Lung lobectomy is the most common operation performed for lung cancer patients. Postoperative atrial fibrillation (POAF) is a common complication following lung lobectomy, with incidence rates ranging from 5.2% to 17.6% (1–4). Although POAF is often considered transient and self-limiting, it is significantly associated with prolonged hospital stays, increased risk of stroke, and higher mortality rates (5–7). Studies have shown that perioperative strategies, such as the administration of magnesium sulfate, preferential use of vasopressors over inotropes, avoidance of red cell transfusion, and video-assisted thoracoscopic surgery, may reduce POAF incidence (3, 8). Therefore, accurate preoperative risk stratification and tailored perioperative management are essential for improving outcomes and quality of life in high-risk lung cancer patients.

Several clinical characteristics, including advanced age, male sex, and procedural invasiveness, have been identified as key predictors of POAF (7, 9). Although these risk factors have been incorporated into predictive models, their performance remains suboptimal, with area under the receiver operating characteristic curve (AUC) values typically below 0.80 (3, 4, 10). Consequently, improving the identification of high-risk patients with POAF is warranted. Epicardial adipose tissue (EAT), a unique fat depot located between the myocardium and the visceral layer of the epicardium, has been suggested to play a role in the development and progression of atrial fibrillation (11). West et al. demonstrated that EAT volume could predict both in-hospital and long-term post-cardiac surgery atrial fibrillation (12).

Radiomics, which extracts high-throughput quantitative features from medical images, can provide disease-related information beyond traditional quantitative features such as volume and density (13, 14). Several studies have highlighted the effectiveness of EAT radiomics models in predicting POAF (15–17). The integration of deep learning (DL) radiomics features, particularly 3D DL features, enhances radiomics by capturing intricate structural patterns, thereby improving predictive model performance (18, 19). Therefore, integrating radiomics and DL analysis of EAT on preoperative CT images may offer a novel and robust approach for predicting POAF following lung lobectomy.

In this study, we aimed to develop and validate a DL radiomics model based on EAT to identify high-risk lung cancer patients for POAF following lung lobectomy. Additionally, we systematically evaluated the impact of various machine learning algorithms and resampling techniques on model performance.

In this study, we constructed a DL radiomics model through DL and radiomics analysis based on preoperative CT images of EAT and validated its ability to identify high-risk lung cancer patients with POAF following lung lobectomy. Compared to the clinical model, the predictive performance of the DL radiomics model, which integrates clinical features and handcrafted and DL radiomics signatures, was significantly improved. Furthermore, resampling techniques and machine learning algorithms did not significantly improve model performance.

The incidence of POAF in center 1 was 3.4%, lower than that reported in previous studies (1–4), which may be attributed to the implementation of enhanced recovery after surgery protocols (22) and a lower proportion of elderly patients. However, in the low-volume center, the incidence of POAF was 8.5%, significantly higher than that in center 1, demonstrating the good transportability of the DL radiomics model. In this study, we identified advanced age (P=0.014) and male sex (P=0.030) as independent risk factors. In the three cohorts, the AUC values of the clinical model were 0.737, 0.716, and 0.630, respectively, consistent with the results of previous studies (3, 4, 10). Furthermore, the Passman score, CHADS₂ score, and CHA₂DS₂-VASc score have been reported to identify high-risk patients with POAF (23–26). However, no significant differences were observed between patients with and without POAF in the three cohorts.

Artificial intelligence, particularly radiomics and deep learning, has rapidly emerged as a transformative field in translational oncology, serving as a critical bridge between medical imaging and precision medicine (27–29). In the context of lung cancer, radiomics has demonstrated broad application prospects in diagnosis, treatment response evaluation, and prognosis prediction (30–32). In recent years, the region of interest in radiomics has extended beyond the tumor itself. Radiomics features of EAT have been shown to effectively predict atrial fibrillation, mortality in acute pulmonary embolism, and myocardial ischemia (33–35). Our previous studies demonstrated that the handcrafted radiomics signatures of EAT achieved acceptable performance in predicting atrial fibrillation after pulmonary endarterectomy and coronary artery bypass grafting (16, 17). Therefore, radiomics analysis of EAT shows potential in identifying POAF in patients undergoing lung lobectomy.

DL radiomics features, particularly 3D DL features, can improve the performance of predictive models (18, 19). However, no relevant studies have been conducted on the DL radiomics analysis of EAT. In our study, we aimed to investigate the incremental value of handcrafted and DL radiomics signatures in POAF prediction. We extracted 1,130 handcrafted radiomics features and 512 DL radiomics features from each VOI, ultimately constructing handcrafted DL radiomics signatures after feature selection. The results of the Delong test and NRI analysis demonstrated that the DL radiomics model, integrating clinical features and handcrafted and DL radiomics signatures, performed satisfactorily in all cohorts, with significantly better predictive ability than the clinical model. However, the biological interpretation of radiomic features remains largely unclear, which limits a deeper understanding of their predictive mechanism and potential clinical translation. A relevant study by Mancio et al. in patients undergoing aortic valve replacement showed that a radiomics model combining EAT features and volume effectively discriminated preoperative POAF patients from those in sinus rhythm, achieving an AUC of 0.80 (95% CI: 0.68–0.92). Furthermore, proteomic analysis in that study revealed that POAF was associated with upregulation of inflammatory and prothrombotic proteins, along with downregulation of cardioprotective proteins with anti-inflammatory and antilipotoxic functions (15). Unfortunately, due to the retrospective nature of our study, we were unable to perform biologically validated assays (e.g., proteomics or histology) to explore the molecular correlates of the EAT radiomic features identified here. In our previous study, EAT segmentation was performed manually, which was time-consuming and labor-intensive. In this study, we employed the TIMESlice software to automatically segment EAT, significantly reducing the segmentation time and facilitating its clinical application.

Currently, imbalanced data remain a significant challenge for prediction models, often leading to prediction misclassification (36). Previous studies have demonstrated that resampling techniques and machine learning algorithms are effective methods for addressing the imbalance problem (37–39). In this study, we adopted the classical linear classification model constructed by the LR algorithm as the baseline model. Additionally, machine learning algorithms such as RF, SVM, and XGBoost, which have been reported to effectively address imbalance problems, were also selected (38, 40). Furthermore, considering the small number of patients with POAF, we employed four oversampling methods to balance the distribution of patients and ultimately constructed 20 different combination models. Currently, most studies only compare numerical values, lacking comprehensive statistical analysis. In our study, we employed the Friedman test and post hoc Nemenyi test to compare the predictive performance of different combinations of resampling techniques and machine learning algorithms, with results visualized using CD plots. The results indicated that neither machine learning algorithms nor resampling techniques significantly improved model performance in terms of AUC. G-mean and F-measure are useful evaluation metrics for imbalanced datasets. Unfortunately, we did not observe significant improvements in terms of G-mean and F-measure with machine learning algorithms and resampling techniques, contrary to previous studies (37). However, the study by Hubert S. Gabryś et al. found no advantage of resampling techniques in improving model performance (41). Due to the specific clinical challenges and the intrinsic characteristics of the data, there is no consensus on the optimal combination of machine learning algorithms and resampling techniques. Further research is required to comprehensively compare the performance of various resampling methods and machine learning algorithms.

This study has several limitations. First, due to the retrospective study design, some potential clinical features associated with POAF were not included, which may have limited the diagnostic performance. Second, it is important to note that the data for this study were retrospectively collected from two centers within a single healthcare network. While this provided internal consistency in imaging protocols, it may limit the generalizability of the findings. Furthermore, the sample size of the external validation cohort was limited, and future large-scale, multi-institutional prospective studies involving external validation cohorts from independent networks are essential to confirm the broad applicability of our model. Third, the relatively low incidence of POAF in our cohort may constrain the statistical power of the analysis and heighten the risk of model overfitting. While we employed resampling methods applied to alleviate class imbalance, further investigation into advanced feature selection and systematic hyperparameter tuning may offer valuable pathways for enhancing model generalizability and performance. Finally, the biological significance of radiomics features, particularly DL features, requires further investigation to enhance understanding and clinical application. The utilization of gradient-weighted class activation maps and perturbation-based explainable artificial intelligence techniques and the integration of radiomic features with other interpretable omics data, such as tissue metabolomics, all represent promising potential avenues for enhancing model interpretability (42–44).

In summary, the DL radiomics model based on preoperative EAT images effectively identified high-risk lung cancer patients with POAF following lung lobectomy and offers a novel tool for risk stratification. Neither machine learning algorithms nor resampling techniques significantly improved model performance in terms of AUC, G-mean, and F-measure.