摘要：

心肌梗死(MI)是心血管疾病(CVD)中常见的临床表现形式, 在发病时具有较高的致命性, 因此心肌梗死的快速定位对于避免死亡至关重要。目前基于心电图的心肌梗死位置定位模型在面对患者间的个体差异时泛化性能不足, 同时传统的基于卷积的模型难以深入挖掘心电图导联与心肌梗死位置之间的关系。为解决这些问题, 提出一种基于特征融合的多通道心肌梗死定位模型FF-ANN, 该模型主要由特征融合模块和自适应的多通道注意力模块组成。通过特征融合模块整合临床知识中的关键波型特征(例如Q波、ST段等), 使模型具有多种感受域, 从而在不同维度上捕捉心肌梗死的特征; 利用自适应的多通道注意力模块对融合后的特征进行重新标定, 通过注意力权重加权对应的特征, 使模型聚焦对预测有重要贡献的导联特征。通过在混合数据集PTB上验证模型的拟合能力, 并使用迁移学习的方法将从PTB数据集中学习到的模型架构迁移到PTBXL数据集中进行泛化能力验证, 结果表明, 与现有研究相比, 该模型在患者间方案下实现了约2.5%的提升, 证明了该模型不仅具有较好的定位性能, 也显示优越的泛化能力, 其架构适用于现实世界中辅助心肌梗死定位的诊断。

关键词: 心肌梗死, 心电图信号, 特征融合, 注意力机制, 定位

Abstract:

Myocardial Infarction (MI) is a common clinical manifestation of Cardiovascular Disease (CVD), which is highly fatal at the onset, making the rapid localization of MI crucial to avoid death. The current Electrocardiogram(ECG)-based MI location localization models have insufficient generalization performance owing to individual differences in patients, whereas traditional convolution-based models make it difficult to deeply explore the relationship between ECG leads and MI location. This study proposes a multi-channel MI localization model based on feature fusion (Feature Fusion-Attention Neural Network(FF-ANN)) to resolve these problems. This model mainly comprises a feature fusion module and a multichannel attention module. The feature fusion module integrates key waveform features (e.g., Q-wave and ST-segment) from clinical knowledge to ensure the model has multiple receptive domains to capture the features of MI at different scales. The multi-channel attention module is used to recalibrate the fused features, and the corresponding features are weighted by attention weights so that the model focuses on lead features that contribute significantly to prediction. This study validates the fitting ability of the model on the hybrid PTB dataset and uses a migration-learning approach to migrate the model architecture learned from the PTB to PTBXL datasets to validate its generalization ability. The model is improved by approximately 2.5% under the between-patient scenario compared to existing studies. The experimental results demonstrate that the model has better localization performance and shows superior generalization ability while demonstrating that its architecture is suitable for the real-world diagnosis of assisted MI localization.

Key words: Myocardial Infarction(MI), Electrocardiogram(ECG) signal, feature fusion, attention mechanism, position