Abstract: In perinatal period, non-invasive Fetal Electrocardiogram (FECG) extraction is currently challenged by masking of target signals by strong Maternal Electrocardiogram (MECG) and environmental noise, difficulty of existing methods in balancing long- and short-term feature interaction, high sensitivity to non-stationary noise, excessive video memory occupation in long-sequence processing, and limited deployment on portable devices. To achieve accurate reconstruction of key clinical morphological features of FECG including P-waves and T-waves, and provide reliable technical support for ultra-early warning of fetal intrauterine distress and prenatal screening of congenital cardiac abnormalities, a novel Dual-Branch Spectral-Adaptive Temporal FECG Extraction Architecture (DSAT-FECG) based on Mamba-UNETR（Mamba and UNEt with TRansformer) was proposed in this work. Traditional Transformer encoder was replaced by the Mamba model in this architecture, and synchronous separation and reconstruction of MECG and FECG were realized through a dual-branch design, with computational complexity of long-sequence processing effectively reduced. A dynamically adaptive spectral block DASB (Dynamically Adaptive Spectral Block) was innovatively designed and combined with an interactive convolutional block ICB (Interactive Convolutional Block), and balanced modeling of global and local temporal features of electrocardiographic signals was achieved via global-local collaborative processing in the frequency domain and adaptive threshold filtering. Meanwhile, a MECG feature elimination mechanism was embedded to actively strip strong interference of maternal signals on FECG during encoding stage. Comparative experiments were conducted on synthetic dataset FECGSYNDB (Fetal ECG Synthetic Database), public real-world datasets ADFECG (Abdominal and Direct Fetal ECG Database) and PCDB (PhysioNet/Computing in Cardiology Challenge Database 2013), as well as a clinical measured dataset, and performance comparison was performed with four existing state-of-the-art FECG extraction methods. Experimental results show that F1-scores of fetal QRS (Q wave, R wave, and S wave complex) complex detection on three types of datasets reach 99.88%, 98.90% and 99.02% respectively, with a maximum improvement of 13.94 percentage points compared with the second-best method; Structural Similarity Index (SSIM) of signal reconstruction reached 98.27%, Peak Signal-to-Noise Ratio (PSNR) reached 45.12 dB, Mean Squared Error (MSE) was as low as 0.0012, and Pearson correlation coefficient (PCC) was as high as 0.986. Accurate reconstruction of key clinical morphological features of FECG including P-waves and T-waves can be achieved with this architecture, and deployment threshold is greatly reduced by single-channel input design. Technical support can be provided for real-time maternal-fetal monitoring on portable devices, and requirements for non-invasive FECG extraction in complex clinical scenarios can be well adapted.

摘要： 针对围产期无创胎儿心电信号(FECG)提取中存在的较强母体心电信号(MECG)与环境噪声掩盖目标信号、现有方法难以兼顾长短期特征交互、对非平稳噪声敏感、长序列处理显存占用高且便携设备部署受限的问题，同时为实现胎儿心电P波、T波等关键临床形态特征的精准重建，为胎儿宫内窘迫超早期预警与先天性心脏异常产前筛查提供可靠技术支撑，提出一种基于 Mamba-UNETR（Mamba and UNEt with TRansformer）的双并行无创胎儿心电提取的双分支频谱自适应时序胎儿心电提取架构DSAT-FECG (Dual-branch Spectral-Adaptive Temporal FECG extraction architecture)。该架构采用 Mamba 模型替代传统 Transformer 编码器，通过双分支同步完成母体与胎儿心电信号的分离重建，降低长序列处理的计算复杂度；创新设计动态自适应频谱块DASB (Dynamically Adaptive Spectral Block)，结合交互式卷积模块ICB(Interactive Convolutional Block)，通过频域全局与局部协同处理与自适应阈值滤波，实现心电信号全局与局部时序特征的平衡建模；嵌入母体心电特征消除机制，在编码阶段主动剥离母体信号对胎儿心电的强干扰。在合成数据集 FECGSYNDB（Fetal ECG Synthetic Database）、公开真实数据集 ADFECG（Abdominal and Direct Fetal ECG Database）与PCDB（PhysioNet/Computing in Cardiology Challenge Database 2013）、临床实测数据集上开展对比实验，与现有的先进胎儿心电提取方法进行性能比对。实验结果表明，该架构在三类数据集上的胎儿 QRS（Q wave, R wave, and S wave complex）波群检测F1分数分别达 99.88%、98.90% 和 99.02%，较次优方法最高提升13.94个百分点；信号重建结构相似性指数（SSIM）达98.27%，峰值信噪比（PSNR）达45.12 dB，均方误差（MSE)低至0.0012，皮尔逊相关系数（PCC)高达 0.986。该架构可精准重建胎儿心电P波与T波等关键临床形态特征，单通道输入设计大幅降低了部署门槛，可为便携式设备实时母胎监护提供技术支撑，适配复杂临床场景下的胎儿心电无创提取需求。