• Citer

Short-term atrial fibrillation onset prediction using machine learning

par Grégoire, Jean-Marie ;Gilon, Cédric ;Marelli, François;Bersini, Hugues ;Groben, Laurent;Nguyen, Thomas ;Deruyter, Bernard;Godart, Pascal ;Carlier, Stéphane
Référence European heart journal. Digital health, 6, 6, page (1159-1168)
Publication Publié, 2025-11-01

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Introduction Integrating machine learning (ML) models into wearable or connected devices to deliver early warning alerts prior to atrial fibrillation (AF) onset may represent an effective preventive strategy. Machine learning algorithms applied to two-lead Holter electrocardiogram (ECG) recordings can support the development of predictive models capable of detecting imminent paroxysmal AF episodes within short-term windows. This approach could facilitate a more targeted 'pill-in-the-pocket' (PITP)-like intervention strategy, potentially enhancing timely therapeutic actions and improving patient outcomes. Aim This study aimed to identify patients currently in sinus rhythm who will experience an AF episode within the subsequent hours by analysing 24-h Holter ECG recordings with ML. Methods We established a novel database comprising 95 871 manually analysed Holter ECG recordings, identifying 1319 episodes of paroxysmal AF from 872 patients. Among these, 835 AF episodes from 506 recordings had more than 60 min of normal sinus rhythm prior to AF onset and more than 10 min of sustained AF following onset. Patients were stratified into five age groups: all patients combined, under 60 years, 60-70 years, 70-80 years, and over 80 years. Additionally, 365 recordings from 347 patients without rhythm abnormalities were identified and classified, from which two ECG segments were selected. Two deep learning (DL) models were trained on raw ECG data to predict AF onset. To compare DL models with traditional ML approaches using heart rate variability (HRV) parameters, we employed a random forest classifier and a gradient-boosted decision tree model (XGBoost, XGB). Results The decision trees models trained on HRV parameters delivered the best predictive performance. The most significant results were observed for episodes lasting more than 5 min of AF, achieving an area under the receiver operating characteristic curve of 0.919 (95% CI: 0.879-0.958) and an area under the precision-recall curve of 0.919 (95% CI: 0.879-0.958) for XGB. At a decision threshold of 0.5, accuracy was 84.5% (81.2-87.8), sensitivity was 83.0% (79.5-86.4), specificity was 86.6% (79.3-93.9), positive predictive value was 90.2% (85.5-94.9), negative predictive value was 78.4% (74.7-82.1), and the F1 score was 86.2% (83.5-89.0). Conclusion These findings indicate that HRV parameters contain crucial information for the short-term prediction of AF onset, supporting preventive strategies. Integration of such predictive models into wearable mHealth technologies could facilitate a PITP-like preventive approach, potentially reducing AF-related morbidity. Prospective studies are warranted to validate these promising results further.