Background: Atrial fibrillation (AF) has escalated in prevalence over the past five decades, emerging as the predominant sustained heart arrhythmia worldwide. This condition significantly elevates the risk of severe cardiovascular diseases (CVD), particularly with silent paroxysmal atrial fibrillation (PAF). Predicting PAF occurrence in individuals without known history and, hence, maintaining normal sinus rhythm (NSR) for long time episodes, is vital for early intervention and averting adverse outcomes. While previous studies predominantly relied on electrocardiogram (ECG) monitoring, heart rate variability (HRV), and premature atrial contractions (PAC) to forecast PAF, the extensive and scattered research calls for a reevaluation on the same dataset and under the same settings. Thus, this study aims to replicate and scrutinize the most pertinent approaches described in the current literature, utilizing a meticulously curated dataset to assess the genuine performance of the generated models.
Method: The study utilized a controlled subset of the Icentia-11k Dataset from PhysioNet's repository. This subset comprised 30-minute long, 200 NSR ECG recordings from 200 distinct patients, 100 presenting occult AF and 100 with no cardiac issues reported. After filtering the ECG recordings to eliminate baseline wander and high-frequency noise, various common features were extracted from the HRV across three domains: time, frequency, and complexity. Subsequently, different machine learning models were evaluated on the extracted features, incorporating information on PACs. Performance metrics such as accuracy (Ac), sensitivity (Se), and specificity (Sp) were computed under a 80:20 training-to-testing ratio
Results: Average Ac, Se, and Sp were respectively 67.50%, 71.76%, 60.91% for time-domain features, 61.79%, 68.82%, 50.91% for frequency-domain features, 60.36%, 74.12%, 39.09% for complexity-domain features, 68.87%, 39.13%, 82.61% using solely PACs information, and 68.93%, 68.24%, 70.00% using all features.
Conclusions: While still open to potential refinements, the study reveals diverse performance across feature domains. Also, while PAC-centric models excel in specificity, they compromise sensitivity.