A novel heart sound segmentation algorithm via multi-feature input and neural network with attention mechanism

Area of Science:

• Cardiology
• Biomedical Engineering
• Signal Processing

Background:

• Heart sound segmentation (HSS) is crucial for assessing heart health by identifying S1, S2 sounds, systole, and diastole.
• Existing neural network methods show promise but can be improved for feature learning.
• Phonocardiogram (PCG) signal analysis is fundamental for non-invasive cardiac diagnostics.

Purpose of the Study:

• To develop a novel and accurate method for heart sound segmentation (HSS).
• To enhance the feature learning process for neural networks in HSS.
• To improve the identification of S1, S2, systole, and diastole from PCG signals.

Main Methods:

• A novel Convolution and Bidirectional Long-Short Term Memory neural network with Attention mechanism (C-LSTM-A) was developed for HSS.
• Incorporated 0.5-order smooth Shannon entropy envelope, instantaneous phase waveform (IPW), and third intrinsic mode function (IMF-3) of PCG signals.
• Utilized the Fuwai Yunnan Cardiovascular Hospital heart sound dataset and the 2016 PhysioNet/CinC Challenge dataset.

Main Results:

• Achieved an average F1-score of 96.85% on the clinical dataset.
• Attained an average F1-score of 95.68% on the PhysioNet/CinC Challenge dataset.
• Demonstrated effectiveness for both normal and common pathological PCG signals.

Conclusions:

• The proposed C-LSTM-A method significantly improves heart sound segmentation accuracy.
• The method's ability to segment fundamental heart sound components (S1, S2) and cycle phases is beneficial for further heart sound classification studies.
• This approach offers a robust tool for analyzing PCG signals and contributes to advancing cardiovascular health diagnostics.