Machine-learning guided differentiation between photoplethysmography waveforms of supraventricular and ventricular...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Distinguishing supraventricular from ventricular arrhythmias using wearable photoplethysmography (PPG) signals remains challenging.
• Investigating the utility of PPG waveforms for arrhythmia origin classification is crucial for non-invasive diagnostics.

Purpose of the Study:

• To assess if a neural network classifier can accurately distinguish the origin of PPG pulse waveforms.
• To evaluate the performance of a convolutional neural network in classifying PPG signals as either supraventricular (atrial pacing) or ventricular.

Main Methods:

• PPG waveforms were recorded from 30 patients during electrophysiological (EP) studies using a wristband device.
• Waveforms were synchronized with ECG and intracardiac electrograms, and labeled as atrial pacing (AP) or ventricular pacing (VP).
• A residual neural network was developed and validated on 25,221 PPG waveform samples.

Main Results:

• The classifier achieved approximately 73% accuracy for AP and 59% for VP on an independent patient level.
• With adaptive, patient-specific annotations, the classifier's accuracy improved to ~97% for AP and ~95% for VP.
• The study included 27 patients (74% female, median age 53 years) with various tachyarrhythmias.

Conclusions:

• A neural network trained on EP-derived PPG data can differentiate between supraventricular and ventricular origins.
• PPG waveforms alone, analyzed by a neural network, show potential for identifying arrhythmia types.
• This approach may offer a non-invasive method for arrhythmia diagnosis using wearable technology.