Non-invasive Fetal ECG Signal Quality Assessment based on Unsupervised Learning Approach

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Non-invasive fetal electrocardiogram (FECG) monitoring is crucial for fetal well-being assessment.
• FECG signal quality directly impacts fetal heart rate (FHR) estimation accuracy.
• Existing supervised learning methods for FECG signal quality assessment (SQA) require extensive labeled datasets, which are scarce.

Purpose of the Study:

• To develop an unsupervised learning-based method for FECG signal quality assessment (SQA).
• To address the limitation of scarce labeled datasets in FECG SQA.
• To accurately distinguish between high and low-quality FECG signal segments.

Main Methods:

• An unsupervised approach utilizing a fully convolutional network (FCN)-based autoencoder (AE) for FECG spectrogram reconstruction.
• Calculating AE-based reconstruction error features to identify signal quality.
• Extracting additional entropy-based features, statistical features, and ECG signal quality indices (SQIs).
• Employing a self-organizing map (SOM) for classifying high and low-quality FECG segments based on extracted features.

Main Results:

• The proposed unsupervised method achieved 98% accuracy in classifying high and low-quality FECG signal segments.
• Demonstrated the effectiveness of AE reconstruction error and other extracted features for SQA.
• Successfully identified high and low-quality FECG segments without reliance on labeled data.

Conclusions:

• The developed unsupervised learning method offers a viable solution for FECG SQA, overcoming the challenge of limited labeled data.
• This approach can enhance the accuracy of FHR estimation by effectively filtering or interpolating low-quality FECG signals.
• The method holds promise for improving the reliability of non-invasive fetal monitoring.