Results of external validation for anatomical site recognition of ablation sites in atrial fibrillation using the CARTONET R12 model

Abstract

Abstract Background CARTONET® is a cloud-based management system designed for retrospective analysis of ablation procedures using the CARTO system. Leveraging a proprietary machine learning model, it automates procedural information analysis. However, the efficacy and accuracy of the CARTONET R12 model's automated location identification remain unverified. Methods This study involved a total of 396 cases. Pulmonary vein isolation (PVI), including cavo-tricuspid isthmus (CTI) ablation, was performed in 313 patients (PVI group). PVI and additional ablation, such as a Box isolation and superior vena cava isolation (SVCI), was performed in 83 patients (PVI+ group). We compared the automated location identification of R12 model and the location identification of arrhythmia experts. Results In this study, 29,422 points were analyzed. The sensitivity of the pulmonary vein carina was low in both PVI group and PVI+ group. Excluding the inferior, anterior, and ridge segmentations, PVI group demonstrated significantly higher sensitivity compared to PVI + group. Regarding the ablation lines for box isolation, both the roof line and post line showed results below 30%. The pulmonary vein carina showed a low PPV in both PVI group and PVI+ group. In segmentations excluding the anterior, PVI group demonstrated significantly higher PPV compared to the group PVI+. Regarding the ablation lines for box isolation, both the roof line and post line showed results below 50%.The R12 model couldn't auto-identify CTI and SVC. Conclusion Excluding the carina, it can be said that in cases where PVI was performed, both PPV and sensitivity demonstrate high accuracy. Conversely, in cases where Box isolation was performed, accuracy decreases due to the low discrimination ability between ablation information near the roof, posterior wall, and inferior wall of the pulmonary veins and the ablation lines of Box isolation. Regarding the right atrium, it was found that the R12 model is entirely incapable of discriminating SVC and CTI, indicating a need for revisions in the current model.The segmentation of anatomyResults of the positive predict value