Research On Knowledge-Guided Classification Algorithm For ECG Data

ECG data plays a major role in the diagnosis of cardiovascular diseases,and its early identification,correct diagnosis and timely treatment are essential to reduce mortality from cardiovascular diseases.The use of computer-aided diagnosis is of great research value as it can reduce physician workload and improve the efficiency of monitoring and detection of cardiovascular diseases.Remote wearable devices allow for continuous monitoring of heart activity,enabling better quality of care for cardiac patients.The large amount of ECG data generated by these devices provides the data base for classification and identification using deep learning techniques.Excellent results have been achieved in the research of ECG classification based on deep learning.Deep learning transforms raw ECG data into higher-level and more abstract representations through nonlinear transformations between multiple hidden layers to achieve disease identification and classification.However,the performance of deep learning is related to the number of parameters,which usually requires large computational and storage resources,and the black-box nature of deep learning poses an obstacle to understanding the model decision process.To explore classification methods to improve the performance and reliability of deep learning based on the domain knowledge of ECG classification for the complexity and causal uncertainty of deep learning models.Specifically,the main research content and work can be summarized as follows.(1)To address the problems of complex convolutional neural network scale and large number of parameters,KecNet,a classification model based on Sinc convolutional layer,is designed.The ability of band-pass filters to effectively separate information in ECG data is used as a priori knowledge,combined with the ability of convolutional neural networks to automatically learn features,to design physically interpretable Sinc-convolutional layers.This convolutional layer learns custom filters parameterized by high and low cutoff frequencies to implement adaptive band-pass filtering.Experiments were performed on the MIT-BIH Atrial Fibrillation database.Experiments were performed on the MIT-BIH Atrial Fibrillation database.The results show that the prior knowledge improves the feature extraction capability and robustness of the whole network,and the structure can achieve more stable and efficient filtering even in noisy environments,while only requiring a smaller number of parameters.(2)To further improve the recognition of AF by KecNet,one of the gold standards for clinical diagnosis of AF:RR interval irregularity was used as a priori knowledge to symbolically represent the temporal characteristics of ECG data.The coefficients of variation reflecting the degree of RR interval dispersion are extracted and added to the network as parameters to improve the ability of the convolutional neural network to represent the long-time correlation features of ECG data.The experimental results show that the RR interval coefficient of variation extracted for the study objectives improves the classification performance of the model by 1.5%-2%.(3)When performing multiple classifications of arrhythmias,it is difficult to generalize the complex pattern of ECG data by a single feature.To combine multidimensional manual features and depth features,WDNet,a classification model based on Wide&Deep structure,is designed.The model consists of a jointly trained generalized linear model and a depth model.The linear model is used to record the combination of domain features with high frequency in the sample to achieve the memory capability of the model.Meanwhile,abstract patterns are extracted from the raw ECG data using deep networks to increase the feature extraction capability of the model.A Softmax classifier is used to integrate two different types of features for classification.Experiments on the MIT-BIH Arrhythmia database showed that WDNet improved classification accuracy by 2-3%and F1 scores by 10-13%compared to a single linear model or convolutional network.Meanwhile,during the validation of different feature subsets,it is found that QRS-related features have the greatest impact on model performance.