Deep source separation for single-channel fetal ECG extraction

Objective: The fetal electrocardiogram (FECG) is critical for monitoring fetal health, however, its extraction remains technically challenging due to strong interference from the maternal electrocardiogram (MECG) in abdominal electrocardiogram (AECG). Therefore, an attention-based generative adversarial network (AGAN) is proposed for source separation of FECG from single-lead AECG signals. Approach: The AGAN architecture uniquely combines two powerful techniques: GAN-style adversarial training for high-quality data generation and attention-based focus mechanisms for intelligent feature selection, leading to superior target signal extraction from complex mixtures. The innovation of the proposed method lies in addressing the amplitude bias issue in multi-objective learning tasks. This work innovatively employs the Hadamard product as the learning objective for the model, preventing the model from favoring high-amplitude components (e.g., MECG) while neglecting low-amplitude yet critical features (e.g., FECG). Main results: Experimental results demonstrate that the proposed method can effectively and simultaneously separate both MECG and FECG components from single-lead AECG signals. When evaluated on the ADFECGDB, B2_LABOUR, and PCDB datasets, the proposed method demonstrated consistent performance, achieving the following SE, PPV, and F1 scores: 96.67%, 97.13%, and 96.90% on ADFECGDB; 95.90%, 96.56%, and 96.22% on B2_LABOUR; and 94.96%, 95.18%, and 95.06% on PCDB. Significance: This study presents a robust method for FECG extraction while simultaneously introducing an innovative data-driven framework for blind source separation problems.