| The incidence rate of cardiovascular disease continues to rise,which has become the primary threat to the health of urban and rural residents,and has evolved into a major livelihood issue.For cardiovascular disease,the key is "early detection,early treatment,and early prevention".Artificial intelligence is applied to computer aided diagnosis scenarios to achieve accurate and real-time monitoring,evaluation,early warning and intervention.ECG annotation is the basis of ECG(Electrocardiography)computer-aided diagnosis,mainly including heartbeat annotation and rhythm annotation,and the cost of heartbeat annotation is much higher than that of rhythm annotation.However,the existing heartbeat detection technology for heartbeat pre-annotation has problems of coarse classification granularity and insufficient training samples,which lead to poor model generalization ability.The atrial fibrillation detection technology used for rhythm pre-annotation has the problem of inadequate representation learning ability of DL(Deep Learning)for short sequence data.Therefore,this thesis proposes "intelligent annotation of ECG data based on human-machine integration",which focuses on the key technologies required for ECG intelligent annotation,and carries out research from improving the generalization ability and representation learning ability of the model,reducing the cost of heartbeat annotation,realizing the intelligent pre-annotation of heartbeat and rhythm,and realizing the intelligent ECG annotation system.The main work and innovations are as follows:(1)To solve the problems of coarse classification granularity and poor generalization ability of the existing beat detection models used for beat pre-annotation,a multi category beat pre-annotation algorithm based on the first stable convergence positioning(Fscp)is proposed from the perspective of data enhancement.In order to ensure the accuracy of the generated data,an generation data querying algorithm based on Fscp is proposed to solve the problem of convergence stability of generative adversarial network for data generation,and the existing generation data quality evaluation framework is improved to comprehensively evaluate and analyze the generation model and data.Finally,the generation model generated 14 types of heart beats,and the maximum mean discrepancy from the real heart beat was as low as 0.01.After the training dataset is enhanced and balanced by the generated data,the accuracy of 14 classifications is 99.28% on the pre-annotation model based on convolutional neural network,which is the highest performance in similar work.(2)In order to solve the problem of efficient quering of samples with high value to be annotated,a DAL(Deep Active Learning)heart beat pre-annotation algorithm based on weak hierarchical query strategy of morphological statistical features is proposed.This algorithm introduces active learning,builds a DAL model,and in order to solve the problem of partial over fitting of DAL model query strategy for ECG data,proposes a query strategy based on weak stratification of morphological statistical features suitable for ECG,and systematically compares and analyzes the performance of the proposed and classical strategies.The experimental results show that compared with the classical strategies,the proposed has the highest stability,not only the accuracy of 14 classification can reach 99%,but also can save 85.7% of the labeled samples.This work realizes representative sampling,which is of great significance to reduce the cost of annotation.(3)In order to solve the problem of ECG rhythm annotation,a pre-annotation algorithm for atrial fibrillation based on gradient set feature extraction is proposed.Due to the poor representation learning ability of deep learning on short sequence data,a feature extraction algorithm based on gradient set is proposed to solve the problem of insufficient representation ability of deep learning on short sequence data.The experiments on several databases show that the feature extraction method based on GDS does not need to detect heartbeat,has high robustness,and is universal to most classifiers.The detection accuracy is 93.5% and the recall is95.7% on the 2-second ECG segment.Compared with similar work,this algorithm has the highest overall performance with 99% accuracy,99.7% recall and 98.2% specificity on the10-second ECG segment.(4)ECG annotation needs not only professional annotation software but also professional ECG doctors.ECG annotation has high professional requirements,cumbersome processes,heavy workloads,and long cycles.The problem of "difficulty in annotation" directly affects the development of the field of ECG intelligent analysis.Therefore,this thesis designs and implements an intelligent annotation system of ECG data based on human-machine integration,which integrates ECG intelligent analysis technology into each stage of annotation work.The system annotation function is complete to meet all annotation needs,and the management function is full to support large-scale annotation work.At the same time,with the increased labeled data,the intelligent self-learning ability of the system also improves.The experimental results show that intelligent annotation can reduce the working time by 89.43% compared with a full manual pattern and the working time by 76.44% compared to a semiautomatic pattern. |
