Arrhythmia Classification And Signal Timescale Study Based On Deep Learning

Arrhythmia is one of the common cardiovascular diseases and poses a serious threat to human health.An electrocardiogram is a standard tool for recording cardiac activity and is widely used to measure the health of the heart.At present,most studies use heart beats to automatically classify arrhythmias,but when using classification based on heartbeats,problems such as RR and PP intervals and compensatory intervals are ignored,resulting in classification errors that prevent these key information from being used.There are also large errors in splitting and intercepting heart shots and annotating by experts.With the combination of deep learning algorithm models and clinical medical data,the application research is becoming more and more mature,and the role of deep learning is increasing.This thesis uses the MIT-BIH dataset as the data base,and uses the heart beat and time slice as the research unit to study arrhythmia,and the main research is the following two aspects:In the classification task of arrhythmia,an algorithm combining Atrous Convolution aggregation and attention Resnet is proposed.Different from ordinary multi-scale convolution operations,the introduction of cavity aggregation module can improve the accuracy of arrhythmia classification while reducing the number of parameters.The introduction of attention residual mechanism allows the model to focus on the ECG information important for arrhythmia classification,avoids the gradient degradation of the model,and shows better antinoise performance.The use of single heartbeat classification to study arrhythmias ignores the information between heartbeats,so we use time fragments as units to discuss the time scale problem in arrhythmia research,and give two ways to label time segments.Considering the particularity of time fragment research,a bidirectional gated recurrent unit(BiGRU)that can process time series information is added to the network model,which can not only pay attention to the ECG information before the target signal,but also take into account the ECG information after the target signal.The advantages and disadvantages of the two labeling methods are discussed,a time scale suitable for arrhythmia research is given,and feature map visualization is performed.