Personalized machine learning framework for early prediction of PAF onset from APC-dominant ECG segments

Paroxysmal Atrial Fibrillation (PAF) is often preceded by Atrial Premature Complexes (APCs), which serve as early indicators of atrial electrical instability. In this study, we propose a personalized machine learning framework for early PA F onset prediction using APC-dominant ECG segments. A set of handcrafted features capturing temporal variability and nonlinear dynamics was extracted from 10-second APC segments and modelled using Support Vector Machine (SVM) and Random Forest (RF) classifiers. A patient-specific 10-fold cross-validation strategy was applied to ensure robust performance across inter-patient variability. Experimental results on the PhysioNet Atrial Fibrillation Database (AFDB) demonstrated strong predictive performance across 15 representative patients. The RF model achieved an average accuracy of 95.84%, precision of 94.68%, recall of 88.55%, and F1-score of 91.00%, outperforming SVM in both robustness and consistency. These findings suggest that interpretable, lightweight machine learning models trained on physiologically meaningful APC events can support early and patient-specific prediction of PAF onset, offering a clinically feasible tool for proactive rhythm monitoring.