Combining computed tomography features of left atrial epicardial and pericoronary adipose tissue with the triglyceride-glucose index to predict the recurrence of atrial fibrillation after radiofrequency catheter ablation: a machine learning study

Abstract

Background: Radiofrequency catheter ablation (RFCA) represents an important treatment option for atrial fibrillation (AF); however, the recurrence rate following surgery is relatively high. This study aimed to predict the recurrence of AF after RFCA using interpretable machine learning models that combined the triglyceride-glucose (TyG) index and the quantification of left atrial epicardial and pericoronary adipose tissue.

Methods: This retrospective study included 325 patients with AF who underwent their first successful RFCA, among whom 79 had confirmed recurrence. The preoperative clinical data of patients were collected, the TyG index was calculated, and computed tomography (CT) image features were quantitatively measured. Multivariate Cox regression analysis was used to identify the independent risk factors for RFCA recurrence, and adjustments being made for various confounding factors. Post-hoc subgroup analysis was conducted to evaluate the predictive value of the TyG index for recurrence in different patient subgroups. Prediction models based on six machine learning algorithms were constructed. The optimal model's features were evaluated using Shapley additive explanations (SHAP).

Results: After adjustment were made for various confounding factors such as comorbidities of AF, Cox regression showed that the volume of left atrial epicardial adipose tissue (LA-EAT), LA-EAT attenuation, left circumflex coronary artery fat attenuation index (LCX-FAI), and the TyG index were independent risk factors for recurrence after RFCA (P<0.001). The support vector machine (SVM) model based on these combined indicators had the best predictive performance, with an area under the curve of 0.793 [95% confidence interval (CI): 0.782-0.805] in the validation set, while its accuracy and positive predictive value were 0.804 and 0.710, respectively. The predictive efficiency of the TyG index appeared to be independent of type 2 diabetes mellitus (T2DM) status (P_interaction=0.660).

Conclusions: The SVM model that integrated the TyG index and quantitative CT imaging variables demonstrated good predictive ability for post-RFCA recurrence in patients with AF. Furthermore, the TyG index appeared capable of predicting recurrence independently of T2DM status.

Keywords: Atrial fibrillation (AF); fat attenuation index (FAI); pericoronary adipose tissue (PCAT); radiofrequency catheter ablation (RFCA); triglyceride-glucose index (TyG index).

Figure 1

The flowchart of the study process. AF, atrial fibrillation; RFCA, radiofrequency catheter ablation; CCTA, coronary computed tomography angiography; PCI, percutaneous coronary intervention; XGBoost, extreme gradient boosting.

Figure 2

Measurement of the volume and attenuation of epicardial adipose tissue in the left atrium (LA-EATV and LA-EAT attenuation, respectively) and the volume and attenuation of the perivascular adipose tissue around the left anterior descending artery (LAD-FAI and LAD-V, respectively). The green area represents the LA-EAT region in the (A) axial, (B) coronal, (C) and sagittal planes. The purple area represents the perivascular adipose tissue around the LAD in the (D) axial, (E) coronal, (F) and sagittal planes. HU, Hounsfield unit; SD, standard deviation; LA-EATV, volume of the left atrial epicardial adipose tissue; LA-EAT, left atrium epicardial adipose tissue; LAD-FAI, left anterior descending artery fat attenuation index; LAD-V, adipose volume surrounding the left anterior descending artery.

Figure 3

A heatmap representation of the Spearman correlation matrix of the variables. Relevant correlations are color-coded based on the strength of the correlation. LA-EATV, volume of the left atrial epicardial adipose tissue; LA-EAT, left atrial epicardial adipose tissue; LCX-FAI, left circumflex coronary artery fat attenuation index; TyG, triglyceride-glucose; BUN, blood urea nitrogen; Scr, serum creatinine; GGT, γ-glutamyltransferase; TC, total cholesterol; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; CHA2DS2-VASc, congestive heart failure, hypertension, age ≥75 years (doubled), diabetes mellitus, prior stroke or transient ischemic attack (doubled), vascular disease, age 65–74 years, female; HAS-BLED, hypertension, abnormal renal/liver function, stroke, bleeding history or predisposition, labile INR, elderly, drugs/alcohol concomitantly; INR, international normalized ratio; BMI, body mass index; TIA, transient ischemic attack; hs-cTnT, high-sensitivity cardiac troponin T; HF, heart failure.

Figure 4

The decision and calibration curve. (A) The decision curve analysis of individual risk factors and combined factors in the multivariate Cox regression. (B) Calibration curve for the 12-month prediction of the combined factors. LA-EATV, volume of the left atrial epicardial adipose tissue; LA-EAT, left atrium epicardial adipose tissue; LCX-FAI, left circumflex coronary artery fat attenuation index; TyG, triglyceride-glucose; OS, overall survival.

Figure 5

Post-hoc subgroup analysis of the TyG index for the primary outcome. PersAF, persistent atrial fibrillation; PAF, paroxysmal atrial fibrillation; HTN, hypertension; T2DM, type 2 diabetes mellitus; TIA, transient ischemic attack; HF, heart failure; hs-cTnT, high-sensitivity cardiac troponin T; OR, odds ratio; CI, confidence interval; TyG, triglyceride-glucose.

Figure 6

ROC curves for the six machine learning models in the training set. AUC, area under the curve; CI, confidence interval; SVM, support vector machine; KNN, k-nearest neighbors; XGBoost, extreme gradient boosting; ROC, receiver operating characteristic.

Figure 7

SVM model explanation via the SHAP method. (A,B) Global model explanation. (C,D) Local model explanation. (A) Summary dot plot. The position of the point along the x-axis represents the actual SHAP value, indicating the impact of specific features on the model output for a particular patient. A higher SHAP value indicates a higher risk of recurrence. Features are distributed along the y-axis according to their importance, and their positions are determined by the average of the absolute SHAP values. The higher the position of a feature is, the more significant its impact on the model. (B) Feature importance bar chart. The width of the bar represents the average absolute SHAP value for each feature over all samples. (C,D) The waterfall plots of a patient without recurrent atrial fibrillation after RFCA and of a patient with recurrence, respectively. The SHAP values in each row quantify the magnitude and direction of the impact that each feature has on the prediction outcome. Features that contributed to an increase in the predicted risk of recurrence are displayed in red; features that contributed to a decrease in the predicted risk of recurrence are displayed in blue. TyG, triglyceride-glucose; LA-EATV, volume of the left atrial epicardial adipose tissue; LA-EAT, left atrium epicardial adipose tissue; LCX-FAI, left circumflex coronary artery fat attenuation index; SVM, support vector machine; SHAP, Shapley Additive explanations; RFCA, radiofrequency catheter ablation.