Predictive value of artificial intelligence and radiomics for atrial fibrillation recurrence after catheter ablation for pulmonary vein isolation

Abstract

Background

Atrial fibrillation (AF) is a common arrhythmia disorder with a high recurrence rate after catheter ablation for pulmonary vein isolation (PVI). Improved preoperative evaluation strategies are needed to enhance prediction accuracy and optimize patient selection for ablation.

Materials and Methods

This study included 311 AF patients who underwent catheter ablation for PVI, stratified into recurrence (n = 160) and non-recurrence (n = 151) groups based on postoperative outcome. The deep learning model SwinUNETR was used to segment the heart region and extract the radiomics features of different areas of the heart. A risk prediction model was built using six machine learning methods. Finally, an interpretable analysis method based on radiomics was provided to capture the relationship between the radiological features of different heart structures and the recurrence of AF.

Results

A feature selection process was applied to determine the most predictive features, resulting in a set of 50 radiomics features and 33 clinical features. The average dice value of the deep learning heart segmentation model was 88.95%, and the area under the curve (AUC) value of the radiomics model for predicting the risk of AF recurrence after PVI was 0.74(95% confidence interval (CI) 0.54, 0.79). The AUC value of the fusion model integrating clinical laboratory indicators and radiomic features was 0.79(95% CI 0.69–0.84). According to the results of the interpretability analysis of the model, multiple radiomics features were determined to be significantly associated with AF recurrence.

Conclusion

This study presents a non-invasive model for predicting post-PVI recurrence and quantifies the contribution of specific cardiac structures to AF recurrence risk.

Keywords

Atrial fibrillation radiomics machine learning deep learning segmentation

Introduction

Atrial fibrillation (AF) is the most common persistent arrhythmia and has a significant impact on patients’ quality of life and survival rate.¹ With age, the incidence of AF gradually increases, and it is widespread among the elderly.^2–4 It not only seriously affects the quality of life of patients but also significantly increases the risk of stroke and heart failure, placing a heavy medical and economic burden on individuals and society. With the ageing of the population, the prevalence of AF is increasing, which makes it an important problem to be solved urgently in the field of cardiovascular diseases.^5–7 Catheter ablation is an important treatment for AF, and circumferential pulmonary vein isolation (PVI) is recognized as the cornerstone of its ablation strategy.^8,9 This ablation strategy aims to restore normal heart rate by destroying arrhythmogenic cardiomyocytes. This is the standard procedure for catheter ablation of AF. The continuous development and popularization of this technology have significantly improved the success rate of curing AF and reduced the risk of complications.^10–12 However, AF recurrence remains a common clinical challenge despite the established efficacy of PVI.^13–15 Studies have shown that the recurrence rate of arrhythmia within one year after PVI can be as high as 20%–50%.^16,17 Currently, the lack of effective predictive tools to identify patients who are unlikely to benefit from PVI further exacerbates the difficulty of clinical management, highlighting the urgent need for such predictive strategies.^18,19 AF recurrence prediction following PVI constitutes a pivotal issue in clinical decision-making. Although multimodal data integration studies have significantly enhanced their predictive accuracy, substantial challenges persist in this domain.²⁰ Moreover, there are very few studies that explore the relationship between cardiac structure and the recurrence of AF after PVI from an imaging perspective. Therefore, exploring more advanced preoperative evaluation strategies can effectively improve the accuracy of recurrence prediction, thereby optimizing patient selection and expanding the pool of eligible candidates who could benefit from this therapeutic intervention.

In recent years, researchers have conducted a large number of studies on the issue of AF recurrence after PVI, revealing a variety of potential factors that affect recurrence.^21–23 However, most of these studies have explored from the perspective of clinical test indicators, with relatively few imaging studies. What is more, most of these studies have remained at the individual level of imaging analysis.^24,25 Therefore, this study enrolled patients with AF who underwent catheter ablation for PVI and explored the relationship between recurrence and the imaging characteristics of the left atrium (LA), left atrial appendage (LAA), left ventricle (LV), right atrium (RA), right ventricle (RV), and pulmonary veins (PVs). In image data processing, deep learning technology was used to train the SwinUNETR segmentation model, which annotated a large volume of imaging data.²⁶ High-throughput image information of cardiac structures were extracted using radiomics methods. Finally, machine learning methods were employed to construct a recurrence prediction model to verify the predictive effect of imaging characteristics on AF recurrence after PVI. This study not only provides a potentially useful imaging-based tool for predicting AF recurrence after PVI but also explores the associations between the characteristics of various cardiac structures and post-PVI recurrence, offering valuable pathophysiological insights. In addition, this paper analyzes the imaging markers that affect the recurrence of AF after PVI from the imaging perspectives of shape, texture, and structure, and proves the relationship between imaging and the recurrence of AF, which is of great significance for the clinical treatment of AF.

Materials and methods

The experimental code for this study was written in Python (version 3.9.7) and R (version 4.2.3). The deep learning framework used was MONAI (version 1.3.0; https://monai.io/), a freely available, community-supported framework specifically designed for medical imaging research. It provides optimized implementations of domain-specific transforms, loss functions, and network architectures, which was chosen to enhance the reproducibility and efficiency of our deep learning segmentation pipeline. The model training was performed on a server with an NVIDIA A800 GPU. Model evaluation was performed using 10-fold cross-validation. The experimental process mainly includes deep learning to segment the region of interest (ROI) in the image, radiomics feature extraction and feature selection, machine learning modeling, and radiomics analysis. The workflow is shown in Figure 1.

Figure 1.

Workflow of deep learning and radiomics prediction models for patients with recurrence after pulmonary vein isolation (PVI).

Data

This study collected 1551 cases of patients with AF who had undergone catheter ablation for PVI in the Cardiac Pacing and Electrophysiology Department. Among all cases, there were 160 cases of AF recurrence after PVI, and 1391 cases of no recurrence. The definition of AF recurrence is that, 3 months after PVI, any episode of AF, atrial flutter, or atrial tachycardia lasting for more than 30 s is recorded through electrocardiogram monitoring. All patients in the non-recurrence group received at least 6 months of follow-up, during which no documented recurrence of arrhythmia was observed. Following the exclusion of cases with incomplete clinical data, the final cohort comprised 906 patients. Among them, 160 underwent a redo PVI procedure for AF recurrence, while 746 had no recurrence after the initial ablation. Due to the large difference in the number of cases and controls, and to avoid the impact of data imbalance and confounding bias on the statistical results, this paper uses propensity score matching (PSM) to reduce the selective bias caused by random assignment and ensure that the distributions of the two groups. Finally, 160 cases of AF recurrence after PVI and 151 cases of no recurrence were retained. Detailed inclusion/exclusion criteria and the enrollment process are shown in Figure 2. Propensity scores were estimated using a multivariate Logistic regression model, and the nearest neighbor method was used for matching. Gender, age, body mass index (BMI), body temperature, respiratory rate, smoking status, drinking status, diabetes, hypertension, and surgical history were used as variables for stratified matching, and matching was performed at a ratio of 1:1. The PSM was performed using the MatchIt package in R. The standardized mean difference (SMD) was used to measure the difference between the two groups, where SMD <0.1 was considered to indicate the strict balance of the matched variables.^27,28 The final matched population was used to evaluate the model's performance in predicting AF recurrence after PVI. All the enrolled patients underwent all preoperative examinations before the first PVI to ensure that the analyzed characteristics were not affected by the outcome of the first ablation and the progression of the disease.

Figure 2.

Inclusion/exclusion criteria and the enrollment process.

Deep learning segments the heart ROI

The left atrial PV computed tomography angiography (CTA) image data used in the study were all acquired using a conventional contrast-enhanced imaging protocol. Given that manual annotation of large-scale medical imaging datasets is both time-consuming and labor-intensive, it is crucial to develop an efficient, accurate, and automated segmentation process. Therefore, during the annotation process, the MONAI Label extension tool on the three-dimensional (3D) slicer (version 5.2.2; https://www.slicer.org/) was used to obtain the initial segmentation results of the various structures of the heart. Then, two experienced radiologists with 6 years of experience carefully examined and optimized the segmentation results layer by layer. The correction process strictly adheres to the anatomical boundary standards. Any disputes are resolved through negotiation or by arbitration by a third senior expert. The method used by MONAI Label is DeepEdit,²⁹ and 50 patients were included in the image annotation process. This method combines the efficiency of automation with the expertise of senior doctors, ensuring the accuracy and repeatability of the annotations. The annotated areas included the LA, LAA, LV, RA, RV, and PV. Of these, each PV begins at the LA opening and continues to the first branch. The ROI segmented in this study are the blood pools. Model training uses the MONAI deep learning framework. The training and test sets are divided 4:1. For SwinUNETR model training, images and masks were resampled with Spacingd, with pixels spaced to (1.5, 1.5, 2.0) mm. Bilinear interpolation was used for images, and nearest neighbor interpolation was used for labels to maintain category integrity. ScaleIntensityRanged was used to linearly scale the gray intensity value of the image to the range of [0,1] to optimize the model training. Spatial, intensity, and rotation transformation data enhancement strategies were introduced to improve the robustness of the model. The 3D image size during model training is 512 × 512 × 512, the batch size is 1, the patch resolution is 96 × 96 × 96, the initial learning rate is 0.0001, and 10,000 iterations are performed using the AdamW optimizer. The initial weights of the backbone network are pre-trained using the SwinUNETR model. Dice is used to evaluate the segmentation effect of the final model. Due to the diverse imaging modes in the heart area and the large differences in the size of the segmented areas, the six heart areas were modeled separately using the SwinUNETR network. In the inference stage, the trained SwinUNETR model was applied to unlabeled CTA images to generate segmentation masks for each structure of the heart in a fully automatic manner. The process is efficiently carried out in batch mode without manual intervention, which greatly improves the efficiency and consistency of labeling. The generated high-quality segmentation masks will be used as the basis for subsequent radiomics analysis to extract quantitative features to support the construction of recurrence risk prediction models.

Radiomics feature extraction and selection

Radiomics is a method to extract high-dimensional, quantitative features describing intensity, shape, and texture patterns from the ROI of medical images.³⁰ These features can analyze disease information more objectively and comprehensively, and play a potential role in promoting disease diagnosis, treatment selection, and prognosis evaluation. In this paper, the Pyradiomics package (version 3.0.1) was used to extract 111 radiomics features for each region. The extracted features mainly include first-order statistical features, shape features, gray-level co-occurrence matrix features, gray-level run-length matrix features, gray-level size zone matrix features (GLSZM), gray-level dependence matrix features (GLDM), and neighboring gray-tone difference matrix features of the original images. The ROI mask of the heart is segmented based on a deep learning model, and the radiomics features of the six regions are extracted separately. Before feature extraction, all images underwent gray-level discretization preprocessing to reduce noise sensitivity and enhance the repeatability of texture features. In this study, the fixed bin width method was adopted, and the bin width was set to 25. This method is based on the widely accepted recommendations in the field of radiomics, aiming to ensure the standardization of the gray scale for all cases, thereby reducing the bias caused by differences in scanning parameters.³¹ The parameters used in radiomics feature extraction are shown in Supplemental Table 1.

In order to improve the generalization ability of the model and avoid the influence of irrelevant features on the performance of the model, the SelectKBest method in scikit-learn was used for feature selection, in which the scoring function used F-statistics to measure the variance of features between different categories. The common radiomics features of the six regions of the heart are used as the evaluation data to determine the optimal number of feature selections. The evaluation models include XGBoost, random forest (RF), Bayesian, K-nearest neighbor (KNN), support vector machine (SVM), and Logistic regression. Among them, Logistic regression and Bayesian models are typical parametric models, which attempt to summarize the data using a fixed and limited number of parameters. While KNN and SVM belong to non-parametric models, their model structures and complexities are flexibly determined by the data itself, and no strong assumptions about the data distribution are made in advance. XGBoost and RF are both extremely powerful and commonly used ensemble learning models. RF is highly praised for its robustness and simplicity, making it an outstanding “ready-to-use” algorithm. XGBoost employs an improvement strategy and integrates advanced regularization and second-order gradient optimization, enabling it to often achieve higher prediction accuracy. However, this powerful learning ability comes with the risk of overfitting on noisy datasets and requires more careful hyperparameter tuning to achieve good generalization performance. In the radiomics feature selection process, based on the performance of the models, it was determined that 50 radiomics features would ultimately be retained for each region, and a total of 300 radiomics features would be retained for the six regions for subsequent model construction.

Risk prediction modeling

In order to establish the most robust prediction model for the recurrence of AF after PVI, XGBoost, RF, Bayesian, KNN, SVM, and Logistic regression were used to compare the final prediction performance of the model. Among them, XGBooST and RF are typical ensemble learning methods, Bayesian and Logistic regression constitute a comparison of probability models to evaluate the gain of the prediction effect by introducing prior knowledge, KNN is a typical non-parametric method, and measurement preserves the distribution characteristics of the data. SVM has unique advantages for small sample nonlinear problems. In the model construction, this study first incorporated 300 radiomics features into the development of the radiomics model. On this basis, the LASSO method was further used to screen the features of preoperative clinical indicators, and finally, 33 clinical variables with significant predictive value were retained. The above radiomics features were combined with the clinical indicators to jointly construct a fusion prediction model. Finally, in order to enhance the interpretability of the prediction model, we employed SHAP values to quantify the contribution of individual radiomics features to the model output. This approach enabled us to assess the importance of the features and reveal how specific anatomical regions affect the prediction results of AF recurrence after PVI. SHAP offers a unified method to quantify feature importance, facilitating the interpretation of anatomy-based radiomics features in predicting post-PVI outcomes.

Statistical analysis

PSM was used to control for potential confounding variables, and frequency distributions and central tendencies were used to describe the characteristics of the study population before and after matching. The dice index was used to measure the performance of the deep learning segmentation model, and a 95% confidence interval (CI) was used to assess the statistical uncertainty. The model results were evaluated using area under the curve (AUC), accuracy, sensitivity, specificity, and F1 value to assess model performance. All the data were standardized using z-score to transform the distribution of the data into a standard normal distribution.

Results

Baseline characteristics

After PSM, 311 cases were finally retained. The covariates in Table 1 are the results of comparisons before and after PSM. The results show that, except for surgical history, the covariates between the two groups have achieved strict balance (SMD < 0.1), and the covariate surgical history has basically achieved strict balance (SMD = 0.103).

Table 1.

Baseline characteristics of the eligible (before-match) and matched study population.

Characteristics	Eligible study population			Matched study population
Characteristics	Case group	Control group	SMD	Case group	Control group	SMD
Total	160	746		160	151
Sex			0.15			0.002
Female (%)	51 (31.9%)	291 (39.0%)		51 (31.9%)	48 (31.8%)
Male (%)	109 (68.1%)	455 (61.0%)		109 (68.1%)	103 (68.2%)
Age (mean (SD))	61.35 (10.97)	62.91 (11.27)	0.141	61.35 (10.97)	60.97 (11.37)	0.034
BMI			0.2			0.091
Normalweight:18.5–23.0	26 (16.2%)	170 (22.8%)		26 (16.2%)	27 (17.9%)
Overweight:23.0–27.5	75 (46.9%)	343 (46.0%)		75 (46.9%)	71 (47.0%)
Obese:>27.5	58 (36.2%)	223 (29.9%)		58 (36.2%)	51 (33.8%)
Underweight:<18.5	1 (0.6%)	10 (1.3%)		1 (0.6%)	2 (1.3%)
Body temperature	36.43 (0.19)	36.46 (0.22)	0.155	36.43 (0.19)	36.43 (0.16)	0.009
Respiratory rate(mean (SD))	18.71 (1.07)	18.70 (1.10)	0.009	18.71 (1.07)	18.81 (1.11)	0.089
Smoking			0.063			0.041
Yes (%)	41 (25.6%)	171 (22.9%)		41 (25.6%)	36 (23.8%)
No (%)	119 (74.4%)	575 (77.1%)		119 (74.4%)	115 (76.2%)
Drinking			0.082			0.03
Yes (%)	33 (20.6%)	130 (17.4%)		33 (20.6%)	33 (21.9%)
No (%)	127 (79.4%)	616 (82.6%)		127 (79.4%)	118 (78.1%)
Diabetes			0.041			0.06
Yes (%)	24 (15.0%)	123 (16.5%)		24 (15.0%)	26 (17.2%)
No (%)	136 (85.0%)	623 (83.5%)		136 (85.0%)	125 (82.8%)
Hypertension			0.145			0.048
Yes (%)	78 (48.8%)	310 (41.6%)		78 (48.8%)	70 (46.4%)
No (%)	82(51.3%)	436 (58.4%)		82 (51.2%)	81 (53.6%)
Operation history			0.178			0.103
Yes (%)	84 (52.5%)	457 (61.3%)		84 (52.5%)	87 (57.6%)
No (%)	76 (47.5%)	289 (38.7%)		76 (47.5%)	64 (42.4%)

BMI: body mass index; SMD: standardized mean difference.

Deep learning segment heart ROI

In the deep learning heart segmentation model construction, the training dataset uses 50 manually annotated 3D computed tomographic (CT) images of the LA and PV, and the manually annotated ROIs include LA, LAA, LV, RA, RV, and PV. Finally, six models are trained using the SwinUNETR network for subsequent labeling of a large number of heart images. Table 2 shows the dice similarity coefficients of the six models in the test set. The results show that the segmentation results of the six models can achieve good performance. The dice similarity coefficients for the LA, LAA, LV, RA, RV, and PV were 96.35%, 87.15%, 89.46%, 82.67%, 92.59%, and 85.45% respectively. The average dice similarity coefficient for the six models was 88.95%, with the best segmentation performance for the LA and RV, at over 90%. The segmentation results of different regions of the heart are shown in Supplemental Figure 1. The results show that the SwinUNETR model established in this paper can better segment the heart ROI, providing an efficient and accurate auxiliary labeling tool for subsequent tasks.

Table 2.

SwinUNETR deep learning segmentation model performance.

Cardiac regions	Dice (%)
LA	96.35
LAA	87.15
LV	89.46
RA	82.67
RV	92.59
PV	85.45

LA: left atrium; LAA: left atrial appendage; LV: left ventricle; RA: right atrium; RV: right ventricle; PV: pulmonary vein.

Feature extraction and selection

Among the 62 clinical indicators, 33 clinical indicators with significant influence were selected using LASSO. The clinical indicators are shown in Supplemental Table 2. Figure 3 shows the LASSO cross-validation curve. In the radiomics feature extraction and selection, 50 features related to AF recurrence were selected from 111 radiomics features. Figure 4 shows a schematic diagram of radiomics feature selection and model selection. According to the results, when the number of radiomics features is 96, the number of common features in the six regions is 50. At this point, the XGBoost model has the best performance among all models, with an AUC value of 73.75%. In addition, it can be seen from the results that when the number of features gradually increases, the performance of different models tends to change in different ways. Among them, the XGBoost, RF, and SVM models show a gradual upward trend, while the Bayesian, KNN, and Logistic models show a downward trend, and the stability of the model weakens.

Figure 3.

LASSO cross-validation curves.

Figure 4.

Schematic diagram of radiomics feature selection and model selection.

Risk prediction modeling

After feature selection, 33 clinical indicators and 300 imaging features representing cardiac structural characteristics were retained to construct a prediction model for the recurrence of AF after PVI. The performance of the radiomics prediction model and the fusion model with clinical indicators is shown in Table 3 and Table 4, respectively. The results of the radiomics model are showed that the AUC values of the XGBoost, RF, Bayesian, KNN, SVM, and Logistic models are 0.74 (95% CI 0.54, 0.79), 0.69 (95% CI 0.56, 0.73), 0.63 (95% CI 0.45, 0.75), 0.58 (95% CI 0.43, 0.71), 0.70 (95% CI 0.55, 0.76), and 0.63 (0.52, 0.76). The receiver operating characteristic (ROC) curves of the radiomics features in the six machine learning prediction models are shown in Figure 5A. The XGBoost model performed best, with an accuracy of 0.65 (95% CI 0.56, 0.70), sensitivity of 0.66 (95% CI 0.45, 0.91), specificity of 0.63 (95% CI 0.37, 0.82), F1 value 0.67 (95% CI 0.52, 0.73), and AUC value 0.74 (95% CI 0.54, 0.79). Table 4 and Figure 5B show the performance of the AF recurrence risk prediction model of the fusion model. The results showed that the performance of the fusion model was improved, and the XGBoost model has the best overall predictive performance and demonstrates a strong advantage over the performance of other models. Its AUC value increased to 0.79 (95% CI 0.69, 0.84). Among other indicators, the accuracy was 0.69 (95% CI 0.60, 0.80), the sensitivity was 0.75 (95% CI 0.58, 0.86), the specificity was 0.63 (95% CI 0.56, 0.85), and the F1 score was 0.71 (95% CI 0.59, 0.80).

Figure 5.

The receiver operating characteristic (ROC) curves of the machine learning prediction models. (A) ROC curve of radiomics model. (B) ROC curve of fusion model.

Table 3.

The performance of the radiomics prediction models.

Model	Metric (95% CI)
Model	Accuracy	Sensitivity	Specificity	F1	AUC
XGBoost	0.65(0.56–0.70)	0.66(0.45–0.91)	0.63(0.37–0.82)	0.67(0.52–0.73)	0.74(0.54–0.79)
RF	0.66(0.55–0.71)	0.66(0.56–0.81)	0.63(0.42–0.73)	0.64(0.58–0.74)	0.69(0.56–0.73)
Bayesian	0.60(0.47–0.71)	0.72(0.47–0.85)	0.47(0.3–0.75)	0.63(0.48–0.75)	0.63(0.45–0.75)
KNN	0.55(0.43–0.69)	0.63(0.51–0.74)	0.47(0.28–0.65)	0.58(0.48–0.71)	0.58(0.43–0.71)
SVM	0.61(0.54–0.68)	0.38(0.26–0.49)	0.87(0.8–0.93)	0.51(0.36–0.59)	0.70(0.55–0.76)
Logistic	0.60(0.49–0.70)	0.66(0.50–0.69)	0.52(0.47–0.84)	0.64(0.5–0.67)	0.63(0.52–0.76)

AUC: area under the curve; RF: random forest; KNN: K-nearest neighbor; SVM: support vector machine.

Table 4.

The performance of the fusion prediction models.

Model	Metric (95% CI)
Model	Accuracy	Sensitivity	Specificity	F1	AUC
XGBoost	0.69(0.60–0.80)	0.75(0.58–0.86)	0.63(0.56–0.85)	0.71(0.59–0.80)	0.79(0.69–0.84)
RF	0.66(0.51–0.77)	0.69(0.58–0.86)	0.63(0.34–0.75)	0.67(0.58–0.77)	0.70(0.58–0.81)
Bayesian	0.63(0.41–0.71)	0.63(0.40–0.80)	0.53(0.24–0.80)	0.62(0.42–0.73)	0.65(0.41–0.85)
KNN	0.58(0.47–0.70)	0.69(0.64–0.81)	0.44(0.25–0.69)	0.63(0.56–0.72)	0.58(0.48–0.72)
SVM	0.63(0.53–0.74)	0.31(0.20–0.53)	0.94(0.72–1.0)	0.45(0.30–0.66)	0.74(0.56–0.81)
Logistic	0.67(0.59–0.75)	0.69(0.58–0.80)	0.66(0.51–0.85)	0.68(0.59–0.76)	0.74(0.60–0.79)

AUC: area under the curve; RF: random forest; KNN: K-nearest neighbor; SVM: support vector machine.

In addition, to evaluate the predictive value of imaging features from different cardiac structures, we constructed XGBoost models for six regions (LA, LAA, LV, RA, RV, and PV), and each model used 50 selected radiomics features. The predictive performance of each region model is shown in Supplemental Figure 2. Among them, the AUC value of PV was the highest, at 0.65 (95% CI 0.55, 0.75), followed by RA, with an AUC value of 0.63 (95% CI 0.52, 0.77). The AUC values of the remaining regions were between 0.50 and 0.56. These results indicate that the radiomics features extracted from the PV and the RA may be more relevant for predicting the recurrence of AF after PVI compared to other cardiac regions.

SHAP analysis

The SHAP analysis method was employed to evaluate the contributions of the radiomics features in different cardiac regions. The most influential feature for each region, as determined by mean absolute SHAP value, was identified: First-order Total Energy (FTE) for the LA, Glszm Zone Percentage (GZP) for the LAA, Glszm Gray Level Variance (GSGLV) for the LV, Glszm Gray Level Nonuniformity (GSGLN) for the RA, Shape Maximum two-dimensional Diameter Slice (SMDS2D) for the RV, and Gldm Dependence Entropy (GDE) for the PV. As shown in Figure 6, the direction and magnitude of the feature contributions have been quantified. In the LA, the positive SHAP value of FTE is associated with an increased risk of recurrence, indicating a positive correlation with LA volume. In contrast, in the LAA, GZP is negatively correlated with recurrence, suggesting that less uniform texture may be associated with a higher risk. Similarly, GSGLV (LV) and SMDS2D (RV) are negatively correlated with recurrence, while GSGLN (RA) and GDE (PV) are positively correlated. Furthermore, in Supplemental Figure 3 are the mean SHAP values for all regions, where GSGLN (RA) and GDE (PV) have significantly greater contributions compared to other features, highlighting their potential role as significant imaging markers of recurrence risk after PVI. The detailed description of the radiomics features is presented in Supplemental Table 3.

Figure 6.

SHAP values were used to represent feature importance.

Discussion

This study enrolled 311 AF patients who underwent PVI to investigate the potential impact of cardiac structural and functional characteristics on recurrence. The main data included baseline clinical characteristics and CTA images. As summarized in Table 2, the used SwinUNETR architecture demonstrated outstanding performance in six of the heart sub-regions of interest, proving its ability to generate reliable segmentation results suitable for quantitative analysis. These results indicate that this model can serve as an effective and intelligent tool for assisting in the efficient annotation of cardiac CTA images. By reducing the burden on human labor and improving repeatability, this method provides a solid foundation for the subsequent extraction and analysis of radiomics features. Accurate segmentation of the heart structure is crucial for advancing the research on imaging-derived biomarkers related to AF recurrence after PVI, and ultimately will support more personalized treatment strategies and improve clinical outcomes.

In this study, 50 radiomics features that were most closely related to the recurrence of AF after PVI were selected for analysis. These features were subsequently incorporated into machine learning-based predictive modeling for each cardiac sub-region. Among the six evaluated models, the XGBoost algorithm demonstrated stronger predictive ability, achieving an AUC of 0.74 in the primary model and further improving to 0.79 in the fusion model. Notably, the performance of XGBoost aligns with a growing body of evidence supporting the efficacy of ensemble tree-based methods in handling high-dimensional radiomic data, often outperforming traditional statistical models.^32,33 AF recurrence after PVI is a multi-factor outcome influenced by complex pathophysiological processes. XGBoost is capable of capturing the complex nonlinear interactions between features, which might explain its robustness in predicting AF recurrence. This finding is consistent with recent research results, which also reported enhancing the utility of the algorithm in cardiovascular imaging applications.^34,35 Furthermore, integrating multimodal data into the model significantly improved the prediction performance, highlighting the value of combining structural information and radiomics information to enhance risk stratification. This approach offers a promising direction for developing more personalized prognostic tools after PVI. By providing an automated, image-based framework, our research contributes to the ongoing efforts to convert imaging biomarkers into clinical practice, potentially helping to identify high-risk patients who may benefit from more intensive monitoring or adjunctive treatment.

The SHAP analysis provided critical insights into the predictive mechanisms underlying the radiomics model by identifying the most influential features for each cardiac region and their directional association with risk of AF recurrence. It is notable that an increase in FTE for LA is associated with a higher recurrence risk, which is consistent with existing clinical knowledge that the LA volume variation is a known risk factor for AF recurrence.³⁶ Similarly, reduced GZP levels within the LAA indicate a more heterogeneous texture and are associated with a higher risk of recurrence, which is consistent with findings linking structural complexity to arrhythmogenicity.³⁷ Furthermore, the SHAP results also revealed novel potential biomarkers. High gray level non-uniformity (GSGLN) in the RA and elevated dependence entropy (GDE) in the PV were strongly and positively correlated with recurrence risk. These features likely reflect increased tissue heterogeneity and structural diversity, potentially indicating pro-arrhythmic structural alterations that facilitate AF recurrence. Interestingly, our analysis revealed some unexpected associations. The higher protection of GSGLV in the LV, indicating greater heterogeneity of texture, may paradoxically reflect a state of adaptive remodeling or compensatory mechanisms. In contrast, lower GSGLV in LV implies higher homogeneity, associated with a higher risk of recurrence. This higher uniformity may indicate more severe, diffuse pathological changes, such as homogeneous fibrosis or myocardial dilatation, that alter the electrophysiological basis and promote AF recurrence after PVI, even if the texture appears homogeneous. Similarly, the larger the SMDS2D in the RV, the more irregular the shape and the greater the degree of dilation, indicating a protective effect. This means that a more regular and well-enclosed RV is actually a risk factor. A regular and more rounded RV is a well-recognized imaging marker of adverse remodeling and elevated pulmonary artery pressure.^38,39 RV pressure overload directly increases the pressure in the RA and alters the blood flow within the heart, thereby providing favorable conditions for the onset and persistence of AF. Moreover, the mean SHAP values highlighted GSGLN (RA) and GDE (PV) as the strongest contributors to the model, suggesting that these features may serve as robust, non-invasive imaging markers for recurrence risk. This underscores radiomics’ capacity to extract pathologically relevant information beyond conventional imaging and supports the integration of interpretable artificial intelligence into clinical tools for improved post-PVI patient stratification.

Limitations

This study has potential limitations. They are therefore subject to biases and confounding that may have influenced our model estimates. Not all patients with AF recurrence after the post-ablation blanking period proceed to a repeat PVI, due to factors such as patient preference, comorbidities, or overall clinical assessment. Therefore, the prediction model constructed in this study may underestimate the risk of recurrence. Moreover, we excluded patients with incomplete data. This may lead to selection bias, for instance, these patients with missing data might have more challenging clinical conditions, poorer prognosis, or different comorbidity characteristics. Therefore, the predictive model and conclusions we have derived may be more applicable to the patient group with complete data. However, when applying them to a broader population, caution is necessary.

There are also limitations from the clinical scan protocols of CTA imaging in the study, with inherent differences in contrast media protocols and scan parameters. Although we have adopted image standardization, this difference still leads to changes in image contrast and noise levels, which remain a significant challenge in cardiac CT radiomics analysis. Beyond these technical variations, a fundamental methodological limitation lies in the source of our imaging features. In the study, the imaging features were directly derived from the blood pool. Although their morphological and texture features might indirectly reflect potential atrial structural remodeling, fibrosis, and hemodynamic changes, the arrhythmogenic matrix is located in the myocardial wall rather than the blood pool. Therefore, directly analyzing the imaging features of the myocardial wall may provide more comprehensive and direct predictive information. Future research will focus on developing advanced image segmentation algorithms capable of precisely segmenting the thin layers of the myocardial wall and extracting richer radiomics features, such as wavelet transform from them.

The study was limited in that, because it was a retrospective analysis, detailed classification of all patients could not be consistently performed in the medical records, preventing us from stratified analysis by AF type or reporting its distribution. It is also not possible to assess whether the predictive performance of our model differs according to AF subtype. Future prospective studies with rigorously collected AF phenotypic data are warranted to validate and generalize our findings. Moreover, factors such as the small size of the test set and the lack of external validation have a certain impact on the generalization ability of the model. In the future, we will further collect multi-center data to verify the predictive effect of the model in clinical practice. These limitations require further clarification.

Conclusion

This study demonstrates that the radiomics features extracted from cardiac CT images can effectively predict the recurrence of AF after PVI. The most informative features for prediction were predominantly concentrated in the RA and PV. Our findings confirm the significant relationship between structural and textural characteristics of these cardiac regions and post-PVI outcomes, providing a non-invasive tool for individualized risk assessment and clinical decision-making.

Supplemental Material

sj-docx-1-dhj-10.1177_20552076251393276 - Supplemental material for Predictive value of artificial intelligence and radiomics for atrial fibrillation recurrence after catheter ablation for pulmonary vein isolation

Supplemental material, sj-docx-1-dhj-10.1177_20552076251393276 for Predictive value of artificial intelligence and radiomics for atrial fibrillation recurrence after catheter ablation for pulmonary vein isolation by Guoxiang Ma, Shuai Shang, Suixia Zhang, Hui Liu, Hulin Li, Baopeng Tang, Yanmei Lu and Kai Wang in DIGITAL HEALTH

Footnotes

Abbreviations

ORCID iDs

Guoxiang Ma

Shuai Shang

Hulin Li

Ethics approval

This study was approved by the Ethics Committee of the First Affiliated Hospital of Xinjiang Medical University (IRB No. 231124-05). As this study was a retrospective study, the requirement for informed consent was waived. However, we protect personal privacy and sensitive information in our research to ensure that personally identifiable data is not compromised, and follow the principles of academic integrity and privacy protection in our research.

Contributorship

All authors contributed to the study conception and design. Data collection was performed by Guoxiang Ma, Kai Wang, Shuai Shang, Yanmei Lu, and Baopeng Tang. The analyses were performed by Guoxiang Ma, Kai Wang, and Shuai Shang. The first draft of the manuscript was written by Guoxiang Ma and Kai Wang. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. This research was supported by a grant from Kai Wang.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by The Key Research and Development Program of Xinjiang Uygur Autonomous Region, The Fundamental Research Operating Funds for Universities Directly Under the Xinjiang Uygur Autonomous Region (grant numbers 2022B03023-2, XJEDU2025J063).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

Data generated or analyzed during the study are available from the corresponding author by request.

Code availability statement

The Python and R code developed for this study is publicly available on GitHub at .

Supplemental material

Supplemental material for this article is available online.

References

Kukendrarajah

Ahmad

Carrington

, et al. External electrical and pharmacological cardioversion for atrial fibrillation, atrial flutter, or atrial tachycardias: a network meta-analysis. Cochrane Database Syst Rev 2024; 6: CD013255.

Lim

Lee

Choi

, et al. Combination of early rhythm control and healthy lifestyle on the risk of stroke in elderly patients with new-onset atrial fibrillation: a nationwide population-based cohort study. Front Cardiovasc Med 2024; 11: 1346414.

Parks

Frankel

Kim

, et al. Management of atrial fibrillation in older adults. Br Med J 2024; 386: e076246.

Rosas Diaz

Troy

Kaplinskiy

, et al. Assessment and management of atrial fibrillation in older adults with frailty. Geriatrics (Basel) 2024; 9: 50.

Newman

O'Meara

Bohm

, et al. Implications of atrial fibrillation for guideline-directed therapy in patients with heart Failure: JACC State-of-the-Art Review. J Am Coll Cardiol 2024; 83: 932–950.

Yang

Heckbert

Ding

, et al. Prevalence of subclinical atrial fibrillation in heart failure with preserved ejection fraction. JACC Heart Fail 2024; 12: 492–504.

Elkholey

Asad

ZUA

Shehata

, et al. Association between atrial fibrillation and heart failure patient reported outcomes across the ejection fraction spectrum. Am Heart J 2024; 273: 61–71.

Haissaguerre

. Spontaneous initiation of atrial fibrillation by ectopic beats originating in the pulmonary veins. Heart Rhythm 2024; 21: 713–714.

Calkins

Hindricks

Cappato

, et al. 2017 HRS/EHRA/ECAS/APHRS/SOLAECE Expert consensus statement on catheter and surgical ablation of atrial fibrillation: executive summary. Europace 2018; 20: 157–208.

10.

Tzeis

Gerstenfeld

Kalman

, et al. 2024 European Heart Rhythm Association/Heart Rhythm Society/Asia Pacific Heart Rhythm Society/Latin American Heart Rhythm Society expert consensus statement on catheter and surgical ablation of atrial fibrillation. Europace 2024; 26: euae043.

11.

Akerstrom

Hutter

Charitakis

, et al. Association between catheter ablation of atrial fibrillation and mortality or stroke. Heart 2024; 110: 163–169.

12.

AlTurki

Joza

Bernier

, et al. Catheter ablation of atrial fibrillation: current and evolving indications. Can J Cardiol 2020; 36: 1685–1689.

13.

Benali

Khairy

Hammache

, et al. Procedure-Related complications of catheter ablation for atrial fibrillation. J Am Coll Cardiol 2023; 81: 2089–2099.

14.

Guo

. Risk and protective factors of recurrence after catheter ablation for atrial fibrillation. Rev Cardiovasc Med 2024; 25: 81.

15.

Zei

. Left atrial calcification after catheter ablation for atrial fibrillation: the Medusa atrium? JACC Clin Electrophysiol 2023; 9: 1118–1120.

16.

Erhard

Metzner

Fink

. Late arrhythmia recurrence after atrial fibrillation ablation: incidence, mechanisms and clinical implications. Herzschrittmacherther Elektrophysiol 2022; 33: 71–76.

17.

Darby

. Recurrent atrial fibrillation after catheter ablation: considerations for repeat ablation and strategies to optimize success. J Atr Fibrillation 2016; 9: 1427.

18.

Stauffer

Knecht

Badertscher

, et al. Repeat catheter ablation after very late recurrence of atrial fibrillation after pulmonary vein isolation. Europace 2024; 26: euae096.

19.

Cheng

Wang

Dai

, et al. Recurrent atrial flutter after radiofrequency ablation for atrial fibrillation in a patient with hypertrophic obstructive cardiomyopathy undergoing repetitive transcoronary ablation of septal hypertrophy. BMC Cardiovasc Disord 2024; 24: 532.

20.

Zhao

Guo

Lei

. Risk factors and nomogram prediction model for recurrence after radiofrequency ablation in patients with persistent atrial fibrillation. BMC Cardiovasc Disord 2025; 25: 578.

21.

Brahier

Kochi

Huang

, et al. Machine learning of cardiac anatomy and the risk of new-onset atrial fibrillation after TAVR. JACC Clin Electrophysiol 2024; 10: 1873–1884.

22.

Giomi

Bernardini

Perini

, et al. Clinical impact of smoking on atrial fibrillation recurrence after pulmonary vein isolation. Int J Cardiol 2024; 413: 132342.

23.

Fei

Zhao

, et al. Factors influencing early recurrence of atrial fibrillation among elderly patients following radiofrequency catheter ablation and the impact of different antiarrhythmic regimens. Front Med (Lausanne) 2024; 11: 1393208.

24.

Valles-Colomer

Rubio Forcada

Soto-Iglesias

, et al. Reproducibility analysis of the computerized tomography angiography-derived left atrial wall thickness maps. J Interv Card Electrophysiol 2023; 66: 1045–1055.

25.

Tonko

Lee

Mannakkara

, et al. Structural phenotyping in atrial fibrillation with combined cardiac CT and atrial MRI: identifying and differentiating individual structural remodelling types in AF. J Cardiovasc Electrophysiol 2024; 35: 1788–1796.

26.

Hatamizadeh

Nath

Tang

, et al. Swin unetr: swin transformers for semantic segmentation of brain tumors in MRI images. In: International MICCAI brainlesion workshop. Cham: Springer International Publishing, 2021, pp.272–284.

27.

Flury

Riedwyl

. Standard distance in univariate and multivariate analysis. Am Stat 1986; 40: 249–251.

28.

Austin

. Using the standardized difference to compare the prevalence of a binary Variable between two groups in observational research. Commun Stat 2009; 38: 1228–1234.

29.

Diaz-Pinto

Mehta

Alle

, et al. Deepedit: deep editable learning for interactive segmentation of 3D medical images. In: MICCAI Workshop on data augmentation, labelling, and imperfections. Cham: Springer Nature Switzerland, 2022, pp.11–21.

30.

Lambin

Rios-Velazquez

Leijenaar

, et al. Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012; 48: 441–446.

31.

Leijenaar

Nalbantov

Carvalho

, et al. The effect of SUV discretization in quantitative FDG-PET radiomics: the need for standardized methodology in tumor texture analysis. Sci Rep 2015; 5: 11075.

32.

Yan

Niimi

Matsunaga

, et al. Preoperatively predicting survival outcome for clinical stage IA pure-solid non-small cell lung cancer by radiomics-based machine learning. J Thorac Cardiovasc Surg 2025; 169: 254–266.e9.

33.

Prinzi

Orlando

Gaglio

, et al. Interpretable radiomic signature for breast microcalcification detection and classification. J Imaging Inform Med 2024; 37: 1038–1053.

34.

Mancio

Pashakhanloo

El-Rewaidy

, et al. Machine learning phenotyping of scarred myocardium from cine in hypertrophic cardiomyopathy. Eur Heart J Cardiovasc Imaging 2022; 23: 532–542.

35.

Kang

Suh

Han

, et al. Performance of prediction models for diagnosing severe aortic stenosis based on aortic valve calcium on cardiac computed tomography: incorporation of radiomics and machine learning. Korean J Radiol 2021; 22: 334–343.

36.

Lee

Hwang

, et al. Lower pulmonary vein-to-left atrium volume ratio predicts poor rhythm outcome after atrial fibrillation catheter ablation. Front Cardiovasc Med 2022; 9: 934168.

37.

Di Biase

Natale

Romero

. Thrombogenic and arrhythmogenic roles of the left atrial appendage in atrial fibrillation. Circulation 2018; 138: 2036–2050.

38.

Mendiola

da Silva Gonçalves Bos

Leichter

, et al. Right ventricular architectural remodeling and functional adaptation in pulmonary hypertension. Circ Heart Fail 2023; 16: e009768.

39.

Rako

Kremer

Yogeswaran

, et al. Adaptive versus maladaptive right ventricular remodelling. ESC Heart Fail 2023; 10: 762–775.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.74 MB