Multi-task deep learning for cardiac rhythm detection in wearable devices.

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence

Background:

• Wearable devices offer continuous physiological monitoring, but detecting abnormal heart rhythms like atrial fibrillation (AF) from noisy wrist-based photoplethysmography (PPG) signals remains challenging.
• Current commercial algorithms for AF detection are proprietary and often struggle with signal noise, limiting their real-world effectiveness.

Purpose of the Study:

• To develop DeepBeat, a multitask deep learning model for real-time atrial fibrillation detection using wearable PPG devices.
• To jointly assess signal quality and detect arrhythmia events, addressing the limitations of existing methods.

Main Methods:

• Trained a multitask deep learning model (DeepBeat) on approximately one million simulated unlabeled and over 500,000 labeled PPG signals from three wearable devices.
• Employed unsupervised transfer learning with convolutional denoising autoencoders and a two-stage training approach to handle data imbalance.
• Validated the model on a prospective cohort of ambulatory participants.

Main Results:

• DeepBeat achieved a significant improvement in AF detection F1 score from 0.54 (single-task model) to 0.96.
• In a prospective cohort, the algorithm demonstrated high performance with 0.98 sensitivity, 0.99 specificity, and 0.93 F1 score.
• The two-stage training effectively addressed challenges associated with unbalanced biomedical datasets.

Conclusions:

• DeepBeat offers a robust and accurate solution for real-time atrial fibrillation detection from wearable PPG devices.
• The multitask learning approach and unsupervised transfer learning significantly enhance AF detection performance and signal quality assessment.
• This method holds promise for improving remote cardiac monitoring and early diagnosis of atrial fibrillation.