Prediction Model for Thermal Lesions in Radiofrequency Ablation Based on an Artificial Neural Network

Radiofrequency cardiac ablation (RFCA) is a widely utilized treatment for atrial fibrillation (AF). However, its therapeutic efficacy can be compromised by either insufficient or excessive ablation, potentially leading to serious adverse effects. Therefore, precise control of the thermal lesion size generated during RFCA is critical for surgical success. Neural network is an implementation method of artificial intelligence, which has a strong ability to learn and adapt to complex data patterns, and shows significant application potential in the field of prediction. This study aimed to construct an artificial neural network (ANN) model capable of predicting the depth, width, and volume of ablation thermal lesions.

A two-branch ANN model was developed to predict lesion size on the basis of four key parameters: RF power, ablation duration, catheter‒tissue contact force, and contact angle. The training dataset for the model was derived from a finite element model of radiofrequency cardiac ablation. The model incorporated two types of RF power; catheter-tissue contact forces of 10 g, 20 g, 30 g, and 40 g; and contact angles of 0°, 45°, and 90°. The test dataset was obtained from ex vivo experiments conducted on a swine model, involving ten sets of experiments.

The finite element model effectively simulated the process of thermal lesion formation during RFCA, generating a substantial amount of effective training data. The ex vivo experiments provided reliable test data. The two-branch ANN model was able to predict the depth, width, and volume of thermal lesions, with errors of 0.1986 mm, 0.7891 mm, and 4.9384 mm³, respectively.

This study introduces a two-branch ANN model that serves as an efficient and reliable tool for predicting lesion size for RFCA. The two-branch ANN model proposed in this study enhances the model's ability to fit complex relationships through activation functions and nonlinear combination features. Compared with other models, it has superior predictive capabilities regarding the depth, width, and volume of ablation thermal lesions.