Automated Segmentation of the Systolic and Diastolic Phases in Wrist Pulse Signal Using Long Short-Term Memory...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Single-period segmentation is crucial for time-domain analysis and feature extraction of pulse signals.
• Existing segmentation methods lack generalization, reliability, and robustness.

Purpose of the Study:

• To develop an advanced period segmentation method for pulse signals using Long Short-Term Memory (LSTM) networks.
• To improve the accuracy and robustness of pulse signal segmentation compared to conventional approaches.

Main Methods:

• Pulse signals underwent preprocessing to eliminate noise and baseline drift.
• The LabelMe tool was utilized to annotate pulse signal periods, defining key landmarks like the main wave start and dicrotic notch.
• A labeled dataset was created and used to train and test an LSTM network, with performance compared against the sum slope function method.

Main Results:

• The LSTM-based method achieved an overall segmentation accuracy of 92.8% across seven types of pulse signals.
• Segmentation accuracies for the systolic phase, diastolic phase, and the entire period surpassed those obtained by the sum slope function method.

Conclusions:

• The LSTM-based approach offers a robust and reliable solution for segmenting pulse signal phases (systolic, diastolic, whole period).
• This research presents a novel methodology for pulse signal segmentation, advancing time-domain analysis techniques.