECG,ST-segment and T-wave changes(ST-T changes)are extremely common.ST-T changes can easily induce ventricular hypertrophy and acute myocardial infarction,which threaten human health seriously.Mostly,ECG classification studies focus on the overall classification of arrhythmia.There are few studies on ST-T disease,also the recognition rate is low.This work has wide research space.Therefore,this article used signal preprocessing and deep learning methods as the technological base and took heartbeat classification as an entry point to improve the ST-T segment recognition rate.Our work based on the MIT-BIH arrhythmia benchmark database,combined with intra-patient and inter-patient clinical evaluation paradigm,and compared the classification performance of diverse model inputs and network frameworks.We found that the original signal as the input,convolutional neural network as the framework can achieve high-accuracy recognition performance.In addition,the introduction of an attention mechanism in the model can improve the recognition rate of abnormal heartbeats which can be aimed at the research of certain specific heart diseases.Referred the above-mentioned research results,this paper designed an ECG record classification model based on clinical ST-T diagnostic rules.Different from the general attention mechanism in the deep learning model,the innovation of this model was proposing an attention mechanism that combined ECG signal characteristics and diagnostic rules,then made some special adjustments.Through the communication with cardiologists,it was founded that when diagnosing diseases such as "ST-T abnormality",doctors tended to focus mainly on the S wave and T wave areas in the ECG.Corresponding to the clinical diagnosis rule,before model training,we manually extracted the waveform annotations in each ECG record,partitioned the entire record,then made different adjustments for different areas.This adjustment was reflected in the enhancement and reductions of the amplitude of the ST region and the non-ST region on the ECG.After experiments,the best clinical parameters were finally obtained.The accuracy rate of the original group(not using the diagnostic rule attention mechanism)was increased from 87.2% to 89.1%,and the recall rate was increased from84% to 91%.The F1-score of the ST-T has been increased from 84% to 86%,and other evaluation indexes have also been improved,which fully proved the effectiveness of this method.Besides,this paper also designed a comparison group with the general attention mechanism.The proposed method is superior to the general attention mechanism in the evaluation indicators.Since the designed classification model is a light network structure based on a convolutional network,with low complexity,high performance,less training time,it is suitable for transplantation on wearable devices or embedding ECG acquisition systems,which is convenient for patient’s out-of-hospital treatment.