基于Lasso回归的颅内静脉和静脉窦血栓形成预后预测模型的构建与验证

Construction and validation of prognostic model for cerebral venous and dural sinus thrombosis based on Lasso regression

目的基于最小绝对收缩和选择算子(Lasso)回归构建颅内静脉和静脉窦血栓形成(CVST)预后的预测模型并验证其效能。方法回顾性分析2012年12月至2023年5月重庆医科大学附属第一医院神经外科收治的98例CVST患者的临床资料。患者的预后通过改良Rankin量表评分(mRS)评估。对比预后良好组(mRS<3分,81例)与预后不良组患者(mRS≥3分,17例)的基线资料、首发症状、入院时根据颅内静脉和硬脑膜窦血栓形成的国际研究建立的风险评分(ISCVT-RS)、入院时颅内静脉血栓形成分级量表(CVT-GS)、入院时格拉斯哥昏迷评分(GCS)、治疗方案、出院时mRS、影像学资料及实验室检查资料。应用二元logistic回归筛选出预测因子构建预后预测模型1,再应用Lasso回归筛选出预测因子构建预测模型2。运用受试者工作特征曲线(ROC)和曲线下面积(AUC)评估两种预测模型的效能。校准曲线用于评估预测模型的校准度。采用Bootstrap自抽样方法对模型进行内部验证。采用决策曲线分析(DCA)方法评估预测模型的临床获益。结果与预后良好组比较,预后不良组意识障碍、肢体乏力、脑实质出血、静脉性梗死、乙状窦血栓形成、横窦血栓形成及颈内静脉血栓形成的患者占比均较高,入院时ISCVT-RS、CVT-GS评分、GCS、外周血白细胞计数、中性粒细胞计数、血糖浓度、D-二聚体均较高,而淋巴细胞计数和血清白蛋白浓度均较低,差异均具有统计学意义(均P<0.05)。多因素logistic回归分析显示,肢体乏力(OR=9.40,95%CI:1.13~78.42)、外周血中性粒细胞计数升高(OR=1.70,95%CI:1.18~2.44)和血糖浓度升高(OR=1.36,95%CI:1.03~1.79)为CVST患者预后的危险因素,而外周血淋巴细胞计数升高(OR=0.05,95%CI:0.01~0.29)以及血清白蛋白浓度升高(OR=0.82,95%CI:0.70~0.96)为CVST患者预后的保护因素,基于以上5个预测因子构建预测模型1。预测模型2基于静脉性梗死、外周血中性粒细胞计数、淋巴细胞计数、血清白蛋白浓度、血糖浓度和D-二聚体6个预测因子构建。两种预测模型的AUC分别为0.95(95%CI:0.90~1.00)、0.94(95%CI:0.89~0.99)。校准曲线显示预测和实际观测之间具有良好的同质性。DCA显示,预测模型2具有更大的标准化净收益。结论基于静脉性梗死、外周血中性粒细胞计数、淋巴细胞计数、血清白蛋白浓度、血糖浓度和D-二聚体构建的Lasso回归预测模型2为CVST患者提供了更符合临床实际的预后预测,可以更加快捷、准确地评估CVST患者的预后,从而指导临床治疗。

Objective To construct a prediction model for the prognosis of cerebral venous and dural sinus thrombosis(CVST)based on least absolute shrinkage and selection operator(Lasso)regression and to verify its effectiveness.Methods The clinical data of 98 CVST patients admitted to the Neurosurgery Department of the First Affiliated Hospital of Chongqing Medical University from December 2012 to May 2023 were retrospectively analyzed.The patients’outcomes were assessed based on the modified Rankin Scale(mRS).We compared the baseline data,symptoms at onset,and symptoms on admission of patients in the good prognosis group(mRS<3 points,81 cases)and the poor prognosis group(mRS≥3 points,17 cases)based on the International Study on Cerebral Vein and Dural Sinus Thrombosis-Risk Score(ISCVT-RS),cerebral venous and dural sinus thrombosis grading scale(CVT-GS)on admission,Glosgow Coma Scale(GCS)score at admission,treatment plan,mRS score on discharge,imaging data and laboratory test data.Binary logistic regression was used to screen out predictors to build a prognosis prediction model 1,and then Lasso regression was used to screen out predictors to build a prediction model 2.The receiver operating characteristic curve(ROC)and area under the curve(AUC)were used to evaluate and compare the performance of the two prediction models.Calibration curves are used to evaluate the calibration of a predictive model.The model was internally verified using Bootstrap self-sampling method.Decision curve analysis(DCA)was used to evaluate the clinical benefit of predictive models.Results Compared with the good prognosis group,the poor prognosis group had higher proportions of patients with consciousness disorder,limb weakness,cerebral parenchymal hemorrhage,venous infarction,sigmoid sinus thrombosis,transverse sinus thrombosis and internal jugular vein thrombosis,and ISCVT-RS score on admission,CVT-GS score,GCS,white blood cell count,neutrophil count,blood glucose concentration,and D-dimer were all increased,while lymphocyte count and serum albumin concentration were all decreased,and the differences were all statistically significant(all P<0.05).Multivariate logistic regression analysis showed that limb weakness(OR=9.40,95%CI:1.13-78.42),elevated peripheral blood neutrophil count(OR=1.70,95%CI:1.18-2.44)and elevated blood glucose concentration(OR=1.36,95%CI:1.03-1.79)were risk factors for the prognosis of patients with CVST,while increased peripheral blood lymphocyte count(OR=0.05,95%CI:0.01-0.19)and serum albumin concentration(OR=0.82,95%CI:0.70-0.96)were protective factors for the prognosis of patients with CVST.The prediction model 1 was constructed based on the above five predictors.Prediction model 2 was constructed based on six predictors:venous infarction,peripheral blood neutrophil count,lymphocyte count,serum albumin concentration,blood glucose concentration and D-dimer.The AUCs of the two prediction models were 0.95(95%CI:0.90-1.00)and 0.94(95%CI:0.89-0.99)respectively.The calibration curve showed good homogeneity between predictions and actual observations.DCA showed that prediction model 2 had a larger normalized net benefit.Conclusion The Lasso regression model based on venous in farction,peripheral blood neutrophil caunt,lymphocyte count,serum albumin concentration,blood glucose concentration and D-dimer provides a more clinically realistic prognosis prediction for CVST patients,and can more quickly and accurately assess the prognosis of CVST patients,thereby guiding clinical treatment.