Early prediction of Paroxysmal Atrial Fibrillation via machine learning and an interpretable fuzzy expert system

Paroxysmal Atrial Fibrillation (PAF) is a transient and often asymptomatic arrhythmia that poses challenges for early diagnosis using standard ECG monitoring. Hence, this study proposes an approach based on signal processing and machine learning for the early prediction of PAF episodes, considering different characteristics extracted from ECG signals, including the presence or absence of P-wave and the variability of RR interval, which are usually related to preceding PAF episodes. The method involves signal pre-processing techniques, such as noise filtering and segmentation, to enhance signal clarity and isolate individual cardiac cycles in the most distant segments from the episode (1 min, 30 s, and 15 s). Several features are extracted from the segments to train various machine learning models, including a fuzzy rule-based expert (FRBS) system, for binary classification of patients as either exhibiting PAF or not. Experimental results on the used ECG database demonstrate that the proposed framework achieves high classification performance, with the best-performing model being Linear Discriminant on the 15-second segment reaching an accuracy of 94.79%, precision of 92.30%, recall of 97.96%, F1-score of 95.05%, and specificity of 91.49%. Notably, the FRBS system, using only three features compared to the full 14-feature set used by the other models, shows competitive performance across all segment lengths and achieves the best results on the 30-second segment, highlighting interpretability for early PAF prediction. • P-wave and RR interval used for early PAF prediction from ECG signals. • Comparative evaluation of AI models using short ECG segments for PAF prediction. • Interpretable fuzzy expert system provides transparent PAF classification.