Using Convolutional Neural Networks for Fetal Heart Sound Segmentation.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Medicine

Background:

• Phonocardiography (heart sound recording) processing is crucial for diagnosing heart disorders.
• Accurate segmentation of phonocardiogram (PCG) signals is essential for most diagnostic methods.
• Traditional methods like hidden semi-Markov models (HSMM) with logistic regression are common but can be improved.

Purpose of the Study:

• To develop and evaluate a U-net-based convolutional neural network (CNN) for segmenting fetal heart sounds from abdominal recordings.
• To compare the performance of the CNN model against established HSMM-based methods for heart sound segmentation.
• To assess the potential of deep learning for improving fetal phonocardiogram analysis.

Main Methods:

• A U-net CNN architecture was employed for heart sound segmentation.
• The model was initially trained on an open pediatric dataset and subsequently fine-tuned using fetal data.
• Performance was evaluated by comparing the CNN against two HSMM methods (a previously developed one and HSMM with logistic regression).

Main Results:

• The best U-net model achieved 92.2% PPV and 84.9% F1 for first heart sound detection, with a mean absolute error of 17.2±26.0 ms.
• For second heart sound detection, the U-net model achieved 88.1% PPV and 88.1% F1, with a mean absolute error of 17.9±12.7 ms.
• While first heart sound detection did not consistently outperform HSMM methods, second heart sound detection accuracy showed significant improvement.

Conclusions:

• Deep learning, specifically U-net CNNs, shows promise for enhancing fetal heart sound segmentation from abdominal recordings.
• The improved accuracy in second heart sound detection suggests potential for more refined diagnostic capabilities.
• Future research could combine specifically trained neural networks with HSMM temporal modeling for even greater accuracy in phonocardiogram analysis.