Leads-Adaptive Fetal Electrocardiogram Extraction Using Attention-Based BiLSTM.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Medicine

Background:

• Extracting fetal electrocardiogram (fECG) from abdominal electrocardiogram (AECG) is crucial for assessing fetal cardiac rhythm.
• Clinical AECG recordings often contain signal anomalies, hindering traditional fECG extraction methods.
• Robust fECG extraction is needed for reliable fetal monitoring.

Purpose of the Study:

• To propose a deep learning-based method for adaptive fECG signal extraction from multi-lead AECG.
• To enhance the accuracy and resilience of fECG extraction in the presence of signal anomalies and channel defects.
• To improve clinical applicability of fECG extraction for fetal cardiac rhythm assessment.

Main Methods:

• A deep learning model utilizing a bidirectional long short-term memory (BiLSTM) architecture.
• Integration of a deep supervision subnetwork and an attention mechanism module.
• The attention module calculates inter-channel relevance weights for adaptive feature fusion and mitigation of defective channels.

Main Results:

• The proposed model effectively extracts fECG signals under various channel defect conditions, validated on public datasets.
• Ablation studies confirmed the attention module's crucial role in improving resilience to channel anomalies.
• Signal masking experiments corroborated the reliability of the extracted fECG signals.

Conclusions:

• The developed deep learning method offers an effective solution for adaptive fECG extraction from multi-lead AECG.
• The attention mechanism significantly enhances the model's robustness against signal anomalies and channel defects.
• This approach holds promise for clinical deployment in fetal cardiac rhythm assessment.