Machine Learning Predicts Heparin Response During AF Ablation

ACTIVATED clotting time prediction using machine learning achieved strong performance in patients with atrial fibrillation undergoing catheter ablation, with a deep neural network model outperforming five alternative algorithms in a retrospective analysis of 1,144 cases.

Activated Clotting Time Prediction Performance

Maintaining appropriate anticoagulation during atrial fibrillation catheter ablation is critical to reduce the risk of thrombotic and bleeding complications.

Researchers analysed clinical data from 1,144 patients who underwent catheter ablation between January 2020 and December 2022 to develop a model capable of predicting activated clotting time (ACT) status following heparin administration.

The study evaluated six machine learning approaches, including Random Forest, XGBoost, CatBoost, deep neural network (DNN), LightGBM, and TabNet.

Patients were categorised into three ACT groups: substandard (<250 seconds), standard (250-300 seconds), and exceeding standard (>300 seconds).

Among all models assessed using fivefold cross validation, the DNN demonstrated the highest performance, with 81% accuracy.

Deep Learning Improves Anticoagulation Prediction

Investigators also explored methods to address class imbalance within the dataset. In a separate DNN analysis, synthetic minority oversampling technique (SMOTE) training outperformed both unbalanced and cost sensitive approaches.

Compared with unbalanced training, SMOTE improved mean accuracy from 78.6% to 81.2%.

The findings indicate that balancing strategies may improve recognition of patients whose ACT values fall within the desired operational range during the procedure.

Heparin Dose Recommendation Remains Challenging

The researchers also assessed whether the model could support heparin dose recommendations.

For initial heparin dosing, predictions demonstrated reasonable agreement with doses administered in clinical practice, with approximately 70% of predictions falling within 1,000 IU of the observed dose.

Performance was less robust for subsequent dose recommendations.

In a post-outcome defined subgroup of 63 cases, a residual based optimisation approach reduced mean absolute error from 4,008 IU to 3,627 IU and increased the 1,000 IU hit rate from 17.5% to 23.8%.

However, R-squared values remained negative, indicating limited absolute predictive accuracy.

The researchers concluded that the framework may serve as an exploratory decision support tool for anticoagulation management during atrial fibrillation ablation.

However, they emphasised that external validation and prospective clinical studies are required before clinical implementation.

Reference

Sun Y et al. Machine learning-based prediction of heparin anticoagulation response and dose recommendation during atrial fibrillation catheter ablation. Sci Rep. 2026;DOI:10.1038/s41598-026-70874-y.

Featured image: Kiryl Lis on Adobe Stock