Introduction: Atrial fibrillation (AF) recurrence is considered a random phenomenon with regards to its initiation, recurrence pattern, and termination. We sought to investigate if and how well the amount and pattern of past AF recurrence can forecast short term (30 day) future AF recurrence.
Methods: We reconstructed the cardiac rhythm histories of 1195 patients with implantable cardiac devices (age 73 ± 10 years, follow-up: 349 ± 40 days, mean AF burden: 11%). The weekly amount (AF burden) and pattern of AF recurrence (AF density) of the 1st 150 days (1st half of the rhythm history) for each patient was used to construct a supervised machine learning algorithm (random forest regression model) with the aim to predict the amount of AF recurrence during days 151–180. Validation of the algorithm was performed on the 2nd half of the observation period of the same patients (AF pattern of days 150–300 to predict AF recurrence in days 301–330). An additional 2nd external validation of the algorithm was performed on a separate patient cohort consisting of 574 patients (age 69 ± 13 years, follow-up: 365 days, mean AF burden: 12%).
Results: In the training cohort, the algorithm accounted for 86% of the observed variance with a mean squared error rate of 0.01. Variable importance measures revealed that the past amount (AF burden) and pattern (AF density) exhibited an important influence on future AF recurrence with decreasing importance for more remote events. AF recurrence up to 6 weeks in the past influenced short term (30 day) future AF recurrence. Validation of the algorithm on the 2nd half of the observation period and the external patient cohort revealed low median AF burden prediction errors of 0.004 [0.002–0.004] and 0.004 [0.003–0.028], respectively (Fig. 1).
Fig. 1
Conclusion: The amount and temporal pattern of past AF recurrence (up to 6 weeks) is associated with short term future AF recurrence. This predictive potential attenuates progressively over time and merits further investigation.
