【作者】 刘磊；

【导师】 魏守水；

【作者基本信息】 山东大学 ， 生物医学工程， 2024， 硕士

【摘要】 心房颤动(Atrial Fibrillation.AF)表现为心房快速无序的收缩和舒张,是一种严重的心房电活动紊乱。阵发性房颤(Paroxysmal Atrial Fibrillation,PAF)作为AF的早期阶段,若不能得到及时治疗,容易演变为持续性房颤,从而加剧心衰和中风的风险。临床上,PAF的发生和自发终止的机制尚不明确,且长时动态心电图(Electrocardiography,ECG)数据量大,分析耗时,使得人工预测PAF尤为困难。针对这一难题,本文基于机器学习方法,提出了 PAF的发生及自发终止预测的算法,为临床上PAF的早期定性和治疗提供了参考。本文的主要研究内容如下:(1)多特征融合的PAF预测研究。一方面,提取了 ECG中时域、频域、时频域和形态学在内的32个人工特征;另一方面,构建了一维和二维深度学习网络,提取了 ECG和时频图中的32个深度特征。将所提特征进行特征融合,并使用四种方法进行了特征排序以筛选出重要性较高的特征。然后对重要特征进行相关性分析,剔除了冗余特征。通过讨论不同分类器和特征组合的预测效果,确定了最佳的实验方法。最终基于多种公开数据集进行了 PAF的预测实验,并采集了山东省立医院的临床数据进行了模型验证。(2)基于双路径网络的PAF自发终止预测研究。构建了一种双路径深度学习网络,两个路径分别提取ECG序列和时频图的特征。网络中使用了深度可分离卷积,并引入了多尺度卷积块来保证特征提取的全面性。然后将两个路径所提特征进行融合和筛选,使用支持向量机进行PAF自发终止的预测。此外,讨论了双路径和特征选择的必要性、时间尺度的影响以及波形预测错误的可能原因。该方法避免了人工特征提取的繁琐性和不全面性,实现了基于10s ECG信号的实时预测,并在多个数据集上表现出良好的泛化能力。(3)网络可解释性研究。针对深度学习的难解释性,使用反卷积网络、遮挡敏感度图和Grad-CAM算法,分别对网络中的ECG序列路径和时频图路径进行了可视化研究,从不同角度探索了 PAF自发终止预测的重要因子,这可能对临床的PAF预测工作具有重要的参考价值。更多还原

【Abstract】 Atrial Fibrillation(AF)is a serious disorder of the electrical activity of the atria,characterized by rapid and disorderly contraction and diastole of the atria.Paroxysmal Atrial Fibrillation(PAF),an early stage of AF,is prone to evolve into persistent AF if it is not treated in time,thereby exacerbating the risk of heart failure and stroke.In clinical practice,the mechanisms underlying the onset and spontaneous termination of PAF are unclear,and the large volume of data and time-consuming analysis of long-duration ambulatory electrocardiography(ECG)make manual prediction of PAF particularly difficult.To address this challenge,this study proposes algorithms for the prediction of the onset and spontaneous termination of PAF,which would provide a reference for the early characterization and treatment of PAF in the clinic.The main research contents of this work are as follows:(1)PAF prediction based on multi-feature fusion approach.On the one hand,32 artificial features including time domain,frequency domain,time-frequency domain and morphology in ECG were extracted;on the other hand,1D and 2D deep learning networks were constructed to extract 32 deep features in ECG and time-frequency maps.The extracted features were subjected to feature fusion,and feature ranking was performed using four methods to filter out the features with higher importance.The important features were then analyzed for correlation and redundant features were eliminated.The best experimental method was determined by discussing the prediction effects of different classifiers and feature combinations.Finally,the prediction experiments of PAF were conducted based on a variety of publicly available datasets,and clinical data from Shandong Provincial Hospital were collected for model validation.(2)The prediction of PAF spontaneous termination based on dual-path network.A two-path deep learning network was constructed with two paths to extract features from ECG sequences and time-frequency maps,respectively.Deeply separable convolution was used in the network and multi-scale convolutional blocks were introduced to ensure the comprehensiveness of feature extraction.The features extracted from the two paths are then fused and filtered for PAF spontaneous termination prediction using Support Vector Machine.In addition,the necessity of dual path and feature selection,the effect of time scale,and the possible causes of waveform prediction errors were discussed.The method avoids the tedious and incomplete nature of manual feature extraction,achieves real-time prediction based on 10s ECG signals,and demonstrates good generalization ability over multiple datasets.(3)Network interpretability study.To address the difficult interpretation problem of deep learning,the ECG sequence paths and time-frequency map paths in the network were visualized using the inverse convolutional network,the occlusion sensitivity map and the Grad-CAM algorithm,respectively.And the important factors for the prediction of spontaneous termination of PAF were explored from different perspectives,which may be an important reference for the PAF prediction in the clinic.更多还原