Multiclass Arrhythmia Classification Using Multimodal Smartwatch Photoplethysmography Signals Collected in Real-Life...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Smartwatches with photoplethysmographic (PPG) sensors offer continuous monitoring for early atrial fibrillation (AF) detection.
• Previous deep learning models for AF detection were limited by controlled environments, short data durations, and difficulty distinguishing AF from premature atrial/ventricular contractions (PAC/PVC).
• Limited datasets for PAC/PVC detection have hindered the performance of current state-of-the-art methods, achieving only 75% sensitivity.

Purpose of the Study:

• To address limitations in AF and PAC/PVC detection using smartwatch PPG data.
• To develop a computationally efficient deep learning model for accurate arrhythmia detection.
• To improve the generalizability of AF and PAC/PVC detection models across different datasets and devices.

Main Methods:

• Utilized data from the NIH-funded Pulsewatch clinical trial, collecting over two weeks of smartwatch PPG data from 106 subjects.
• Developed a 1D bi-directional Gated Recurrent Unit deep learning model incorporating multi-modal inputs (PPG, accelerometer, heart rate).
• Classified data into normal sinus rhythm, AF, and PAC/PVC categories.

Main Results:

• Achieved 83% sensitivity for PAC/PVC detection and 97.31% accuracy for AF detection, significantly outperforming prior methods.
• The model demonstrated 14 times greater computational efficiency and 2.7 times faster processing speed.
• Validated generalizability on external datasets, achieving macro-averaged AUROC values of 96.22% and 94.17%.

Conclusions:

• A lightweight, multimodal deep learning model can accurately differentiate PAC/PVC from AF, reducing false positives.
• Enhanced accuracy in detecting both AF and PAC/PVC can increase clinical and public trust in smartwatch-based AF monitoring.