摘要
目的探讨联合临床病理特征和人工智能(AI)算法构建的胃癌生存预测模型的效能及临床意义。方法选取2016年6月至2018年5月丽水市人民医院就诊的200例胃癌患者的病理信息,筛选与生存结局高度相关的指标。采用10倍交叉验证法将200例患者以2∶8的比例分为建模队列40例和验证队列160例。与生存结局高度相关的指标采用随机梯度提升(gbm)、广义线性模型(glmnet)、逻辑回归模型(plr)、径向基函数核支持向量机(svmRadial)、贝叶斯模型(naive_bayes)和随机森林模型(ranger)6种AI算法在建模队列中构建胃癌生存预测模型,并在验证队列中验证模型的预测效能。结果200例患者中存活组109例,死亡组91例。肿瘤最大径、淋巴结转移、肿瘤位置、神经浸润和TNM分期在存活组和死亡组的分布差异均有统计学意义(均P<0.05),且与生存结局均相关(均P<0.05)。ROC曲线显示单项指标预测患者生存的AUC均>0.500。综合比较6种算法的多维度考量指标发现,基于svmRadial算法下的5项病理特征组合生存预测模型综合效能最佳,AUC为0.817,灵敏度为0.762,特异度为0.833,准确度为0.795。在验证队列中svmRadial算法下的5项病理特征组合生存预测模型AUC为0.624。结论AI技术能够有效提升5项病理特征联合检测的预测效能,从多维度综合分析患者情况,具有优异的辅助潜能,可为改善胃癌患者的预后管理提供研究思路和理论基础。
Objective To construct a survival prediction model for gastric cancer(GC)with artificial intelligence(AI)algorithms and to verify its performance efficiency.Methods Clinicopathological information of 200 patients with gastric cancer treated in Lishui People's Hospital from June 2016 to May 2018 was analyzed,and the indicators highly correlated with survival outcome were screened out.Patients were divided into a training cohort(40 cases)and a testing cohort(160 cases)at a ratio of 2∶8 by 10-fold cross-validation.GC survival prediction models were constructed based on pathological indexes in the training cohort by 6 AI algorithms,including stochastic gradient boosting(gbm),generalized linear model(glmnet),penalized logistic regression(plr),support vector machines with radial basis function Kernel(svmRadial),naive_bayes and random forest(ranger);and the prediction efficiency of models was verified in the testing cohort.Results Of the 200 patients,109 survived and 91 died.There were significant differences in tumor size,lymph node metastasis,tumor location,nerve invasion,and TNM stage between the survival group and the death group(all P<0.05),and there indexes were correlated with the survival outcomes of patients(all P<0.05).ROC curve showed that the AUC of the single index for predicting the survival of patients was all>0.500.Comprehensive comparison of the multi-dimensional considerations of the six algorithms showed that the 5MP survival prediction model based on the svmRadial algorithm had the best comprehensive performance with an AUC of 0.817,sensitivity of 0.762,specificity of 0.833 and accuracy of 0.795.The AUC of the survival prediction model based on the combination of 5 clinicopathological features constructed with the svmRadial algorithm in the validation queue was 0.624.Conclusion The prediction model based on the combination of clinicopathological features and constructed by AI technology has excellent auxiliary potential in predicting the prognosis of gastric cancer patients.
作者
周璐青
廖旭慧
曹学全
蒋苏甜
ZHOU Luqing;LIAO Xuhui;CAO Xuequan;JIANG Sutian(Graduate School,Zhejiang Chinese Medical University,Hangzhou 310053,China;不详)
出处
《浙江医学》
CAS
2024年第3期262-268,共7页
Zhejiang Medical Journal
基金
丽水市科技计划项目(2023GYX41)。
关键词
胃癌
病理特征
人工智能
生存预测
Gastric cancer
Clinicopathological features
Artificial intelligence
Survival prediction