Machine learning has been shown to be effective for QT analysis, especially in patients with both acquired and congenital long QT syndrome. The relationship between QT and RR intervals (QT-dynamicity) has not yet been used for the detection and forecast of atrial fibrillation episodes.

Study the importance of ECG delineation features and especially QT-dynamicity for the detection and forecast of paroxysmal AF episodes.

24H Holter ECG recordings from 88 patients were used, allowing an in-depth analysis of the transition from sinus rhythm to the AF episodes. Raw ECG signals were delineated using a wavelet-based signal processing technique. The dataset is composed of more than 17 million values for each patient. A selection of delineation features was performed from a statistical analysis and literature review. A total of 44 ECG features were chosen (e.g. interval and wave durations and amplitude). A machine learning model (XGBoost) was trained with a Bayesian selection of hyperparameters for different windows. We used a 5-fold cross-validation method for model validation.

Mean age of the patients was 75.9±11.9 (range 50–99), number of episodes per patient was 2.3±2.2 (range 1–11) and CHA2DS2-VASc score was 2.9±1.7 (range 1–9). For the detection, we obtained an area under the receiver operating curve (AUROC) of 0.988 (CI 95%: 0.987–0.989) and an accuracy of 95% using a 30s window. For the forecast, we obtained an AUROC of 0.739 (0.712–0.766) and an accuracy of 74% using a 120s window. For the detection, features related to RR intervals were the most important, followed by those on QT intervals. For the forecast, QT dynamicity as assessed by the Spearman's correlation of the QT-RR slope was the best predictor (Fig. 1).

In addition to RR intervals, QT intervals and QT-dynamicity are important predictors for AF detection and short-term forecast. These data suggest that ventricular repolarization changes play a relevant role in the triggering of AF episodes.