Fetal ECG Extraction on Time-Frequency Domain Using Conditional GAN

Abstract

Fetal electrocardiogram (ECG) analysis plays an important role in assessing fetal heart rate, rhythm, and detecting potential cardiac anomalies. However, obtaining fetal ECG remains challenging due to its inherently low amplitude and susceptibility to maternal signal interference. In this paper, we propose a novel approach for robust fetal ECG extraction using a Conditional Generative Adversarial Network (cGAN) that operate on time-frequency domain of abdominal ECG signals instead of raw 1D abdominal ECG signals. By utilizing the frequency domain, our model is able to capture intricate time-varying patterns often obscured by noise and interference in the time domain. Moreover, cGAN leverages prior knowledge about fECG signal structures, enhancing the accuracy of fECG reconstruction. Experimentations on real-world ECG dataset validates the efficacy of our model in accurately extracting fECG signals, achieving high structural similarity score (SSIM) and low mean squared error (MSE) when compared with corresponding ground truth test sets. The approach shows superiority over conventional methods, demonstrating robustness to noise and interference. All in all, this work presents a promising avenue for advancing non-invasive fECG extraction techniques and its potential applications in clinical settings.