Atrial fibrillation detection from raw photoplethysmography waveforms: A deep learning application.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Atrial fibrillation (AF) is a stroke risk, often asymptomatic, necessitating accurate detection.
• Smart devices use photoplethysmography (PPG) for AF screening, but accuracy needs improvement to minimize false positives.

Purpose of the Study:

• To evaluate if a deep learning algorithm using raw, smartwatch-derived PPG waveforms can better discriminate AF from normal sinus rhythm compared to algorithms using only heart rate data.

Main Methods:

• 51 patients with AF underwent PPG sensing via wrist-worn fitness trackers.
• Electrocardiograms served as the reference standard.
• Evaluated accuracy using heart rate variability, a long short-term memory (LSTM) network on heart rate data, and a deep neural network (DNN) on raw PPG data.

Main Results:

• The deep neural network (DNN) model using raw PPG data achieved the highest accuracy.
• DNN model yielded an area under the receiver operating characteristic curve (AUC) of 0.983 (sensitivity 0.985, specificity 0.880).
• This outperformed conventional heart rate variability (AUC 0.717) and LSTM models (AUC 0.954).

Conclusions:

• A deep learning model leveraging raw PPG signals achieves high accuracy in AF detection.
• This approach surpasses traditional methods relying solely on heart rate data for AF screening.