Research On Detection Methods Of Record-level And Beat-level Electrocardiogram Abnormalities

Cardiovascular diseases pose a serious threat to human health and cause a heavy economic burden to the society.At present,the prevention and treatment of cardiovascular diseases in China is facing many challenges,such as the continuous increase of incidence rate and the insufficient supply of medical resources.To tackle these challenges,remote heart health monitoring technology based on wearable electrocardiogram(ECG)hardware and intelligent diagnosis algorithms may provide an effective and economical solution,and is expected to play an important role in the future medical and health service system.ECG signal can be used to detect a variety of cardiac abnormal states.As different abnormal states may have different occur timing and durations,it thus requires algorithms to detect abnormalities at different time granularities.In the heart health monitoring scenario,the massive incoming monitoring data far exceeds the processing capacity of manual verification,so the requirement for the detection accuracy is stricter.But the performance of current algorithms is limited by a series of factors,such as noise interference,the subjective deviation of human experts in annotations and the lack of annotated samples.Therefore,this thesis studies new detection algorithms for ECG abnormalities at record level(coarse-grained)and beat level(fine-grained),in order to further improve the detection performance on a series of key tasks.The main research contents are as follows:(1)In order to provide supports for heartbeat localization and feature extraction,problems in ECG wave detection are studied,and a QRS complex detection method considering the probability distribution of RR intervals(reflecting the time interval between adjacent beats)is proposed.Different from previous detection methods that pay too much attention to the ECG waveforms,this study constructs a dynamic Bayesian network to incorporate the ECG waveform factors and the heart rate factors(reflected by RR interval)into a unified probability model,and then applies it to the reasoning of QRS-complex positions,so as to improve the robustness of the algorithm to noise.In addition,to adapt the model to the individual differences of patients,an unsupervised parameter optimization method based on expecta tion maximization is designed.To improve the reasoning efficiency of the model,several simplification strategies are introduced into the model to make the reasoning have linear time complexity.The experimental results show that the proposed method has better performance than other methods on datasets with strong noise interference,and shows strong robustness in noise stress test.(2)In order to detect atrial fibrillation(AF)based on single lead ECG signal with poor signal quality,a single-lead ECG record classification method based on clustering feature is proposed.Different from previous methods that directly extract the features of beat waveforms and RR intervals from an ECG record,this study first clusters the beat waveforms and RR intervals of the record to obtain representative and high-quality beat waveforms and RR intervals,and then extracts the features from the representative waveforms and RR intervals.In the feature extraction,this study also considers the temporal distribution characteristics of beat waveform s and RR intervals.The extracted features not only reflect the variation of waveforms and RR intervals with time,but also reflect the interaction between waveforms and RR intervals.Based on the extracted features and XGBoost algorithm,a single-lead ECG classifier is constructed to distinguish between four classes: normal sinus rhythm,AF,other rhythm,and too noisy signal.The verification results on the dataset collected from portable ECG equipment show that the algorithm achieves state-of-the-art performance,and the clustering features have a significant contribution to the improvement of model performance.(3)In order to detect various abnormalities that may exist simultaneously in an ECG record,the multi-label classification method for ECG records is further studied.As the sample size of ECG data collected by a single institution(single-source data)is usually limited,ECG data from different institutions(multi-source data)are used for model training.Different from the previous methods that ignore the problem of label incompleteness,this study proposes a category-hierarchy-based label completion method and a weak supervised learning method for incomplete labels to tackle the problem of label incompleteness of multi-source ECG data.On this basis,a deep-learning-based multi-label classification model for ECG records is constructed,where the category attention mechanism is designed to extract the relevant features for each category.In addition,considering different misclassifications may have different actual costs,a cost-sensitive thresholding method is designed to minimize the classification cost of the model and reduce the impact of category imbalance of the training data during the process of transforming the prediction probabilities into labels.The cross-dataset evaluation results show that the proposed label completion method has substantial repair effect on real datasets,and the proposed mechanisms play a significant role in improving the model performance in multi-label ECG classification.(4)The above classification methods in recording level can detect the abnormal conditions existing in the recording,but they can not determine the abnormal type on a specific heartbeat,so they are not suitable to detect beat-level abnormalities such as ectopic beats.Therefore,the method of beat-by-beat ectopic beat detection is further studied.Unlike previous methods that rely on supervis ed learning on small-scale finely-labeled datasets,this study constructs a beat-by-beat ectopic beat detection model based on weak supervised learning,which can use large amounts of coarsely-labeled ECG data to improve the generalization ability for the detection.By introducing the mapping relationship from local prediction to overall prediction into the deep neural network,the overall annotations of ECG records can be used to guide the model to learn the predictions for local beats.In the aspect of feature extraction,the priori-knowledge-based heartrate features and deep-learning-based waveform features are combined to further improve the recognition ability of the model for abnormal beats.In addition,to deal with the ill-posed problem in weakly supervised learning,a two-stage training strategy combining supervised pretraining and weakly-supervised optimization is proposed.The experimental results on multiple datasets show that the proposed model with weak supervised learning on large and coarsely-annotated datasets has much better performance than other methods with supervised learning on small and finely-annotated datasets in the detection of ectopic beats.