A model for atrial fibrillation detection based on differentiation and compression of interbeat interval sequences

N. S. Markov<sup>ab

a</sup> Institute of Immunology and Physiology, Ekaterinburg
^b Ural Federal University named after the First President of Russia B. N. Yeltsin, Ekaterinburg

Abstract: Atrial fibrillation is the most common arrhythmia with a major impact on public health. This paper presents a model for automatic detection of atrial fibrillation episodes in ECG, using information compression and numerical differentiation for classification of beat-to-beat interval sequences. The core of the model is normalized compression distance based on the theory of universal similarity metrics. To enable class discrimination by compression we consider finite-difference representation of interval sequences with subsequent quantization procedure. In particular, we introduce a simple $\Delta$5RR-interval representation which improves the sensitivity of the model to heart rhythm fluctuations. Our model achieves 96.37% sensitivity, 97.74% specificity and 0.935 MCC in 8x5-fold cross-validation on the MIT-BIH AFDB dataset using a segment window of 128 R-peaks. The particular advantage of the model is the classification quality in a few-shot learning setting, i.e., a training set with a small number of sequence observations can be used for classification of sufficiently large test sets.

Keywords: normalized compression distance, few-shot learning, atrial fibrillation detection, heart rate, RR-interval sequences.

DOI: 10.15514/ISPRAS-2025-37(2)-21