Deep source separation for single-channel fetal ECG extraction

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Fetal electrocardiogram (FECG) extraction from abdominal signals is crucial for fetal health monitoring.
• Maternal electrocardiogram (MECG) interference significantly challenges FECG isolation in abdominal electrocardiogram (AECG) signals.
• Existing methods struggle with the high amplitude of MECG, often overshadowing the critical FECG signal.

Purpose of the Study:

• To propose an attention-based generative adversarial network (AGAN) for robust FECG source separation from single-lead AECG.
• To address the amplitude bias issue in multi-objective learning for signal separation.
• To develop a novel framework for blind source separation applicable to biomedical signals.

Main Methods:

• An AGAN architecture combining GAN-style adversarial training and attention mechanisms was developed.
• The Hadamard product was innovatively employed as the learning objective to mitigate amplitude bias.
• The model was trained and evaluated on the ADFECGDB, B2_LABOUR, and PCDB datasets.

Main Results:

• The AGAN method effectively separated both MECG and FECG components simultaneously from single-lead AECG.
• Consistent high performance was achieved across multiple datasets, with SE, PPV, and F1 scores exceeding 94%.
• Specific scores included 96.67% SE, 97.13% PPV, and 96.90% F1 on ADFECGDB.

Conclusions:

• The proposed AGAN presents a robust and effective solution for FECG extraction from AECG signals.
• This study introduces an innovative data-driven framework for addressing blind source separation challenges.
• The method holds significant potential for advancing non-invasive fetal monitoring technologies.