Identifying High-Risk Features Associated with Limited Overall Survival in Esophageal Cancer Patients with Vascular Invasion Using Machine Learning

Background Vascular invasion is a critical pathological feature associated with poor prognosis in esophageal cancer, yet individualized risk prediction in this subgroup remains limited. Methods We retrospectively analyzed clinical data from 318 esophageal cancer patients with confirmed vascular invasion who underwent surgical resection at Shandong University Qilu Hospital between January 2019 and December 2022. Eight machine learning models were constructed using clinical, pathological, and laboratory features. The Gradient Boosting Machine (GBM) model was selected based on superior performance in discrimination, calibration, and decision curve analysis. Internal validation and survival stratification were conducted to assess robustness and clinical utility. Results Eight variables were identified as independent prognostic factors: nerve invasion, invasion of the fibrous outer membrane, T stage, N stage, BMI, white blood cell count, squamous cell carcinoma antigen level, and total number of lymph nodes dissected. The GBM model achieved the highest time-dependent AUCs (1-year: 0.987, 2-year: 0.971, 3-year: 0.976) and demonstrated consistent calibration and net clinical benefit. Survival analysis based on GBM risk scores revealed significant stratification between risk groups (p < 0.001). Conclusion Our GBM-based model provides accurate and interpretable prognostic predictions for esophageal cancer patients with vascular invasion. It offers valuable guidance for individualized clinical decision-making and warrants further validation in multicenter prospective cohorts.