Introduction: Atrial fibrillation (AF) is an under-detected cardiac arrhythmia with high risk of stroke and mortality. The pathophysiological pathway that leads to the onset of AF is largely unknown. An AI-ECG model demonstrated to detect latent AF from single ECGs in sinus rhythm (SR) within 31 days prior to AF onset. We hypothesise that the time evolution of the AI-ECG model score can predict future AF and improve overall risk predictions as compared to the standalone AI-ECG score.
Methods: The score of the AI-ECG AF model is mapped to patient specific ECG trajectories. A convolutional neural network is trained to differentiate trajectories related to AF patients from those of healthy individuals. Patients were included if they had at least three SR ECGs spaced one year within a five-year window. Included non-AF patients have no history of AF and at least five years follow up. ECG scores within the same one-year boundary were averaged. The results were compared to the AI-ECG AF risk score of the most recent ECG within the window.
Results: 20 948 patients were eligible given the inclusion criteria. The trajectory model resulted in an area under the receiver operator curve (AUC) of 0.80 (95% CI 0.78-0.82) for prediction of AF up to two years prior to onset. Standalone AI-ECG AF risk of the most recent ECG resulted in an AUC of 0.77 (95% CI 0.75-0.79). Using trajectory analysis, the prediction window of AF could be prolonged with clinically grade performance to up to two years.
Conclusion: Trajectory analysis of an AI-ECG AF risk score can improve and prolong prediction of AF as compared to standalone risk assessment. Results emphasise the impact of considering long term patterns of AF risk for better contextualization and stratification of patients at risk.