A10 - Enhancing Fetal ECG Extraction with Multi-Scale Feature Fusion and Spatiotemporal Attention for Early CHD Detection

Introduction: : Congenital heart disease (CHD) remains a leading cause of neonatal morbidity and mortality, underscoring the critical need for early and accurate screening to improve clinical outcomes. Maternal abdominal electrocardiography (aECG), as a non-invasive monitoring technique, offers significant promise for fetal cardiac assessment. However, the extraction of fetal ECG (fECG) signals is particularly challenging due to their low amplitude and susceptibility to interference from strong maternal ECG (mECG) signals and ambient noise, limiting the effectiveness of traditional extraction methods. To address these challenges, we propose an innovative deep learning framework based on a graph-structured multi-scale Transformer. This approach is designed to efficiently and accurately extract fetal cardiac features from high-noise, multi-channel maternal aECG recordings, enabling the generation of high-fidelity fECG signals. The resulting framework provides a robust foundation for early risk screening and timely intervention in cases of fetal congenital heart disease.

Materials and

Methods: : The core of the proposed model is a novel deep learning framework that seamlessly integrates Graph Neural Networks (GNNs) with an advanced Transformer architecture. GNNs are employed to structurally model multi-lead abdominal ECG (aECG) signals, effectively capturing dynamic physiological correlations and spatial relationships among leads through sophisticated graph construction strategies. Building upon this foundation, we introduce a Multi-scale Maternal Temporal Feature Extraction (MSTFE) module, which leverages multi-scale causal convolutions and gated units to robustly capture the intricate temporal dependencies present in complex maternal signals.

At the heart of the framework is a Transformer-based encoder-decoder architecture, responsible for deep temporal encoding and autoregressive generation of the fetal ECG (fECG) signal. The decoder incorporates a guided cross-attention mechanism that explicitly decodes and reconstructs the fECG from the encoded maternal physiological states, further enhanced by multi-scale temporal convolutions to refine temporal attention. This innovative design enables highly accurate, interpretable, and physiologically faithful generation of fECG signals, representing a significant advancement over conventional approaches.

Results, Conclusions, and Discussions:: Results
The proposed graph-structured multi-scale Transformer model demonstrates superior performance in extracting fetal ECG (fECG) signals from complex maternal abdominal ECG data. Experimental findings indicate that the model effectively mitigates high noise interference and maternal signal contamination, consistently generating high-fidelity fECG waveforms with strong physiological realism. Performance benchmarks reveal that the extracted fECG surpasses or matches traditional and deep learning baselines on key metrics, including signal-to-noise ratio and morphological similarity (R², RMSE) to ground-truth fECG. Interpretability analyses, such as attention and feature activation maps, further validate the model’s mechanism for isolating critical signal components and detecting weak fetal signals.

Conclusions
This work presents an innovative graph-structured Transformer architecture that integrates multi-scale feature representations and guided attention for accurate fetal ECG extraction and early congenital heart disease (CHD) risk screening. By modeling spatial inter-lead dependencies with GNNs, capturing complex maternal signal dynamics through a multi-scale temporal module, and leveraging guided cross-attention and probabilistic outputs, the framework significantly improves the accuracy, physiological fidelity, and reliability of fECG extraction. The approach effectively addresses the limitations of conventional techniques in handling noise and signal overlap, offering a robust, non-invasive tool for early CHD screening and improved clinical management of high-risk pregnancies.

Discussions
The core innovation of this study is the synergistic integration of Graph Neural Networks and Transformer-based attention mechanisms to overcome the distinct challenges of fECG extraction. Unlike prior methods that treat multi-lead aECG signals as independent channels, this framework captures synergistic and differential information among leads, enabling a more comprehensive understanding of signal origins. The guided cross-attention design allows for intelligent isolation of weak fetal signals, while enhancing model interpretability. The multi-scale feature extraction module further improves the capture of complex ECG morphologies. The model also demonstrates potential for multi-modal data integration and robust CHD risk assessment. Future work will focus on large-scale clinical validation and extending the approach to other biomedical signal processing domains with similar complexity.

Acknowledgements and/or References (Optional)::