A Fast, Lightweight, and Generalizable Deep Neural Network for the Detection of Atrial Fibrillation

Atrial fibrillation (AFib) represents a critical diagnostic challenge in clinical cardiology, calling for automated detection systems capable of robust performance across diverse clinical environments. We present a computationally efficient deep neural network architecture for AFib detection that demonstrates exceptional generalizability despite training on a modest dataset. Our convolutional neural network, comprising 17 million parameters, was trained on 67,432 12-lead electrocardiograms and subsequently validated on over 1.1 million ECG recordings spanning six independent public datasets. On CPU, the model processes a 10-second ECG in 200 ms, enabling real-time inference capabilities.

Our model achieves state-of-the-art performance, with an average area under the receiver operating characteristic curve (AUROC) of 0.97, an average sensitivity of 0.83, and an average specificity of 0.96 across six external validation cohorts. These metrics rank among the highest reported in the literature, while preserving computational efficiency suitable for resource-constrained environments.

A key innovation of our approach is the implementation of channel-masking methodology, enabling seamless operation across variable lead configurations without model retraining. This flexibility allows deployment from single-lead ambulatory monitors to comprehensive 12-lead clinical systems using identical network weights. Gradient-based saliency analysis confirms the model’s attention to physiologically relevant features, particularly P-wave morphology and lead II characteristics, thereby enhancing clinical interpretability and trustworthiness.

Our findings also establish that a single, well-curated small training dataset can yield a compact yet highly generalizable AFib detection system suitable for deployment across diverse clinical settings, from critical care monitoring to ambulatory screening applications. The combination of robust cross-dataset performance, computational efficiency, and clinical interpretability positions this approach as a viable solution for largescale AFib diagnosis and monitoring programs.