[目的]基于上海市社区人群低剂量CT(low dose computed tomography,LDCT)肺癌筛查项目数据,构建肺癌风险预测模型,为我国LDCT筛查高危人群的界定及后续追踪提供科学依据.[方法]选取2013年8月至2017年12月参与上海市闵行区肺癌LDCT筛查的合格人群24 530人,收集LDCT筛查信息、肺癌风险评估问卷信息、肺癌发病信息.采用Cox比例风险回归法共构建了两套风险预测模型:基本模型(n=24 530)纳入性别、筛查年龄、吸烟史、家族史、是否检出结节;LDCT筛查模型(n=3 649)纳入吸烟史、家族史、筛查是否阳性、结节性质、结节大小.将人群按7∶3的比例随机分为训练集和验证集,使用受试者工作特征曲线的曲线下面积(area under the curve,AUC)评价区分度,绘制校准曲线评估模型的校准度,利用十倍交叉验证方法进行预测模型的内部验证.[结果]24 530名研究对象的结节检出率为17.5％,LDCT筛查阳性率为12.0％,中位结节大小6.0 mm(P25,P75:4.0,10.0 mm).在中位随访9.8年(P25,P75:8.4,11.4年)期间,共发现新发肺癌病例503例(男性342例,女性161例).训练集中,基本模型预测1、3、5年肺癌发生风险的AUC分别为0.883、0.800 和 0.828,LDCT筛查模型的AUC分别为0.826、0.803和0.804,模型区分能力均较好.基本模型和LDCT筛查模型的校准曲线显示,模型拟合度均良好.十倍交叉验证结果显示,基本模型的平均AUC为0.783,标准误为0.012;LDCT筛查模型的平均AUC为0.796,标准误为0.017;模型预测效果均稳定.[结论]该研究建立了基于社区人群LDCT筛查的肺癌风险预测模型,其在判别能力和预测准确性方面具有良好的性能,有助于肺癌LDCT筛查高危个体的识别及筛查后健康管理.
Construction of Lung Cancer Risk Prediction Model Based on Low Dose Computed Tomography Screening in Shanghai Com-munity Population
[Objective]To develop a risk predictive model for lung cancer based on a community low dose computed tomography(LDCT)screening program.[Methods]A total of 24 530 eligible participants of the organized lung cancer screening program in Minhang District of Shanghai dur-ing August 2013 and December 2017 were included.Data of LDCT results,questionnaire-based risk assessment,and incidence of lung cancer were collected and two risk prediction models were developed.The basic model(n=24 530)included gender,age at screening,smoking,family his-tory of lung cancer,and nodule detection status;and the LDCT screening model(n=3 649)in-cluded smoking,family history of lung cancer,results of LDCT(positive/negative),feature and size of detected nodules.The study population was randomly divided into training(70％)and vali-dation(30％)sets.The area under the receiver operating characteristic curve(AUC)was used to evaluate differentiation,the calibration curves were profiled to assess the calibration of the mod-els,and the ten-fold cross-validation method was applied for internal validation of the predictive models.[Results]Among 24 530 eligible participants,lung nodules were detected by LDCT in 17.5％subjects,with a positive rate of 12％.The median diameter of the nodules was 6.0 mm[P25,P75:4.0,10.0 mm].During a median of 9.8 years of follow-up(P25,P75:8.4,11.4 years),503 subjects(342 male and 161 female)were diagnosed with lung cancer.In the training set,the AUCs of the basic model were 0.883,0.800 and 0.828,respectively for predicting lung cancer risk within 1-,3-and 5-year,while those for the LDCT screening model were 0.826,0.803 and 0.804,respectively.Both models exhibited good discriminatory ability and calibration.Ten-fold cross-validation results revealed an average AUC of 0.783 with a standard error of 0.012 for the basic model,and an average AUC of 0.796 with a standard error of 0.017 for the LDCT screening model.[Conclusion]The risk predictive models constructed in this study perform well in predict-ing lung cancer risk,which have great potential for more targeted offers for LDCT screening in the populations and for further health management after screening.
community populationlow dosecomputed tomographyscreeninglung can-cerrisk predictionShanghai
NETL
NSTL
万方数据
