Directional statistics-inspired end-to-end atrial fibrillation detection model based on ECG rhythm

First, the RRI was transformed into a p-dimensional Poincaré plot representation, which reveals that the uniformity in the angular component of the Poincaré plot can depict the irregularity of the RRI, while the magnitude of the radial component represents the variability of the RRI. Then, the Shifted Spherical Projection (SSP) method was proposed to simultaneously quantify the irregularity and variability of the RRI. Furthermore, a learnable feature extraction module was devised to capture the irregularity and variability of the RRI, and it was integrated into the end-to-end model for AF detection. Finally, the effectiveness of the end-to-end model was validated through cross-dataset validation using a simulated RRI database (SIMU) and eight additional real-world databases. [Display omitted] Numerous algorithms designed for detecting atrial fibrillation (AF) often exhibit limitations in extracting essential rhythm features, leading to challenges in accurately discerning ectopic beats and consequently yielding suboptimal detection performance. In this study, we explored the distribution patterns of R-R intervals (RRIs) in both AF and specific rhythms, such as atrial premature beats (APBs) and normal sinus rhythm (NSR). For the first time, we employed the Sobolev test statistics, a method used in directional statistics to assess spherical uniformity, to quantify the irregularity and variability characteristics of RRIs. We developed an end-to-end learnable model for detecting AF by leveraging this approach. A cross-dataset validation method is employed to train and test the proposed model. This involved the use of a simulated dataset and eight distinct real-world databases. Notably, when trained on the Computing in Cardiology Challenge 2017 (C2017) and tested on the MIT-BIH Atrial Fibrillation Database (AFDB), our model, designed to take only a sequence of 32 RRIs as input, achieved a sensitivity of 96.2%, specificity of 98.2%, and accuracy of 97.3%. The results illustrate its competitive standing against the existing methods for AF detection and its enhanced resilience to ectopic beats. Unlike existing deep learning-based AF detection models, our model prioritizes interpretability, boasts lower computational complexity (with fewer than 2000 learnable parameters), and demonstrates superior generalization capabilities. This will help improve the quality of long-term real-time monitoring and management of AF, reduce the burden on clinicians, and ultimately imp