Extracting fetal heart signals from Doppler using semi-supervised convolutional neural networks

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Fetal Monitoring Technologies

Background:

• Cardiotocography (CTG) is vital for fetal well-being assessment, requiring clear Doppler ultrasound (DUS) signals.
• Current fetal heart rate (FHR) algorithms struggle with signal differentiation, leading to monitoring gaps.
• Limited data availability hinders the application of artificial intelligence (AI) for DUS signal classification.

Purpose of the Study:

• To evaluate semi-supervised learning's effectiveness in improving DUS signal classification accuracy.
• To address the challenge of limited data in developing AI models for fetal monitoring.
• To enhance the reliability and accuracy of fetal well-being assessments using AI.

Main Methods:

• Developed an AI model using semi-supervised learning for classifying DUS signals (fetal heartbeat, artifacts, etc.).
• Utilized a dataset comprising 9,600 labeled and 48,000 unlabeled DUS signal data points.
• Compared the performance of the semi-supervised model against a supervised learning model.

Main Results:

• The semi-supervised learning model achieved an average classification accuracy of 80.9%.
• Semi-supervised learning consistently outperformed the supervised learning model in DUS signal classification.
• Demonstrated high generalization accuracy with a limited dataset.

Conclusions:

• Semi-supervised learning is effective in enhancing AI model accuracy for DUS signal classification.
• This AI approach can improve fetal monitoring quality and reduce development effort.
• The findings suggest a promising method for more reliable fetal well-being assessments.