Aim: Identification of patients at risk to develop atrial fibrillation (AF) can enable more frequent follow-up and lead to earlier detection of AF. The aim of this project was to perform long-term AF prediction based on a novel set of ECG-derived features describing the occurrences and characteristics of supraventricular arrhythmias present before the first detected AF episode. The project employed data from the two large prospective AF screening studies, STROKESTOP I and STROKESTOP II.
Methods: In total, 12199 patients aged 75/76 performed repeated single-lead ECG measurements for 30 seconds during a period of 2-4 weeks. Arrhythmia detection was performed and subject to quality control to ensure that only reliable detections were included in the analysis. For each detected arrhythmia, a set of features were calculated including the degree of prematurity and burden of supraventricular ectopic beats, as well as the intervals between arrhythmic events. Following feature extraction, a 1D-convolutional neural network was employed to predict the long-term risk of developing AF. STROKESTOP I was used for training and STROKESTOP II was used for testing. Survival analysis by means of risk stratification curves was employed for evaluation of the risk of developing AF within a 7-year period.
Results: The best prediction model led to an ROC of 0.64 both in the training and in the test set. At the end of the observation period, the risk stratification curves for this model and for the test set showed 93% and 85% probabilities of not developing AF for the low-risk and high-risk groups, respectively. These results correspond to a weighted F1 score of 0.72.
Conclusion: The results show that ECG-derived features based on supraventricular arrhythmias detected before AF contribute to improved risk stratification for future AF diagnosis.