Machine-Learning Classification of Pulse Waveform Quality.

Area of Science:

• Biomedical Engineering
• Physiological Monitoring
• Machine Learning Applications

Background:

• Wearable devices offer vital physiological monitoring, but accurate pulse waveform estimation is challenging due to motion artifacts from insufficient skin-sensor contact.
• Extracting high-quality pulse data is crucial for reliable index calculations in noninvasive monitoring.

Purpose of the Study:

• To evaluate machine learning's effectiveness in differentiating high-quality from low-quality pulse waveforms.
• To assess the impact of varying skin-surface contact pressures on pulse waveform quality.

Main Methods:

• Radial blood pressure waveform (BPW) signals were recorded noninvasively from healthy subjects using a strain-gauge transducer.
• Pulse data were collected at appropriate (67.80 ± 1.55 mmHg) and higher (151.80 ± 3.19 mmHg) contact pressures.
• Eight machine learning algorithms analyzed 40 harmonic pulse indices, including amplitude proportions and phase angles.

Main Results:

• Significant differences in BPW indices were observed between appropriate and higher contact pressures.
• Appropriate pressure ensured device stability and tissue physiological integrity.
• Machine learning demonstrated excellent discrimination performance (random-forest AUC = 0.96) in distinguishing pulse quality.

Conclusions:

• Machine learning provides an effective method for automatic screening of pulse waveform quality.
• This approach enhances the reliability of noninvasive physiological data acquisition using wearable devices.
• Further studies can explore other interfering factors using similar methodologies.