【摘要】 为提高阵发性房颤(paroxysmal atrial fibrillation, PAF)的识别能力，本研究提出一种基于卷积网络的多分辨率心电图(electrocardiogram, ECG)理解框架。该框架通过同时利用局部的高分辨率形态特征和全局的低分辨率节律特征，可始终保持高分辨率特征并不断引入低分辨率特征分支。通过不断整合各分支的特征，高分辨率分支可辨别P波形态变化，低分辨率分支可检测RR间期的节律变化，从而实现PAF定位、房颤分类和QRS波定位多个任务。本研究使用CPSC 2021-Train数据库训练模型，并使用两个临床ECG数据库测试。两个数据库上的PAF定位分数分别为1.818 2和3.487 0;房颤分类和QRS波定位在两个数据库上的F1分数均值分别为88.36%和99.47%。说明本研究方法具有良好的PAF端点和QRS波定位性能。更多还原

【Abstract】 In order to enhance paroxysmal atrial fibrillation(PAF) localization, we proposed a convolutional network-based multi-resolution ECG understanding framework. By harnessing both local high-resolution morphological features and global low-resolution rhythmic characteristics, the framework consistently maintained high-resolution features while progressively incorporated low-resolution feature branches. Through continuous integration of features from each branch, the high-resolution branch discriminated changes in P-wave morphology, while the low-resolution branch detected rhythmic alterations in RR intervals, thereby facilitated multiple tasks including PAF localization, AF classification, and QRS-wave localization.We trained the model on the CPSC 2021-Train database and conducted tests using two clinical ECG databases. The PAF localization scores on the two databases were 1.818 2 and 3.487 0, AF classification and QRS-wave localization achieved mean F1 scores of 88.36% and 99.47%, respectively. These results affirm the efficacy of our approach in PAF endpoints and QRS-wave localization.更多还原