基于机器学习的多中心胰腺癌术后复发风险预测模型建立和外部验证.pdfVIP

受试者工作特征曲线下面积为0.733；k-近邻算法模型预测术后2年复发风险的准

确率及受试者工作特征曲线下面积最高，分别为75.9%和0.661。

研究结论：

基于机器学习，本研究通过整合胰腺导管腺癌患者围手术期的临床病理特征建

立术后复发风险预测模型，并验证了相应模型。这些预测模型将会为胰腺导管腺

癌患者个性化术后监测方案的制定提供指导。

关键词：胰腺导管腺癌；机器学习；根治性切除；复发风险；预测模型

Multi-institutionalDevelopmentandExternalValidationof

machinelearning-basedmodelstopredictrelapseriskof

pancreaticductaladenocarcinomaafterradicalresection

ZhejiangUniversitySchoolofMedicineGeneralsurgery

PostgraduateLouJianyao

SupervisorWuYulian

Abstract

Background:

Pancreaticductaladenocarcinoma(PDAC)isakindofhighlyfatalcancer.Surgical

resectionistheonlypotentiallycurativetreatmentforPDACandthesurvivalof

patientsafterradicalresectioniscloselyrelatedtorelapse.Weaimedtodevelopmodels

topredicttheriskofrelapseusingmachinelearningmethodsbasedonmultipleclinical

parameters.

Methods:

Datawerecollectedandanalysedof262PDACpatientswhounderwentradical

resectionat3institutionsbetweenJanuary2013andDecember2017.Wedeveloped

severalpredictivemodelstopredict1-and2-yearrelapseriskusingmachinelearning

approachesincludinglogisticregression(LR),randomforest(RF),supportingvector

machine(SVM),gradientboostingmachine(GBM),k-nearestneighbor(KNN)and

neuralnetwork(NN).Meanwhile,wecomparedthesepredictivemodelsandfiguredout

bestmodelforeachterm.

Results:

Machinelearningtechniquesweresuperiortoconventionalregression-based

analysesinpredictingriskofrelapseofPDACafterradicalresection.Amongthem,the

Conclusion:

Bymachinelearning,thisstudyhasdevelopedandvalidatedcomprehe