RawECGNet: Deep Learning Generalization for Atrial Fibrillation Detection From the Raw ECG.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning models for atrial fibrillation (AF) detection using rhythm analysis achieve high performance.
• Rhythm-based approaches overlook crucial morphological information in ECG waveforms, potentially limiting accuracy.
• Atrial flutter (AFl) detection also benefits from comprehensive ECG analysis.

Purpose of the Study:

• To develop and evaluate RawECGNet, a novel deep learning model for detecting AF and AFl episodes using raw, single-lead ECG data.
• To assess the generalizability of RawECGNet across diverse datasets with geographical, ethnic, and lead position variations.
• To benchmark RawECGNet against a state-of-the-art rhythm-based model, ArNet2.

Main Methods:

• Developed RawECGNet, a deep learning model processing raw, single-lead ECG signals.
• Evaluated RawECGNet on two external datasets (RBDB and SHDB) to test generalization.
• Compared RawECGNet's performance against ArNet2, a model using only rhythm information.

Main Results:

• RawECGNet achieved F1 scores of 0.91-0.94 in RBDB and 0.93 in SHDB.
• ArNet2 achieved F1 scores of 0.89-0.91 in RBDB and 0.91 in SHDB.
• RawECGNet demonstrated superior and generalizable performance in detecting AF and AFl episodes.

Conclusions:

• RawECGNet is a high-performance, generalizable algorithm for AF and AFl detection.
• The model effectively leverages both rhythm and morphological ECG information.
• RawECGNet offers an advancement over rhythm-only deep learning approaches for arrhythmia detection.