基于机器学习算法的胰腺鳞状细胞癌预后模型的构建与验证

Developing and validating nomograms for predicting survival in patients with pancreatic squamous cell carcinoma based upon machine learning

目的胰腺鳞状细胞癌(pancreatic squamous cell carcinoma,PSCC)预后差,缺乏个体化的预后预测工具。研究通过SEER数据库中的大样本真实世界数据,基于机器学习算法,构建PSCC病人预后列线图,旨在精准化、个体化评价PSCC病人的预后,为临床决策制定提供参考。方法提取SEER数据库中2000~2019年期间经病理学确诊的PSCC病人的临床资料,以7∶3的比例随机划分为训练集和验证集,采用多因素Cox比例风险模型、LASSO回归模型和随机生存森林模型筛选变量,构建预测3、6、12个月肿瘤特异性生存期(cancer-specific survival,CSS)和总生存期(overall survival,OS)的Nomogram模型,利用一致性指数、受试者操作特征曲线操作、校准曲线、生存曲线、决策曲线分析对模型进行验证和评估。结果本研究共纳入367例病人,其中训练集256例,验证集111例。训练集和验证集病人的中位随访时间分别为3(1,7)个月和2(1,8)个月。两组间的基线特征均衡(均P>0.05)。多因素Cox比例风险模型显示:肿瘤大小、T分期、M分期、手术、化疗是OS和CSS的独立影响因素(均P<0.05)。LASSO回归模型显示:M分期、手术、化疗和OS、CSS相关。随机生存森林模型显示,影响OS的重要性评分前4位变量分别为化疗、M分期、手术和年龄,而影响CSS的重要性评分前4位变量分别为化疗、M分期、手术和肿瘤大小。基于这些因素所构建的Nomogram用于预测病人3、6个月的OS和CSS。验证结果表明:对于OS,训练集和验证集中一致性指数分别为0.753(95%CI:0.720~0.790)和0.723(95%CI:0.660~0.780);对于CSS,两者分别为0.749(95%CI:0.720~0.780)和0.721(95%CI:0.660~0.780)。受试者操作特征曲线操作显示:对于3个月OS,训练集和验证集的曲线下面积(AUC)分别为79.8%和75.9%;对于6个月OS,训练集和验证集的AUC为78.9%和76.8%;对于12个月OS,训练集和验证集的AUC为78.7%和77.5%;对于3个月CSS,训练集和验证集的AUC为79.3%和76.3%;对于6个月CSS,训练集和验证集的AUC为78.6%和76.9%;对于12个月CSS,训练集和验证集的AUC为77.8%和77.4%。校准曲线均靠近理想的45°参考线,表现出良好的一致性。结论年龄、M分期、肿瘤部位、手术和化疗是病人预后的独立影响因素。研究构建的Nomogram预测模型具有良好的预测价值,有利于临床对PSCC病人选择个性化治疗。

Objective Pancreatic Squamous Cell Carcinoma(PSCC)has a poor prognosis and it lacks individualized prognostic tools.This study aimed to construct prognostic nomograms for PSCC patients based upon machine learning and using large-scale real-world data from the database of SEER,provide precise and individualized prognostic assessments and offer valuable references for clinical decision-making.Methods From 2000 to 2019,the relevant clinical data of 367 pathologically diagnosed PSCC patients were extracted from the database of SEER.They were randomized by a ratio of 7∶3 into training(n=256)and verification(n=111)sets.Multivariate Cox proportional hazard model,LASSO regression and random survival forest model were utilized for identifying independent prognostic factors for patient survival.These factors were utilized for constructing nomograms for predicting cancer specific survival(CSS)and total survival(OS)at Month 3/6.Subsequently,the models were internally and externally validated in training and validation sets by concordance index(C-index),receiver operating characteristic(ROC)and calibration curves for assessing their accuracy and predictive capacity.Results The median follow-up period in training and verification sets were 3(1,7)and 2(1,8)month.Baseline profiles were comparable between two groups(all P>0.05).Multivariate Cox proportional hazard model analysis indicated that tumor size,M/N stage,surgery and chemotherapy were independent influencing factors for OS/CSS.LASSO regression analysis revealed that M stage,surgery and chemotherapy were associated with OS/CSS.For OS,top four scoring variables for via random survival forest model were chemotherapy,M stage,surgery and age;For CSS,chemotherapy,M stage,surgery and tumor size.Nomograms for predicting OS/CSS at Month 3/6 were developed based upon these independent prognostic factors.Validation results showed that C-index for OS in training and verification sets were 0.753(95%CI:0.720-0.790)and 0.723(95%CI:0.660-0.780)and for CSS 0.749(95%CI:0.720-0.780)and 0.721(95%CI:0.660-0.780).ROC curve analysis indicated that AUC values for OS in training and verification sets were 79.8%and 75.9%at Month 3,78.9%and 76.8%at Month 6 and 78.7%and 77.5%at Month 12;for CSS,79.3%and 76.3%at Month 3,78.6%and 76.9%at Month 6 and 77.4%and 78.4%at Month 12 respectively.Calibration curve analysis demonstrated a decent agreement between predicted and actual OS/CSS.Both were closely situated near ideal 45°reference line,demonstrating a high consistency.Conclusion Age,M stage,tumor size,surgery and chemotherapy are independent prognostic factors for PSCC patients.And the above constructed nomogram prediction models exhibit favorable predictive value and facilitate personalized therapeutics for PSCC patients in clinical practices.