Research And Application Of Convolution And Self-Attention Mechanism Based Multi-lead ECG Classification

Heart disease is a major cause of thre at to human health,and ECG is the most commonly used clinical screening tool for cardiovascular disease.With the booming development of artificial intelligence technology,computer-aided ECG classification and diagnosis technology can greatly reduce the workload of physicians and alleviate the problem of medical resource constraints.However,existing methods do not fully utilize the rich lead information of the most common multi-lead ECG data in clinical practice,and thus the recall of some diseases is low.Meanwhile,for the ECG multi-label classification task,most studies have only conducted experiments on small-scale public datasets,lacking the validation of a large amount of clinical data,and the robustness is difficult to guarantee.To solve the problems of existing work,this thesis conducts algorithmic research and application development on the problem of feature representation and multi-label classification of multi-lead ECG,proposes a learning method of ECG representation and multi-label classification model based on convolution and self-attention mechanism,designs and develops an ECG intelligent diagnosis system for clinical needs,and verifies the effectiveness of the system on a large-scale clinical data set.The work in this thesis specifically includes:(1)Convolution-based Supervised Contrastive Learning of ECG Representations.In this thesis,we investigate the ECG feature representation,exploit the ECG signal characteristics,and propose a Convolution-based Supervised Contrastive Learning of ECG Representation Method(CSCL-ECGR).Aiming at the downstream task of ECG classification,CSCL-ECGR uses the supervised contrastive learning method to pretrain the ECG data so that the label information directly guides the pre-training process of the convolutional attention encoder to improve the category differentiability of the ECG representation and help the downstream classifier to establish clear decision boundaries in the feature space.In the fine-tuning stage,CSCL-ECGR exploits the focal loss function to mine difficult sample information and alleviate the data imbalance problem.In this thesis,comparative experiments on a clinical dataset for ECG binary classification task show that the class differentiability of ECG representations obtained by supervised contrastive learning is significantly improved over unsupervised contrast learning pre-training,and the classification performance of CSCL-ECGR surpasses that of unsupervised contrastive learning and directly supervised trained models,respectively.(2)Lead-Aware Hierarchical Transformer and Convolution Fusion Network.To solve the multi-lead ECG multi-label classification problem,this thesis proposes a Lead-aware Hierarchical Transformer and Convolution Fusion Network(LHTC-Net).For the characteristics of ECG signals which are long time sequences,the hierarchical Transformer network built based on the self-attentive mechanism can capture the longrange dependent features across multiple heart beats,while the attentional convolutional network uses convolution to introduce the inductive bias of the translational invariance of heart rhythm patterns and focuses on the local waveform features through the CBAM to finally fuse the global and local features to improve the comprehensive classification performance.To address the characteristics of some rhythm types with lead specificity,this thesis proposes a lead-aware mechanism,which uses lead-independent segmentation and window-based self-attention mechanism to achieve adaptive extraction of key information of specific leads and improve the classification performance of rhythm types with lead characteristics.Full experiments on public and clinical datasets show that the performance of LHTC-Net outperforms all five baseline methods,and the adaptive extraction of lead-specific key information by the lead-aware mechanism is demonstrated visually through case studies.(3)ECG intelligent diagnosis system.Based on clinical requirements,in this thesis we develop and deploy an ECG intelligent diagnosis system,which integrates CSCL-ECGR and LHTC-Net models to build an ECG comprehensive diagnosis model,and supports remote diagnosis service for GPU inference server devices and local diagnosis service for PC devices,providing a platform for the algorithm model to create application value.The ECG comprehensive diagnosis model is trained on large-scale clinical data containing about 477,000 data,and supports diagnostic services for 30 common heart rhythm types.The accuracy test results show that the detection rate of 16 rhythm types exceeds 85%,among which the detection rate of 11 rhythm types exceeds 90%,which verifies its high recall rate accuracy and can meet the actual clinical needs.The stress test results of remote diagnosis service on the server and the performance test results of local diagnosis service on the PC ensure the availability of the online service of ECG intelligent diagnosis software.