Machine learning-based prediction of heparin anticoagulation response and dose recommendation during atrial fibrillation catheter ablation

This study investigates activated clotting time monitoring and individualized heparin dosing during atrial fibrillation catheter ablation. Data from 1175 patients treated at the First Affiliated Hospital with Nanjing Medical University between January 2020 and December 2022 were collected; after eligibility screening and preprocessing, 1144 cases were analyzed. Random Forest, XGBoost, CatBoost, DNN, LightGBM, and TabNet were evaluated using five-fold cross-validation, with each case serving once in a held-out validation fold. In the six-model comparison, DNN achieved a mean accuracy of 0.81, a Kappa coefficient of 0.68, macro-recall of 0.75, and macro-F1 of 0.72. In a separate five-fold DNN class-balancing analysis, the unbalanced mean accuracy was 0.786 and standard-category recall was 0.485; SMOTE-DNN achieved a mean accuracy of 0.812, Kappa of 0.683, macro-recall of 0.758, macro-F1 of 0.752, and standard-category recall of 0.654. For initial-dose agreement, the mean error was 728 IU, R-squared was 0.68, and 70% of predictions were within 1000 IU of the administered initial dose. In a separate matched ablation involving a 63-case post-outcome-defined subgroup, residual-based sample exclusion reduced dose MAE from 4008 to 3627 IU and increased the 1000-IU hit rate from 17.5 to 23.8%; however, R-squared remained negative, indicating limited absolute accuracy for subsequent-dose recommendation.