| Electrocardiograms are widely used to measure the health of the human heart.However,both standard 12-lead ECG and 24-hour ambulatory ECG rely on manual analysis by medical professionals,which is time-consuming and labor-intensive.Due to time constraints,abnormal states of the patient’s heart are highly likely to be undetectable because they are not in the onset phase.With the development of deep learning wearable intelligent ECG monitoring technology,there has been much work to achieve continuous monitoring and automatic diagnosis of the heart in daily health monitoring scenarios.Most of the research involves complex feature extraction of ECG data and subsequent classifier design to discriminate multiple types of arrhythmic signals.However,the heart is a complex organ,and many different and new types of arrhythmias may occur at any time,which are not part of the original training set.Therefore,implementing ECG signal abnormality detection in health monitoring scenarios is a more effective and practically feasible essential technique.This paper mainly considers two aspects to propose improvements to the ECG signal abnormality detection method.Firstly,we build a user-specific feature learning model based on the specificity features exhibited by different individual patient data.Secondly,to effectively screen for ECG abnormalities that never appear in training sets in real health monitoring scenarios.This paper applies contrast learning-related techniques to ECG signal abnormality detection and proposes two anomaly detection models..The main research contents and innovations of this paper includes:First,given the problem that significant individual differences in ECG signals generally affect the accuracy of computer-aided diagnosis,a user-specific ECG abnormality detection model is proposed.The model uses a contrastive learning technique,which integrates individual data-specific attention into a probability-based noise-contrastive estimation loss function,and uses individual user-specific data features as "hidden" labels to help calculate the model’s prediction loss.Moreover,the triplet loss function is integrated for a specific downstream ECG signal abnormality detection task.The model learns to construct the embedding space during the training process,maximizing the similarity between normal ECG samples as much as possible and minimizing that between the normal ECG samples.The similarity between samples and abnormal samples improves the accuracy of anomaly detection.In this chapter,statistical analysis is carried out on the publicly available dataset MIT-BIH arrhythmia dataset to perform functional validation of the model and compare the experimental results of different research methods.With only a small amount of labeled data in the test set,the abnormality detection method in this chapter achieves a considerable improvement in accuracy,precision,and sensitivity,which can reach 95.26%,97.42%,and 92.23%,respectively.Second,current anomaly detection algorithms based on closed-set recognition lead to a model that cannot cope well with the ECG signal anomalies that may never appear in training sets.We propose an open-set recognition-oriented ECG signal anomaly detection model.First,the method refers to supervised contrast learning,using the label information to set normal ECG pairs as similar,aiming to make the neural network model construct tightly distributed clusters of normal samples in the embedding space.Any abnormal types are projected outside the tight clusters of normal samples.By establishing a threshold on the similarity,we can capture the open space of abnormal classes and improve the effectiveness of abnormality detection.Finally,experiments and analyses on the PTB-XL dataset show that its accuracy can reach more than 97%.In the ablation experiments,the performance is substantially improved under different openness settings by comparing the proposed method with the supervised endto-end model.It has shown good reliability and effectiveness in anomaly detection. |
PDF Full Text Request