Foetal ECG recovery using dynamic neural networks.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Maternal-Fetal Medicine

Background:

• Non-invasive fetal electrocardiography (ECG) is crucial for monitoring fetal well-being during pregnancy.
• Existing methods face challenges like limited data, lack of performance metrics, and underutilization of adaptive techniques.
• Automatic fetal ECG recovery is a key application requiring robust methodologies.

Purpose of the Study:

• To address limitations in fetal ECG recovery by generating synthetic data and employing advanced adaptive methods.
• To develop and evaluate a model selection strategy using numerical and statistical measures.
• To enhance fetal ECG recovery accuracy using non-linear adaptive techniques, specifically neural networks.

Main Methods:

• Generation of synthetic fetal ECG registers and analysis of noise influence.
• Implementation of a model selection method using correlation coefficient and ANOVA.
• Integration of Finite Impulse Response (FIR) and gamma neural networks into Adaptive Noise Cancellation (ANC) schemes.
• Benchmarking neural networks against Least Mean Squares (LMS) and Normalized LMS (NLMS) algorithms.

Main Results:

• Fetal-maternal signal-to-noise ratio (SNR) is critical for model identification in synthetic data.
• Neural networks significantly outperform LMS-based algorithms when electromyogram interference is high.
• ANOVA tests reveal statistical differences between neural and LMS models in complex SNR scenarios.
• FIR neural networks provide the optimal balance between model complexity and recovery performance.

Conclusions:

• Advanced neural network models, particularly FIR, offer superior fetal ECG recovery compared to traditional adaptive filters.
• The proposed methodology for model selection and the use of synthetic data improve the robustness of fetal ECG analysis.
• This work contributes improved techniques for non-invasive fetal monitoring and well-being assessment.