Noise-Robust Atrial Fibrillation Detection from Garment-Type Wearable Holter Electrocardiogram Monitoring Using R–R Interval-Based Deep Learning: Algorithm Development and Validation Study (Preprint)

BACKGROUND Atrial fibrillation (AF) is a significant contributor to cardioembolic stroke, necessitating early and precise detection of AF to mitigate associated risks. Long-term Holter electrocardiography (ECG) monitoring using garment-type wearable devices produces large volumes of single-lead data with various noise artifacts. Deep learning has achieved high performance in AF detection from ECG data; however, many deep learning studies report strong performance on curated datasets or noise-controlled recordings. Comparatively fewer approaches have been developed and evaluated with an explicit strategy to maintain diagnostic accuracy in noise-included real-world wearable Holter ECG data. An alternative representation using the R–R interval (RRI) time series may reduce the dependence on waveform morphology and provide a computationally efficient pathway for robust AF screening in noisy recordings. OBJECTIVE This study aims to develop a computationally efficient, noise-robust deep learning model that leverages the irregularity of the RRI in noisy wearable monitoring environments. We evaluated the impact of the analysis window length on model performance. METHODS Single-lead Holter ECG data from 117 patients at the University of Osaka Hospital were analyzed, excluding those with atrial tachycardia/flutter. The RRIs were extracted, segmented into 1.5-, 3-, and 6-min windows, and transformed into two-dimensional histogram images. A ResNet-34–based two-dimensional convolutional neural network (2D-CNN) was trained for three-class classification. The model performance was evaluated using five-fold inter-patient cross-validation and externally validated using the MIT-BIH AF Database. Patient-level AF burden was defined as the proportion of AF duration relative to total analyzable recording time per patient; agreement between cardiologist-derived and model-estimated AF burden was assessed using Pearson’s correlation coefficient and linear regression. RESULTS Of 129 monitored patients (Feb 1, 2023–Nov 20, 2025), 117 were analyzed. In the internal validation, the 3-min window had superior performance (accuracy, 96.9%; AF sensitivity, 97.0%; AF specificity, 98.2%). External validation corroborated this balance (accuracy, 96.1%; AF sensitivity, 93.3%; and AF specificity, 98.7%). The 3-min model exhibited an exceptionally high correlation with the reference AF burden (r = 0.988, R² = 0.976). CONCLUSIONS The RRI-based 2D-CNN achieved high AF classification accuracy and excellent agreement with AF burden. By utilizing RRI features and a noise-adaptive training strategy, a 3-min RRI window has emerged as a practical solution for efficient AF screening in a garment-type Holter ECG.