Background: While AI-ECG derived biomarkers have been shown to improve prediction of atrial fibrillation (AF) when added to traditional risk score models such as CHARGE-AF, existing studies predominantly rely on linear combinations of an AI-ECG output and patient clinical variables in their analyses, potentially masking nonlinear relationships.
Aim: To determine whether an AI-ECG biomarker adds significant independent prognostic value for 5-year incident AF when accounting for complex nonlinear relationships and confounding by patient clinical variables.
Methods: We derived a retrospective cohort from the MIMIC-IV database to predict the outcome of incident AF within 5 years. The logit of a pretrained AI-ECG AF prognosis model and the CHARGE-AF variables were used as features. Several linear and nonlinear models were trained (e.g. Logistic Regression, XGBoost, Random Forest) with and without the AI-ECG biomarker, and performance was compared via AUC-ROC with bootstrapped 95% confidence intervals. Additionally, we applied the Double Machine Learning (DML) framework using an interactive regression model to estimate the average predictive effect of a positive AI-ECG prognosis when controlling for nonlinear confounding.
Results: Validation showed improved discrimination when including the AI-ECG biomarker in both Logistic Regression (0.76 [0.74,0.79] vs. 0.74 [0.71, 0.76]) and Random Forest models (0.78 [0.75, 0.80] vs. 0.75 [0.73, 0.78]), with similar improvements observed in all other evaluated models. Further, DML analysis revealed a statistically significant average predictive effect of a positive AI-ECG prognosis (0.42 [0.35, 0.50], p<0.001), which was underestimated by a naive additive linear probability model (0.33 [0.28, 0.38], p<0.001).
Conclusion: An AI-ECG biomarker for incident AF provides significant independent prognostic value and is complementary to traditional risk models, even when accounting for nonlinear clinical interactions.