Prediction of atrial fibrillation over time using artificial intelligence based on ECG features and clinical risk factors

Abstract Introduction Accurate prediction of new onset atrial fibrillation (AF) could enable early diagnosis, treatment and prevention. Several predictive scores for new onset AF have been proposed, however, none are frequently used in clinical practice. AF prediction using artificial intelligence (AI) can utilize both clinical risk factors and ECG features. Purpose This study aimed to predict development of AF based on CHA₂DS₂-VA score and ECG features. Methods ECG and clinical data were obtained from the STROKESTOP I & II datasets, two large AF screening studies. Patients were instructed to perform one-lead ECG recording using a Zenicor device for 30 seconds at least twice daily for two weeks. Data on development of AF was obtained from the national patient registry after a minimum of 5 years follow up. Patients who died before the end of the study without receiving an AF diagnosis were excluded. The models were developed using data from the STROKESTOP I study and externally validated on the STROKESTOP II dataset. A threshold for CHA₂DS₂-VA score for AF prediction was calculated using Youden’s J statistic on the area under curve (AUC). For ECG data analysis, the ECGs were preprocessed using ECG Parser from Cardio Lund, machine learning based quality control and feature extraction. Upon detection of premature atrial contractions (PAC) in the ECG recordings, 5 features were considered: (1-2) the median and minimum prematurity of the beat with respect to the median RR interval, (3-4) the number of beats to the next ectopic beat from the previous beat (median and minimum), and (5) the percentage of ectopic beats. These features were employed to train two logistic regression classifiers, the first based on ECG-derived features only and the second based on ECG-derived features and CHA₂DS₂-VA score. Results The STROKESTOP I and II studies included ECG data from 5,529 and 5,685 patients, respectively, all aged 75 or 76 years. In total, 740 patients from the STROKESTOP I study and 551 patients from the STROKESTOP II study developed AF within seven years. The CHA₂DS₂-VA model predicted AF-onset with an area under curve (AUC) performance of 0.57 over the course of seven years and with varying performance over time. The ECG-based model consistently performed better than the CHA₂DS₂-VA model, with a notable improvement in performance during the first 1.5 years. Over a seven-year period, it achieved an AUC of 0.62. Performance improved further when using a model trained on both ECG features and CHA₂DS₂-VA score, where the predictive ability improved to an AUC of 0.64. Conclusion A combination of both ECG features and clinical risk factors achieved an improvement over ECG features or clinical risk factors alone for long term AF prediction. It is noted that ECG features alone also had a better predictive ability than CHA₂DS₂-VA score alone.