Atrial fibrillation prediction in patients with hypertrophic cardiomyopathy based on long-term follow-up data and machine learning model.

Background: Hypertrophic cardiomyopathy (HCM) significantly increases the risk of new-onset atrial fibrillation (AF) through distinct pathogenic mechanisms. This study develops a machine learning (ML) model to improve AF prediction in patients with HCM and investigates electrophysiological abnormalities and non-traditional factors as preclinical predictors. The findings aim to inform early warning systems and precision management strategies for AF in this patient cohort. Methods: This retrospective cohort study consecutively enrolled patients diagnosed with HCM who had no prior history of AF from 2014 to 2023. A two-step feature selection procedure involving the Boruta algorithm and LASSO regression was implemented, and four machine-learning based algorithms were trained and compared against the reference model of the HCM-AF score. The Shapley Additive Explanations (SHAP) method was utilized for interpretability. Results: A total of 225 patients (annual incidence ratio 3.42%, 225/1014) developed new-onset AF during the median follow-up period of 6.50 years. Twelve clinical variables were identified through the feature selection process, with the random forest model demonstrating the best overall performance with area under the receiver-operating characteristic curve (AUC) 0.770 and 95% confidence interval (CI) of 0.699-0.840, outperforming the reference model (AUC 0.692, 95% CI 0.598-0.786) (Delong's Conclusions: The Random Forest model outperformed the HCM-AF score with robust predictive ability, validating P-wave indices as preclinical AF biomarkers through ML model for the first time. This integrated model offers a practical tool for early AF risk stratification in HCM, highlighting the clinical value of early symptomatic monitoring. Future multicenter validation is needed.