Research On Automatic Atrial Fibrillation Detection Based On Fusion Of Domain Knowledge And Multi-view Features

Atrial fibrillation(AF)is a serious and life-threatening tachyarrhythmia that increases the risk of stroke and heart failure.Early detection can improve the effectiveness of treatment and prevent serious complications.However,the manual diagnosis of ECG signals requires high professional knowledge and experience,and intermittent atrial fibrillation requires long-term monitoring and identification.Subjective judgment can easily lead to misdiagnosis and missed diagnosis.Therefore,it is of great significance to develop an automatic diagnosis system.The traditional atrial fibrillation detection algorithm using electrocardiogram(ECG)relies on manual feature extraction,which has certain limitations.Therefore,in recent years,the automatic detection of arrhythmia based on deep learning technology has become an important research problem.However,the current automatic diagnosis model methods usually have problems such as the inability to fully and efficiently express multi-domain features,the lack of classification accuracy,the lack of clinical knowledge guidance,and the lack of interpretability.In this paper,a new deep learning network-based atrial fibrillation detection algorithm is proposed,which improves the detection accuracy and verifies the reliability of the algorithm in a series of subsequent experiments.The main contents of the paper include the following two points:(1)Aiming at the detection and classification of ECG signals of onset atrial fibrillation,this paper proposes a knowledge-guided multi-channel feature-enhanced atrial fibrillation detection model.It combines clinical medical knowledge and diagnostic criteria to extract time,shape and rhythm features from ECG signals for feature representation.By emphasizing the separate channel processing of atrial activity frequency and heart rhythm characteristics,the model is guided to focus on key parts of the ECG data.The attention mechanism fuses the three channel features to obtain a more accurate and stable decision model.Apply clinical knowledge in detection algorithms to improve confidence in results and interpretability of models.(2)Aiming at the detection of paroxysmal atrial fibrillation that is difficult to distinguish in traditional clinical medicine,this paper proposes a wide and deep multi-feature fusion network framework based on Transformer.It combines the one-dimensional and two-dimensional features of the signal,and the complementary fusion of deep features and shallow handcrafted features to improve the accuracy and comprehensiveness of signal representation.The proposal and application of this method improves the accuracy of the detection and classification of atrial fibrillation population,reduces the missed diagnosis caused by paroxysmal seizures,and can intervene and treat patients earlier,and reduce the adverse consequences caused by atrial fibrillation attacks.This paper proposes two deep learning models to solve the detection and classification of patients with paroxysmal atrial fibrillation,and compares multiple groups based on the Physio Net Computing in Cardiology Challenge 2017 dataset,the MIT-BIH AFDB database and other public datasets Experiment to train and validate the model.The final test results show that the two algorithms have good performance,which are significantly improved compared with most of the algorithms currently used.