Abstract:Monitoring the electrical pulse of fetal heart through a non-invasive fetal electrocardiogram (fECG) can easily detect abnormalities in the developing heart to significantly reduce the infant mortality rate and post-natal complications. Due to the overlapping of maternal and fetal R-peaks, the low amplitude of the fECG, systematic and ambient noises, typical signal extraction methods, such as adaptive filters, independent component analysis, empirical mode decomposition, etc., are unable to produce satisfactory fECG. While some techniques can produce accurate QRS waves, they often ignore other important aspects of the ECG. Our approach, which is based on 1D CycleGAN, can reconstruct the fECG signal from the mECG signal while maintaining the morphology due to extensive preprocessing and appropriate framework. The performance of our solution was evaluated by combining two available datasets from Physionet, "Abdominal and Direct Fetal ECG Database" and "Fetal electrocardiograms, direct and abdominal with reference heartbeat annotations", where it achieved an average PCC and Spectral-Correlation score of 88.4% and 89.4%, respectively. It detects the fQRS of the signal with accuracy, precision, recall and F1 score of 92.6%, 97.6%, 94.8% and 96.4%, respectively. It can also accurately produce the estimation of fetal heart rate and R-R interval with an error of 0.25% and 0.27%, respectively. The main contribution of our work is that, unlike similar studies, it can retain the morphology of the ECG signal with high fidelity. The accuracy of our solution for fetal heart rate and R-R interval length is comparable to existing state-of-the-art techniques. This makes it a highly effective tool for early diagnosis of fetal heart diseases and regular health checkups of the fetus.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to accurately separate the fetal electrocardiogram (fECG) signal from the mother's electrocardiogram (mECG) in non - invasive fetal electrocardiogram (fECG) extraction while maintaining the morphological characteristics of the signal. Specifically, because the electrocardiogram signals of the mother and the fetus overlap in time, and the amplitude of fECG is low, coupled with the influence of system and environmental noise, traditional signal processing methods such as adaptive filters, independent component analysis, and empirical mode decomposition cannot produce satisfactory fECG signals. In addition, although some techniques can generate accurate QRS waves, they often ignore other important parts in ECG, such as P - waves, ST - segments, and T - waves. To solve these problems, this research proposes a deep - learning technique based on 1D CycleGAN, aiming to reconstruct the fECG signal from the mECG signal while maintaining its morphological characteristics. Through extensive pre - processing and an appropriate design framework, this method can effectively extract the fetal electrocardiogram from the mother's electrocardiogram, providing an efficient tool for the early diagnosis of fetal heart diseases and routine fetal health examinations. ### Main contributions 1. **Morphology preservation**: Unlike existing research that mainly focuses on fQRS detection, this method can reconstruct the entire fECG signal, including information such as P - waves, S - waves, T - waves, and PR, ST, and QT intervals, which is crucial for diagnosing various fetal diseases. 2. **High - performance fQRS detection**: In terms of fQRS detection, the performance of this method is comparable to that of the existing state - of - the - art techniques. 3. **New weighted loss function**: A new weighted loss function is used, which significantly improves the quality of the generated fECG signal. 4. **Powerful pre - processing module**: The pre - processing module includes techniques for removing artifacts, such as filter lag, transient response, etc. 5. **High - precision fetal heart rate and heart rate variability measurement**: It can accurately measure fetal heart rate and heart rate variability indicators from the generated fECG signal. 6. **Diversity of data sets**: Two different real - world data sets are used, making the framework more robust and not affected by factors such as experimental settings, electrode positions, and recording devices. ### Method overview This research uses the 1D CycleGAN framework to achieve the mapping from mECG to fECG through two generators and two discriminators. The goal of the generators is to extract fECG from mECG, while the discriminators are used to evaluate the similarity between the generated fECG signal and the real fECG signal. To ensure that the generated fECG signal is not only close to the real signal in the frequency spectrum, but also consistent in time and power, the researchers introduced three additional loss terms: spectral loss, time loss, and power loss. These loss terms are combined with the traditional L1 loss to form a comprehensive adversarial loss function, thereby improving the quality of the generated signal. ### Data sets Two data sets are used in the research: 1. **Abdominal and Direct fECG (A&D fECG) Database**: It contains 4 - channel mECG signals of 5 pregnant women and fECG signals recorded from the fetal scalp. 2. **Fetal electrocardiograms, direct and abdominal with reference heartbeat annotations**: It is divided into two parts: delivery signals and pregnancy signals, which respectively contain 4 - channel mECG signals of 12 pregnant women and 10 pregnant women and the corresponding fECG signals. ### Pre - processing The pre - processing steps include resampling, baseline correction, filtering, segmentation, and normalization. Through these steps, the researchers can effectively remove noise and artifacts and retain the morphological characteristics of the signal. ### Model architecture 1. **CycleGAN architecture**: It includes two generators and two discriminators. The generators are responsible for extracting fECG from mECG, and the discriminators evaluate the authenticity of the generated signal. 2. **Loss function**: In addition to the traditional L1 loss, spectral loss, time loss, and power loss are also introduced to ensure that the generated fECG signal is close to the real signal in multiple dimensions. ### Training algorithm The researchers provide a general training algorithm to train CycleGAN through mini - batch gradient descent method, gradually optimizing the performance of generators and discriminators. In conclusion, this research through innovative deep - learning techniques and elaborate design...

A Novel Deep Learning Technique for Morphology Preserved Fetal ECG Extraction from Mother ECG using 1D-CycleGAN

Morphology Extraction of Fetal Electrocardiogram by Slow-Fast LSTM Network

An Efficient ECG Denoising Method by Fusing ECA-Net and CycleGAN.

Fetal ECG Extraction from Maternal ECG using Attention-based CycleGAN

Enhancing Fetal Electrocardiogram Signal Extraction Accuracy through a CycleGAN Utilizing Combined CNN–BiLSTM Architecture

Morphology extraction of fetal ECG using temporal CNN-based nonlinear adaptive noise cancelling

Extracting Fetal Heart Rate from Abdominal ECGs Based on Fast Multivariate Empirical Mode Decomposition

Integrating Contrastive Learning and Cycle Generative Adversarial Networks for Non-invasive Fetal ECG Extraction

A deep learning framework for noninvasive fetal ECG signal extraction

Correlation-Aware Attention CycleGAN for Accurate Fetal ECG Extraction

A novel fetal ecg signal extraction from maternal ecg signal using conditional generative adversarial networks (CGAN)

Recovery of the fetal electrocardiogram for morphological analysis from two trans-abdominal channels via optimal shrinkage

Diagnosis of fetal arrhythmia in echocardiography imaging using deep learning with cyclic loss

A Deep Learning Approach for the Assessment of Signal Quality of Non-Invasive Foetal Electrocardiography

A Novel Machine Learning-Based Prediction Method for Early Detection and Diagnosis of Congenital Heart Disease Using ECG Signal Processing

Single-lead Noninvasive Fetal ECG Extraction by Means of Combining Clustering and Principal Components Analysis

A Novel Deep Learning Approach for Early Detection of Cardiovascular Diseases from ECG Signals

Noninvasive Extraction of Maternal and Fetal ECG using Periodic Progressive FastICA Peel-off

A novel hybrid method for calculating the fetal heart rate from the non-invasive abdominal Electrocardiogram signal

rECGnition_v1.0: Arrhythmia detection using cardiologist-inspired multi-modal architecture incorporating demographic attributes in ECG

Fetal QRS Detection in Noninvasive Abdominal Electrocardiograms Using Principal Component Analysis and Discrete Wavelet Transforms with Signal Quality Estimation