Atrial Fibrillation (AF) onset prediction could enable new preventive strategies based on improved overdrive pacing in patients with cardiac implantable devices (CIED). The role of autonomic nervous system (ANS) imbalance in the initiation of AF is now well established. Heart rate fragmentation (HRF), Poincaré plots and second-order difference of RR intervals plots (SODP) are markers of ANS dysfunction and a predictor of AF.

Determine whether normal sinus rhythm (NSR) windows closed to the AF onset are distinguishable from NSR windows further away (forecast). Compare these two types of windows with NSR windows from healthy subjects (identification of patient at-risk).

We screened our database of 53,086 Holter records to identify paroxysmal AF episodes. For AF patients, we selected the 1-hour window at the start of the record if no sign of AF was found in the window, and 1-hour before all AF onsets if there was no previous AF in the window. For healthy subjects, i.e. without any sign of AF or any cardiac condition in the record, we selected the first hour and the twelfth hour of the record. We extracted the RR intervals and selected geometric features from Poincaré plot and SODP, and HRF features. We used a machine learning (ML) technique (boosted trees) to compare the three types of windows. We tested the performance of the algorithm with 10-fold patient cross-validation.

We identified 491 AF records from 445 patients (820 AF episodes with 1-hour NSR before the onset). We found 349 records from 342 healthy subjects. The model achieved an accuracy of 0.66 (95% CI 0.63–0.68) in categorizing the three window types. The confusion matrix is presented in Fig. 1. For a binary classification between the two types of AF windows, the model achieved an accuracy of 0.67 (0.64–0.70). The best accuracy was achieved in a binary classification between windows close to AF onset and healthy windows: the model achieving an accuracy of 0.85 (0.83–0.87). The most important features for all tasks are from the Poincaré (SD1 and SD2) and the mean length of HRF segments.

Our ML model was able to discriminate three types of NSR windows (close to AF, distant from AF and healthy subjects). These results may lead to new opportunities for AF onset prediction and AF screening, which could improve patient care by allowing earlier intervention. The forecasting task could help optimize the overdrive atrial pacing strategy in CIED. The identification task could allow a pill-in-the-pocket strategy to avoid AF crises.