中国医科大学硕士学位论文
III
Abstract
Objective: With the increasing incidence of cardiovascular diseases year by year, it is
urgent to prevent and diagnose cardiovascular diseases. It is very inefficient and
challenging to only rely on doctors to diagnose ECG signals. Therefore, home portable
monitors are becoming more and more popular. Based on the fact that the portable monitor
can detect and classify ECG abnormalities in the home, this paper proposes a convolutional
autoencoder algorithm with skip connection and a multi-scale convolutional classification
network based on attention mechanism. The autoencoder uses the relationship between
reconstruction error and threshold to make abnormal judgment on ECG data types that
have not been learned by the model, so as to inform the user of the occurrence of abnormal
conditions; and for the types of ECG signals learned by the model, the classification model
will point out the specific types of ECG beats, so that patients can know their illness at
home. When the ECG anomaly detection and classification model is applied to the family
portable ECG monitor, the patients can find the condition and go to the hospital in time,
which can reduce the incidence of heart disease and provide clinical diagnosis basis for
doctors.
Methods: The study is divided into two parts: anomaly detection and classification. The
anomaly detection part adds skip connection on the basis of convolution autoencoder, and
it uses the relationship between the error between input data and reconstructed data and
the threshold to detect abnormal ECG beat. If the reconstruction error is greater than the
threshold, it is determined as abnormal ECG beat. After eliminating abnormal ECG beats,
the ECG data learned by the anomaly detection network will then enter the multi-scale
convolution classification network with mixed attention mechanism to realize the
classification of four ECG beats. In this paper, the MIT-BIH arrhythmia database is
preprocessed by wavelet denoising, cardiac beat interception and normalization. Normal
class (N), right bundle branch block (R), left bundle branch block (L) and atrial premature
beat (A) are selected as the training set for model training. Ventricular premature beat (V)
is regarded as abnormal ECG data and detected in the test stage; The four types of ECG