Feature extraction tool using temporal landmarks in arterial blood pressure and photoplethysmography waveforms

Area of Science:

• Biomedical Engineering
• Cardiovascular Physiology
• Medical Informatics

Background:

• Arterial (ABP) and photoplethysmography (PPG) waveforms contain rich physiological information.
• Extracting features from these waveforms is crucial for clinical applications and machine learning.
• Manual feature extraction is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and validate an automatic tool for feature extraction from cardiac cycles in ABP and PPG waveforms.
• To assess the tool's performance in landmark detection and feature extraction accuracy.
• To evaluate the tool's potential for clinical applications and machine learning model development.

Main Methods:

• Developed an automatic feature extraction tool identifying key landmarks (systolic onset/peak, dicrotic notch, diastolic peak) in cardiac cycles.
• Extracted 852 features per cycle across time, statistical, and frequency domains.
• Validated landmark detection on the MLORD dataset (17,327 patients) and real-time patient monitor data.

Main Results:

• The tool demonstrated robust performance in detecting all four landmarks across datasets and waveform types.
• Achieved average F1-scores above 97% and error rates below 4% compared to expert annotations.
• Successfully extracted a comprehensive set of 852 features per cardiac cycle.

Conclusions:

• The automatic feature extraction tool exhibits high accuracy and robustness for ABP and PPG waveforms.
• This tool can significantly support the clinical use of waveform-derived features.
• Facilitates the development of advanced, feature-based machine learning models for diverse clinical applications.