Construction and Validation of a Predictive Model for Postoperative Stent Occlusion in Patients Undergoing Iliac Vein Stenting Based on an Explainable Machine Learning Model

Yan Jiang¹Dan Nie²Lifeng Zhang²Xia Tang¹Haolu Li¹Long Li³He Hu Qiang⁴Yang Liu⁵Weijian Mao⁶Zhiwei Xiong⁷Chengyong Jin^1*

• ¹Ya 'an Traditional Chinese Medicine Hospital, Ya'an, China
• ²General Hospital of Western Theater Command, Chengdu, China
• ³The 945th Hospital of the Joint Logistics Support Force, PLA, Ya'an, China
• ⁴The Affiliated Hospital of Southwest Medical University, Luzhou, China
• ⁵Affiliated Hospital of Chengdu University of Traditional Chinese Medicine, Chengdu, China
• ⁶The Fourth People's Hospital of Zigong City, Zigong, China
• ⁷Panzhihua Central Hospital, Panzhihua, China

[Abstract] Objective: This study aims to develop an interpretable machine learning model for predicting post-operative iliac venous stent occlusion risk. Methods: Employing a retrospective cohort design, data from 826 patients across seven hospitals (January 2017-June 2024) were incorporated with stratified sampling into training (n=661) and test sets (n=165), ensuring no significant baseline characteristic differences (all P>0.05). An AutoML framework was constructed using the Improved Sequoia Optimization Algorithm (ISequoiaOA), integrated with LASSO feature selection and SHAP interpretability analysis; model evaluation incorporated six core metrics (including AUC/PR-AUC), calibration performance, and Decision Curve Analysis (DCA). Results: In independent testing-set validation, the AutoML model demonstrated superior robustness: ROC-AUC reached 0.9251 and PR-AUC 0.8712. Decision curve analysis confirmed significantly higher clinical net benefit across a wide threshold probability range (1%-87%) compared to conventional approaches, indicating exceptional generalizability. Calibration curves revealed the lowest Brier score (0.123) in the test set, further validating predictive accuracy. Outperforming comparative models (e.g., XGBoost [ROC-AUC 0.8203] and LightGBM [PR-AUC 0.7806]), AutoML dominated across all metrics including accuracy (0.7417) and F1-score (0.7559). Concurrently, SHAP analysis quantified critical feature contributions: Pathogenic triad (DVT+Cockett+PE); Hemodynamic thresholds (common femoral and external iliac vein recanalization rates both <70%); Stent geometric parameters (diameter >14mm/inferior vena cava segment length >20mm); With CRP>10mg/L and D-dimer>1.5mg/L coexistence elevating occlusion risk. Conclusion: The occlusion prediction system integrating AutoML with explainable AI successfully quantifies multi-level interactions, surpassing traditional predictive dimensions to provide evidence-based support for personalized anticoagulation and stent optimization.