TL;DR
This paper introduces a novel method for extracting fetal electrocardiograms (FECG) from abdominal signals by combining multiple adaptive filters. Their proposed "Combined Recursive Least Squares" (CRLS) filter significantly improves FECG extraction accuracy, especially in noisy environments, outperforming existing state-of-the-art methods.
What Problem Does This Paper Address?
Monitoring fetal health during pregnancy is crucial, and the fetal electrocardiogram (FECG) is a key diagnostic tool. However, extracting a clean FECG signal non-invasively from the mother's abdomen is a significant challenge. Abdominal signals are a complex mix of the weak FECG, the much stronger maternal electrocardiogram (MECG), respiratory noise, and electrical interference. The MECG's amplitude can be 2-10 times higher than the FECG, making it incredibly difficult to isolate the fetal signal. Existing methods often struggle to achieve high accuracy, particularly in low signal-to-noise ratio (SNR) scenarios.
Methods
The core innovation of this paper is the use of an affine combination of adaptive filters to extract FECG. Instead of relying on a single adaptive filter, they combine the strengths of multiple filters. Specifically, they explore combinations of Least Mean Square (LMS) and Recursive Least Squares (RLS) filters.
The paper proposes three main combinations:
- Combined Least Mean Square (CLMS): Combines two LMS filters with different step sizes (one fast, one slow) to balance convergence speed and stationary performance.
- Combined Recursive Least Squares (CRLS): Combines two RLS filters, each with a different forgetting factor to adapt to rapid changes and maintain performance in stationary situations.
- RLS-LMS: Combines an RLS filter (for faster convergence) with an LMS filter (for better tracking ability).
The "affine combination" means that the output of each filter is weighted and summed to produce the final output. The weights are adaptively adjusted to minimize the overall error. The authors found that CRLS, which combines two RLS filters, demonstrated the most promising results.
Datasets
The authors utilized a public dataset named "Cutaneous potential recordings of a pregnant woman" from DaIsy. This dataset contains eight single-channel recordings, with the first five channels being abdominal signals and the remaining three being thoracic signals. For their experiments, they primarily used the second abdominal channel as the input signal and the eighth thoracic channel as the reference signal.
Key Insights and Ideas
- The power of combination: Isolating weak signals like FECG from strong interference is inherently difficult for single filters. Combining multiple adaptive filters, each with different strengths, allows for a more robust and accurate extraction.
- RLS shines in non-stationary environments: The paper highlights that RLS filters are generally better suited for non-stationary signals like ECG compared to LMS, due to their ability to minimize weighted squared error sums. This makes them a strong candidate for FECG extraction.
- CRLS as the champion: The CRLS filter, by combining two RLS filters with different forgetting factors, proved to be the most effective in suppressing noise and extracting the FECG signal, especially in challenging low SNR conditions.
- SNR interpretation: The paper cleverly reinterprets the standard SNR metric. For FECG extraction, a lower SNR value is desirable, indicating that the filter has successfully suppressed the dominant MECG noise, leaving a cleaner FECG signal.
Key Results
- Superior noise suppression: The CRLS filter achieved the lowest SNR (most negative dB value), indicating superior suppression of maternal ECG and other noise sources compared to single filters (NLMS, RLS) and other combinations (CLMS, RLS-LMS).
- Improved performance metrics: When compared to the state-of-the-art MSF-ANC method on a challenging low-SNR channel (Channel 2), CRLS demonstrated significant improvements:
- Sensitivity: +3.58%
- Accuracy: +2.39%
- F1 Score: +1.36%
- Qualitative waveform analysis: Visual inspection of the extracted FECG waveforms showed that CRLS effectively eliminated the high-amplitude MECG components, resulting in a cleaner FECG signal that was more easily identifiable and periodic. Single filters, in contrast, often left residual MECG that obscured the FECG.
Limitations and Future Work
- Dataset limitations: The study used a single public dataset. While it provided valuable insights, testing on a more diverse range of datasets with varying signal characteristics and noise levels would further validate the CRLS approach.
- Computational complexity: While not explicitly detailed as a limitation, adaptive filters, especially RLS, can be computationally more intensive than simpler methods. Future work could explore optimizations for real-time implementation.
- Generalizability: The authors suggest that the affine combination algorithm is highly efficient for extracting FECG in high-noise scenarios and could be a promising direction for detecting weak signals in other complex environments beyond FECG. This opens avenues for applying the CRLS approach to other challenging signal processing tasks.
Summary updated