Accurate estimation of lesion metrics in radiofrequency ablation using machine-learning model

Conventional parameters for estimating lesions in radiofrequency catheter ablation (RFCA), such as ablation energy (AE), contact force, and impedance variation, often yield suboptimal results. This study aimed to develop a machine learning model to improve the accuracy of lesion metric estimation in RFCA.

RF energies (30-50W) were applied to excised ventricular myocardium using RFCA with contact forces of 10g or 20g for durations between 10 and 180 seconds, with both perpendicular and parallel catheter orientations. Correlations between total AE, force-time integral, impedance-drop, and lesion metrics were evaluated and compared to machine learning model predictions, using eXtreme Gradient Boosting (XGBoost). The dataset was split for training (75%) and validation (25%). Feature importance for each lesion metric was also assessed.

A total of 1,142 ablations were analyzed. Total AE had the strongest correlation with max depth, max length, and volume (r² = 0.63, 0.50, 0.69), followed by force-time integral (r² = 0.54, 0.45, 0.62) and impedance-drop (r² = 0.33, 0.45, 0.31). Impedance drop was most strongly associated with surface area (r² = 0.48). The machine learning model accurately predicted lesion metrics: r² = 0.87 for depth, 0.82 for length, 0.86 for volume, and 0.69 for surface area, with low root mean square error values. Total AE and ablation duration were key predictors, with impedance drop contributing more to surface area and length predictions.

Machine learning using multiple RFCA parameters improves lesion metric predictions, enhancing lesion estimation beyond conventional metrics, potentially improving procedural guidance and safety.