Heart Sound Segmentation-An Event Detection Approach Using Deep Recurrent Neural Networks

Area of Science:

• Cardiology
• Biomedical Engineering
• Signal Processing

Background:

• Accurate segmentation of heart sounds is crucial for diagnosing cardiac conditions.
• Existing methods often rely on prior knowledge of state durations, limiting their applicability to irregular heart rhythms.

Purpose of the Study:

• To develop an event detection approach for accurately identifying the heart sound sequence (S1-systole-S2-diastole) without assuming state durations.
• To create a method applicable to recordings with cardiac arrhythmia and extendable to other heart sound events.

Main Methods:

• Utilized the 2016 PhysioNet/CinC Challenge dataset with heart sound recordings and annotations.
• Extracted spectral and envelope features from heart sound recordings.
• Employed deep recurrent neural network (DRNN) architectures with virtual adversarial training, dropout, and data augmentation for regularization.

Main Results:

• Achieved an average F1-score of approximately 96% for the four events of the state sequence on an independent test set.
• Demonstrated state-of-the-art performance when compared to existing methods.

Conclusions:

• The proposed event detection methodology using DRNNs offers a robust and accurate approach for heart sound segmentation.
• The method's ability to function without prior state duration information makes it versatile for various cardiac conditions, including arrhythmias.