Discussion
Atrial fibrillation is the most common sustained arrhythmia that increases with age and presents with a wide spectrum of symptoms and severity, including reentry theory and focal agitation[17,18]. Despite shorter ablation times, faster balloon cooling, and longer thawing times, ablation of atrial fibrillation with a second-generation cryoballoon is associated with higher success rates of pulmonary venous dissociation, acute and long-term PV isolation rates are high, AE rates are similar, and atrial fibrillation is absent[19]. Therefore, the data we collected and collated were all studies and analyses of AF patients after cryoablation.
In this study, prognostic variables were not simply analyzed by clinical index (CLI); imaging data was used to construct radioactivity index (RSI). With the joint action of clinical and radioactivity indicators, the C-index and ROC curve results of the formed risk score model showed a significant improvement in the prediction effect on the whole set, training set, validation set 1, or validation set 2. Meyre et al. (2019)[20] performed Cox regression analysis adjusted for risk factors for routine admission using clinical data to calculate the risk ratio (HR) and obtained a C-statistic of 0.64 (95% CI: 0.61-0.66). Similarly, Peng et al. (2019)[21] constructed radiomic features based on features extracted from PET and CT images in the training set to predict disease-free survival (DFS). To this end, the radioactivity indexes (RSI) of AF patients after cryoablation were integrated into this paper, and the original clinical indicators (CLI) were combined to form the comprehensive risk score (TCRS). From an effects standpoint, the combination of RSI and CLI indicators showed a significant increase compared to the AUC and C-index of a single indicator. In the ROC curve of the training set, the AUC of RSI was 0.942 and the AUC of CLI was 0.862, but the AUC of TCRS obtained by linearly weighting the two indicators was 0.955. Similarly, in the validation set 1, the TCRS index formed by TCRS showed superior prognostic performance. In order to improve the accuracy of the model, a second validation set was set up to witness the TCRS constructed from the nonlinear survival model. The results of the C-index also confirm the accuracy of the model, such as RSI (C-index: 0.8894; 95%CI: 0.8166-0.9621), CLI (C-index: 0.8431; 95%CI: 0.7466-0.9395), and TCRS (C-index: 0.9072; 95%CI: 0.8281-0.9864) in validation set 2.
Researchers have traditionally used Lasso Cox analysis to perform survival analyses for various diseases. Bieging et al. (2018)[22] selected only shape parameters using the Lasso method and factor analysis, adding them to a Cox regression model that included multiple clinical parameters and LA fibrosis (C-index: 0.68-0.72). Other scholars have used the nomogram method to analyze the postoperative prognosis of patients with AF. Zhou et al. (2021)[23] explored the risk factors for recurrence of atrial fibrillation (AF) in patients after radio frequency ablation and constructed a targeted nomogram prediction model (AUC=0.852). Dong et al. (2022)[24] used the least absolute shrinkage and selection operator regression for variable screening and a multi-variable Cox survival model for nomogram development, obtaining an AUC of between 0.855 and 0.863 in the development and validation cohorts. Our study combines the nonlinear part of machine learning and the linear part of the Cox model to obtain a nonlinear proportional risk survival model, enabling the construction of the risk score of AF patients to divide them into high- and low-risk groups. First, from the perspective of ROC curve, TCRS constructed from non-linear survival models had better prognosis whether in the training set (AUC: 0.955 vs 0.664), validation set 1 (AUC: 0.920 vs 0.548) or validation set 2 (AUC: 0.945 vs 0.591). The closer the AUC is to 1, the better the prediction. Second, the results were more credible through the auxiliary verification of C-index, such as RSI (C-index: 0.9125; 95%CI: 0.8596-0.9653), CLI (C-index: 0.8479; 95%CI: 0.7656-0.9303) and TCRS (C-index: 0.9454; 95%CI: 0.9219-0.9689).
The innovation of this paper lies in the construction of an optimal non-linear survival model using a variety of machine learning models, and the demonstration of its superior prognostic performance. While Katzman et al. (2016)[15] proposed a combination of nonlinear and linear models in theory, their prognostic effect lacked empirical research. In contrast, the prognostic performance of the nonlinear proportional hazard survival model developed in this study was significantly improved. From the model correction diagram, the predicted value and the real value are basically in the same straight line, and there is only a small error (Fig. 5). It can also be seen from the forest plot that the higher the TCRS value of the indicator we constructed, the higher the risk of recurrence, and the P value is less than 0.001, indicating that the result is significant (Fig. 7). The constructed risk score index was then used to divide patients into high and low risk groups using X-tile software. The Kaplan-Meier curve showed a significant difference between the two groups, and the log-rank test was less than 0.001, indicating clear differences in survival time and survival rate between the high and low risk groups. Both the training set and the validation set showed that there were significant differences in the high and low analysis groups divided by TCRS (Fig. 6). This predicted non-linear proportional hazard survival model can serve as a quantitative means to assess the high and low risk of atrial fibrillation recurrence in cryoablation patients.