摘要
Background Machine learning(ML)is increasingly being used to predict prognosis.This study aimed to establish and compare various ML models for predicting five-year all-cause mortality in patients with mixed gastric cancer,focusing on identifying significant prognostic features.Methods We developed five ML models using a follow-up database of mixed gastric cancer patients.The model with the highest performance,the Light Gradient Boosting Machine(LGBM),was selected to predict five-year allcause mortality.The log-rank test was employed to evaluate the divergence of Kaplan-Meier curves.Results The LGBM model excelled in predicting and stratifying patient risk,achieving an average AUC of 0.887(SD:0.079),an F1-score of 0.790(SD:0.060),and a Brier score of 0.152(SD:0.036)in the training set.In the test set,it achieved a maximum AUC of 0.835(95%CI:0.823-0.848)and a minimum Brier score of 0.172(95% CI:0.166-0.177).Higher machine learning scores correlated with increased mortality risk.Key prognostic factors included pN stage,pTNM stage,tumor diameter,surgical radicality,Fbg levels,and biomarkers CA199,CA724,and CA125.Conclusion This study successfully applied ML techniques to predict five-year all-cause mortality in patients with mixed gastric cancer.The LGBM model proved effective in both prediction and risk stratification,highlighting critical prognostic factors that clinicians should consider in patient management.
基金
supported by Nn10 program of Harbin Medical University Cancer Hospital,China(No.Nn10 PY 2017-03).
