Remote Atrial Fibrillation Burden Estimation Using Deep Recurrent Neural Network.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Atrial fibrillation burden (AFB) is the percentage of time spent in atrial fibrillation (AF).
• AFB offers greater prognostic value than a binary AF diagnosis.
• Estimating AFB from long-term continuous recordings is crucial for patient management.

Purpose of the Study:

• To evaluate the ability to estimate AFB from long-term continuous ECG recordings.
• To introduce a novel deep recurrent neural network (DRNN) approach for AFB estimation.
• To compare the DRNN model's performance against a gradient boosting model.

Main Methods:

• Developed and evaluated a DRNN model (ArNet) on 68,800 hours of ECG data from 2,891 patients.
• Utilized 24-hour beat-to-beat time series from a single portable ECG channel.
• Benchmarked ArNet against a gradient boosting (XGB) model using features like sample entropy and AFEvidence.

Main Results:

• The DRNN model (ArNet) achieved a median absolute AF burden estimation error of 1.2 (0.1-6.7) on the test set.
• The XGB model had a median error of 2.8 (0.9-11.7) for AF individuals.
• Generalization performance on the PhysioNet LTAF database was consistent, with ArNet showing lower error rates.

Conclusions:

• Deep recurrent neural networks (DRNNs) can feasibly estimate AFB from 24-hour beat-to-beat interval time series.
• This data-driven approach facilitates robust remote diagnosis and phenotyping of atrial fibrillation.
• The findings support the clinical utility of DRNNs in managing patients with AF.