A photoplethysmogram-based 1D-CNN algorithm for automated atrial fibrillation detection.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Atrial fibrillation (AF) is a common arrhythmia.
• Traditional AF diagnosis relies on electrocardiograms (ECG).
• Photoplethysmogram (PPG) signals from wearables offer a non-invasive alternative for monitoring.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for AF detection using PPG signals.
• To assess the performance of the algorithm against existing state-of-the-art methods.
• To explore model optimization techniques like pruning and binarization for computational efficiency.

Main Methods:

• A 1-dimensional convolutional neural network (CNN) was developed using the MIMIC III Waveform database.
• The CNN model was trained and validated for AF detection from PPG signals.
• Model pruning and binarization were applied to assess computational complexity reduction.

Main Results:

• The developed CNN model achieved 98.27% accuracy and 97.78% F1-score for AF detection.
• The pruned network (60% sparsity) maintained high performance with 97.26% accuracy and 96.40% F1-score.
• The binarized network demonstrated feasibility with 94.51% accuracy and 93.29% F1-score.

Conclusions:

• Deep learning-based AF detection from PPG signals is highly accurate and feasible using wearable devices.
• The proposed method offers a promising approach for continuous telemonitoring and early AF detection.
• Model optimization techniques can reduce computational load without significant performance degradation.