Dynamic prediction of paroxysmal atrial fibrillation onset using longitudinal sample entropy in joint models

Atrial fibrillation (AF) is the most common cardiac arrhythmia and is associated with a five-fold increased risk of stroke. Early prediction of AF onset could improve care for at-risk patients. Existing predictive models often rely on clinical risk scores or machine learning approaches using, for instance, heart rate variability (HRV) features. Among these, sample entropy (SampEn), a quantitative measure of signal complexity, has shown promise as a predictor of AF onset. In this study, we proposed a joint modeling approach that incorporates both baseline covariate and the longitudinal trajectory of SampEn to estimate the risk of AF onset within the next 12 hours.

We developed several joint models using varying model structural complexity, particularly in modeling the longitudinal process. We evaluated models performance using bootstrap replications on the publicly available IRIDIA-AF dataset. We used time-dependent area under the curve (AUC), sensitivity, and specificity, together with other calibration and accuracy measures to assess predictive performance. Additionally, we illustrated individual prediction profiles for selected patient records.

The best-performing model, which used natural cubic splines in the longitudinal submodel, achieved an AUC of 64.4% and a sensitivity of 77.63%. A simpler model using a linear longitudinal trajectory achieved the highest specificity of 77.94%.

These results demonstrate the potential of joint models for short-term AF risk prediction, providing not only binary classification but also dynamic, individualized risk estimates over time. They enable updated predictions and personalized monitoring of patient risk as new longitudinal data become available.