Fetal R-peak detection using a swin transformer network with dynamic encoding and parallel decoding.

Area of Science:

• Biomedical Engineering
• Cardiovascular Technology
• Artificial Intelligence in Healthcare

Background:

• Fetal heart rate (FHR) monitoring is crucial for assessing fetal well-being.
• Accurate FHR detection from abdominal electrocardiogram (fECG) signals is challenging due to noise and interference.
• Existing methods often struggle with signal quality and require complex preprocessing.

Purpose of the Study:

• To develop an advanced deep learning model for robust FHR detection from abdominal ECG data.
• To improve the accuracy and reliability of FHR estimation compared to existing techniques.
• To create a foundation for enhanced maternal and fetal monitoring systems.

Main Methods:

• A dynamic encoding and parallel decoding Swin transformer-based network (DPSTFR-Net) was designed for FHR detection.
• A cascaded Sparse Low-Rank and Kernel Recursive Least Squares (CSKL) filter was employed for signal preprocessing to remove noise.
• Adaptive position encoding and parallel decoder contextual information were utilized to enhance classification accuracy.

Main Results:

• DPSTFR-Net achieved high performance on benchmark datasets (PCDB and ADFECGDB).
• Accuracy reached up to 97.52%, precision up to 98.25%, and recall up to 97.35%.
• The method demonstrated effective noise reduction and reliable FHR estimation from abdominal ECG.

Conclusions:

• The proposed DPSTFR-Net effectively estimates FHR from abdominal ECG signals, outperforming conventional methods.
• The network's ability to handle noisy signals makes it suitable for real-world applications.
• This technology holds potential for commercialization in long-term maternal and fetal monitoring systems.