R-R Interval Histogram-Based Deep Learning for 3-Class Atrial Fibrillation Screening in Garment-Type Wearable Holter...

Area of Science:

• Cardiology and Medical Technology
• Artificial Intelligence in Healthcare
• Signal Processing for Biosensors

Background:

• Wearable Holter electrocardiographic (ECG) monitoring for long-term atrial fibrillation (AF) detection is often compromised by noise.
• Deep learning models show promise for AF detection but require explicit strategies for handling noise in wearable ECG data.
• R-R interval (RRI) time series offer an alternative representation that may reduce reliance on waveform morphology for AF screening in noisy environments.

Purpose of the Study:

• To develop and evaluate a 3-class, noise-aware framework using RRI time series for AF screening.
• To explicitly differentiate between AF, non-AF, and uninterpretable noise segments within wearable ECG recordings.
• To assess the impact of different analysis window lengths (1.5, 3, and 6 minutes) on the framework's performance.

Main Methods:

• Analysis of single-lead garment-type wearable Holter ECG data from 117 patients.
• Conversion of RRI segments into 2D histogram images (time vs. RRI-derived heart rate) for 1.5-, 3-, and 6-minute windows.
• Training a ResNet-34-based 2D convolutional neural network for 3-class classification, validated internally and externally using the MIT-BIH AF Database.

Main Results:

• Internal validation showed high performance for 1.5- and 3-minute windows (96.6% accuracy).
• External validation demonstrated the 3-minute window achieved the highest overall performance (97.3% accuracy, 96.9% AF sensitivity, 97.7% AF specificity).
• High correlation was observed at the patient level between reference AF burden and model-estimated AF burden across all window lengths.

Conclusions:

• The RRI-based 2D convolutional neural network framework effectively classifies AF with high accuracy.
• Explicitly separating noise using a 3-class framework improves AF screening in noisy wearable ECG data.
• A 3-minute RRI window offers a favorable balance of performance for AF screening in garment-type Holter ECG monitoring.