Robust Fetal ECG Extraction From Abdominal Surface Potentials Using a Variational Autoencoder

The accurate extraction of fetal electrocardiogram (fECG) signals from noninvasive maternal abdominal recordings is essential for reliable and continuous fetal health monitoring. However, due to the dominance of maternal signals and the presence of various noise sources, this task remains technically challenging. In this study, we propose a deep learning-based approach that leverages a variational autoencoder (VAE) to achieve robust separation of fetal components and accurate detection of QRS complexes. This capability is crucial for estimating fetal heart rate and detecting potential abnormalities. The proposed method was evaluated on two publicly available datasets: the PhysioNet Challenge 2013 dataset and the Non-Invasive Fetal ECG Database (NIFECGDB). To validate its effectiveness, we compared the VAE-based approach against several baseline methods, including CycleGAN, W-NET, R2W-NET, and a combination of principal component analysis with discrete wavelet transform (PCA+DWT). Evaluation metrics included F1 score, precision, recall, and accuracy. Our method achieved superior performance, with F1 scores of 98.69% on the PhysioNet dataset and 99.61% on NIFECGDB. These results demonstrate the potential of the proposed VAE-based approach for practical use in noninvasive fetal monitoring systems, offering advantages in accuracy, robustness, and generalizability across different recording conditions.