南方医科大学学报

南方医科大学学报 ›› 2022, Vol. 42 ›› Issue (11): 1672-1680.doi: 10.12122/j.issn.1673-4254.2022.11.11

• • 上一篇    下一篇

胎儿心电信号的无创提取：基于时间卷积编解码网络

曹 石，巩 高，肖 慧，方威扬，阙与清，陈超敏   

  1. 南方医科大学生物医学工程学院，广东 广州 510515；南昌大学第一附属医院，江西 南昌 330006
  • 出版日期:2022-11-20 发布日期:2022-11-30

Fetal ECG extraction using temporal convolutional encoder-decoder network

CAO Shi, GONG Gao, XIAO Hui, FANG Weiyang, QUE Yuqing, CHEN Chaomin   

  1. College of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China; First Affiliated Hospital of Nanchang University, Nanchang 330006, China
  • Online:2022-11-20 Published:2022-11-30

RichHTML

30

PDF

238

摘要： 目的 实现从孕妇腹壁混合心电信号中提取微弱的胎儿心电信号，为准确估计胎儿心率、分析胎儿心电波形等提供基础。方法 利用深度卷积网络（deep CNN）优越的非线性映射能力，本文提出了一种基于时间卷积编解码网络的非线性自适应噪声消除（nonlinear ANC）提取框架，以实现胎儿心电信号的有效提取。首先构建适用于处理胎儿心电信号的深度时间卷积网络（TCED-Net）模型作为非线性映射工具；然后以孕妇胸部心电信号为参考，利用该模型估计孕妇腹壁混合心电信号中的母体心电成分；最后从腹壁混合信号中减去所估计的母体心电成分，以得到完整的胎儿心电信号。实验利用合成心电数据（FECGSYNDB）和临床心电数据（NIFECGDB、PCDB）对方法性能进行测试与对比。结果 本文方法在FECGSYNDB上的胎儿R峰检测精度（[F1]值）、均方误差（MSE）和质量信噪比（qSNR）分别达到98.89 %，0.20和7.84；在NIFECGDB上的[F1]值达到99.1%；在 PCDB 上的[F1]值达到 98.61%。在不同数据集中较之 EKF（[F1=]93.84%）、ES-RNN([F1] =97.20% )和 AECG-DecompNet（[F1]=95.43%）等现有性能最佳的算法，本文方法的R峰检测精度指标分别高出5.05%、1.9%和3.18%，均优于现有最佳方法。结论 与现有算法相比，本文方法可以提取出更为清晰的胎儿心电信号，对孕期进行有效的胎儿健康监护具有一定的应用价值。

关键词: 胎儿监护；自适应滤波；时间卷积神经网络；无创胎儿心电图；胎心率信号

Abstract: Objective To extract weak fetal ECG signals from mixed ECG signals recorded from maternal abdominal wall for accurate analysis of fetal heart rate and fetal ECG patterns. Methods By exploiting the superior nonlinear mapping ability of deep convolutional network, we developed a nonlinear adaptive noise cancelling (nonlinear ANC) extraction framework based on a temporal convolutional encoder-decoder network for extracting fetal ECG signals. We first constructed a deep temporal convolutional network (TCED-Net) model for fetal ECG signal extraction, and with the maternal chest ECG signal as the reference signal, the maternal ECG component in the abdominal mixed signal was estimated using this model. The estimated maternal ECG component was subtracted from the mixed abdominal ECG signals to obtain the fetal ECG component. Experimental analyses were performed using synthetic ECG signals (FECGSYNDB) and clinical ECG signals (NIFECGDB, PCDB) to test the performance of the propose method. Results The results of experiments on the FECGSYNDB dataset showed that the proposed approach achieved good performance in F1-score (98.89%), mean-square-error (MSE; 0.20) and quality signal-to-noise ratio (qSNR; 7.84). The F1- score reached 99.1% on the NIFECGDB dataset and 98.61% on the PCDB dataset. The R peak detection accuracy index of the proposed method was higher than the existing best-performing algorithms such as EKF (F1=93.84%), ES-RNN (F1=97.20%) and AECG-DecompNet (F1=95.43%) by 5.05%, 1.9% and 3.18%, respectively. Conclusion The fetal ECG signals extracted using the proposed method are clearer than those by the existing algorithms, suggesting the potential value this method for effective fetal health monitoring during pregnancy.

Key words: fetal monitoring; adaptive filtering; temporal convolutional neural networks; non-invasive fetal electrocardiogram; fetal heart rate