基于自适应位置编码的心电图重构算法

Adaptive position encoding for ECG reconstruction

可穿戴心电检测的主要挑战是较多导联影响被测者的身体活动,如果减少导联会使心电数据信息减少使检测效果变差。为了平衡被测者日常穿戴舒适性和检测准确性,笔者设计了一个基于Transformer Encoder的自适应相对位置编码重构算法,通过前后重叠的切片方式使相邻片段的信息具备关联性,相对位置编码时加入可训练参数对任意片段进行重构,从而有效地提取位置信息。用3个导联的EGG信号重构标准12导联EGG信号,实验结果表明,重构的ECG数据均方根误差低至0.02758,平均相关系数高达98.43%,显示出本文算法在应用于可穿戴心电检测设备的应用前景。

The main challenge in wearable electrocardiogram(ECG)detection is that multiple leads for detection affects the subject's physical activity,but reducing leads will weaken detection performance due to the loss of ECG data.An adaptive position encoding based reconstruction algorithm using Transformer Encoder is proposed to balance wearing comfort and detection accuracy.The information of adjacent slices is correlated through slice overlapping.When encoding relative positions,trainable parameters are added to reconstruct any slice for effectively extracting position information.The standard 12-lead ECG signal is reconstructed from 3-lead ECG signals,and the experimental results show that the root mean square error and average correlation coefficient of the reconstructed ECG data are 0.02758 and 98.43%,indicating that the proposed algorithm has high application prospect in wearable ECG monitoring equipment.