Construction of a risk prediction model for postoperative atrial fibrillation in lung cancer patients based on multi-dimensional feature fusion and ensemble learning

Ziwei Gong¹Slamuguli Hella²Jing Shi²Xinya Liu^3*

• ¹Xinjiang Medical University College of Public Health, 823037, Urumqi, China
• ²Cancer Hospital of Xinjiang Medical University, Urumqi, China
• ³The Fifth Affiliated Hospital of Xinjiang Medical University, Urumqi, China

Surgery remains a cornerstone in lung cancer treatment, yet a subset of patients face high risks of recurrence or mortality postoperatively. Poor prognosis significantly shortens survival time, underscoring an urgent clinical need to accurately identify high-risk individuals. To address this, numerous studies have focused on constructing risk prediction models that integrate multidimensional data (clinical, pathological, and emerging biomarkers) to quantify postoperative adverse event probabilities, guiding personalized adjuvant therapy and enhancing follow-up management. To investigate risk factors for postoperative atrial fibrillation (POAF) in lung cancer patients and develop/validate a predictive model based on multi-dimensional feature fusion and ensemble learning. This retrospective cohort study analyzed 369 lung cancer patients undergoing surgical resection at Xinjiang Medical University Affiliated Tumor Hospital (2019– 2024). Univariate analysis screened potential risk factors, followed by multivariable logistic regression to confirm independent predictors.Nine machine learning algorithms were employed to build predictive models, among which the top three performers were selected for ensemble modeling via weighted averaging, resulting in the final risk prediction model. Multivariate analysis revealed three independent predictors of POAF: cardiac insufficiency (OR=64.55, 95%CI 2.41–1727.70), ventricular rate (OR=1.17, 95%CI 1.1–1.25), and elevated N-terminal pro-B-type natriuretic peptide (NT-proBNP, OR=1.005, 95%CI 1–1.009). The Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Machine (GBM) demonstrated the highest accuracy (ACC=0.9041, 0.9178, and 0.9178, respectively). The ensemble model srg-LCPOAF further improved ACC to 0.9452, significantly outperforming individual algorithms. This study is the first to integrate cardiopulmonary function, biomarkers, and surgical parameters into an ensemble model (srg-LCPOAF), providing evidence-based support for early intervention in high-risk POAF patients.