基于SEER数据库构建食管癌肺转移风险预测的机器学习模型OA
Development and validation of a machine learning model for predicting lung metastasis risk in esophageal cancer based on the SEER database
目的:基于美国监测、流行病学和最终结果(SEER)数据库构建食管癌肺转移机器学习预测模型,开发可交互在线风险计算器,为临床个体化风险分层提供量化依据.方法:提取 SEER 数据库 2010 年 1 月—2019 年 12 月确诊的 12 187 例食管癌患者资料,以肺转移为结局,纳入性别、年龄、种族等8项临床病理特征.按7:1.5:1.5进行结局分层随机抽样,划分为训练集、验证集、测试集.构建6种机器学习模型,依托Optuna框架完成超参数寻优,以受试者工作特征曲线下面积(AUC)结合准确率、灵敏度、特异度筛选最优模型,并开发配套在线计算器.结果:全队列肺转移发生率 6.32%(770/12 187).多因素 Logistic 回归证实男性、美洲印第安人/阿拉斯加原住民种族、食管鳞状细胞癌(ESCC)、肿瘤分化程度较低、T 分期升高及N分期升高为食管癌肺转移的独立危险因素.极端梯度提升(XGBoost)模型区分效能最优,验证集AUC 为0.7611(95%CI:0.7091~0.8131),测试集 AUC 为 0.7535(95%CI:0.7013~0.8057),准确率为 78.13%,灵敏度为 62.93%,特异度为 79.16%.Shapley 加性解释(SHAP)分析显示,T 分期、N 分期、组织学类型为核心预测因子.结论:本研究构建的XGBoost模型对食管癌肺转移具备良好区分能力,配套在线计算器可辅助高危人群分层管理,但现阶段仅适用于风险分层参考,其校准性能与临床价值仍需外部队列进一步验证优化.
Objective:To construct machine learning prediction models for lung metastasis in esophageal cancer based on the Surveillance,Epidemiology,and End Results(SEER)database,and to develop an interactive online risk calculator,and to provide a quantitative basis for individualized clinical risk stratification.Methods:Data on 12,187 patients diagnosed with esophageal cancer from January 2010 to December 2019 were extracted from the SEER database.With lung metastasis as the outcome,eight clinicopathological features including sex,age,and race were included.The dataset was divided into training,validation,and test sets at a ratio of 7:1.5:1.5 using outcome-stratified random sampling.Six machine learning models were constructed,with hyperparameter optimization performed using the Optuna framework.The optimal model was selected based on the area under the receiver operating characteristic curve(AUC)combined with accuracy,sensitivity,and specificity,and a corresponding online calculator was developed.Results:The overall incidence of lung metastasis in the entire cohort was 6.32%(770/12,187).Multivariate logistic regression confirmed that male sex,American Indian/Alaska Native race,esophageal squamous cell carcinoma(ESCC),lower tumor differentiation,higher T stage,and higher N stage were independent risk factors for lung metastasis.The extreme gradient boosting(XGBoost)model had the best discriminative performance,with an AUC of 0.7611(95%CI:0.7091-0.8131)in the validation set and an AUC of 0.7535(95%CI:0.7013-0.8057)in the test set,an accuracy of 78.13%,a sensitivity of 62.93%,and a specificity of 79.16%.Shapley additive explanations(SHAP)analysis revealed that T stage,N stage,and histological type were the most important predictive factors.Conclusion:The XGBoost model constructed in this study has favorable discriminative ability for lung metastasis in esophageal cancer,and the accompanying online calculator can assist in stratified management of high-risk populations.However,at the current stage,it is only suitable as a reference for risk stratification,and its calibration performance and clinical value still require further validation and optimization in external cohorts.
折涛;周银曦;王海慧;王斌;刘力强
空军军医大学唐都医院胸腔外科 陕西 西安 710038空军军医大学唐都医院胸腔外科 陕西 西安 710038空军军医大学研究生院研究生三大队七队 陕西 西安 710032西安医学院研究生工作部 陕西 西安 710021空军军医大学唐都医院胸腔外科 陕西 西安 710038
医药卫生
食管癌肺转移机器学习XGBoostSEER数据库风险预测
Esophageal CancerLung MetastasisMachine LearningXGBoostSEER DatabaseRisk Prediction
《机器人外科学杂志(中英文)》 2026 (9)
1650-1662,13
评论