Machine learning identifies esophageal luminal temperature patterns associated with thermal injury in catheter ablation for atrial fibrillation.

Abstract

Luminal esophageal temperature (LET) monitoring during atrial fibrillation (AF) ablation is widely used to reduce the incidence of endoscopically detected esophageal lesion (EDEL). We sought to assess whether specific patterns of LET variation are associated with EDEL. A high-fidelity multisensor probe was used to record LET in AF patients undergoing radiofrequency ablation (RFA) or cryoballoon ablation (CBA). Explainable machine learning and SHapley Additive exPlanations (SHAP) analysis were used to predict EDEL and assess feature importance. A total of 94 patients (38.3% persistent AF, 71.3% male, 72 RFA, and 22 CBA) were included. EDEL was detected in 11 patients (10 RFA and one CBA). In the RFA group, the highest LET recorded was similar between patients with and without EDEL (40.6 [40.1-41]°C vs. 40.2 [39.1-40.9]°C; p = .313), however, the rate of LET rise for the highest recorded peak was higher (0.08 [0.03-0.12]°C/s vs. 0.02 [0.01-0.05]°C/s; p = .033), and the area under the curve (AUC) for the highest peak was smaller (412.5 [206.8-634.1] vs. 588.6 [380.4-861.1]; p = .047) in patients who had EDEL. In case of CBA, the patient with EDEL had a faster LET decline (0.12 vs. 0.07 [0.02-0.14]°C/s), and a smaller AUC for the lowest trough (2491.3 vs. 2629.3 [1712.6-5283.2]). SHAP analysis revealed that a rate of LET change higher than 0.05°C/s and an AUC less than 600 were more predictive of EDEL in RFA. The rate of LET change and AUC for the recorded temperature predicted EDEL, whereas absolute peak temperatures did not.