子痫前期不良妊娠结局的影响因素和风险预测模型建立与验证

Analysis of influencing factors of adverse pregnancy outcomes in preeclampsia and establishment and validation of risk prediction model

摘要:

目的： 探讨子痫前期（preeclampsia，PE）不良妊娠结局的影响因素并构建风险预测模型。方法： 回顾性选取2018年6月至2020年12月苏州大学附属苏州九院和南通大学附属医院收治的PE患者作为建模集（模型开发），以患者入院48 h内是否出现不良妊娠结局划分为不良组、非不良组。对2组各项指标进行单因素筛选，再行多因素logistic回归分析PE不良妊娠结局的影响因素。基于筛选结果，利用R语言构建风险预测列线图模型。采用受试者工作特征（receiver operating characteristic，ＲＯＣ）曲线分析和拟合优度偏差性检验评价模型的表现。采用Bootstrap法（自抽样法）验证并制作校准图；采用决策曲线评价模型的临床获益率。选取2021年1月至2022年3月苏州大学附属苏州九院收治的PE患者作为验证集。结果： 共纳入381例PE患者作为建模集，其中126例发生不良妊娠结局，255例未发生不良妊娠结局；共纳入102例PE患者作为验证集，其中34例发生不良妊娠结局，68例未发生不良妊娠结局。Logistic回归分析显示：入院孕周越小（OR=2.672，95%CI=1.495～5.153）、临床症状数目越多（OR=2.643，95%CI=1.394～4.917）、24 h蛋白尿定量越高（OR=3.662，95%CI=1.982～7.604）、血小板计数越低（OR=2.396，95%CI=1.307～4.653）、D-二聚体越高（OR=2.929，95%CI=1.728～5.843）、miR-21表达量越高（OR=4.302，95%CI=2.426～9.185）均是PE患者发生不良妊娠结局的影响因素（P<0.05）。基于上述6个因素构建风险预测列线图模型，模型的ROC曲线下面积（area under the curve，AUC）为0.912（95%CI=0.864～0.956），最佳截断值（阈概率）为0.32，此时的灵敏度和特异度分别0.889、0.845；拟合优度偏差性检验（χ²=4.214，P=0.837）；偏差校准曲线平均绝对误差为0.013。验证集的AUC为0.904（95%CI=0.842～0.936），灵敏度和特异度分别0.874、0.823；拟合优度偏差性检验（χ²=3.720，P=0.729）；偏差校准曲线平均绝对误差为0.021。当决策曲线中阈概率值设为32.0%，建模集和验证集的临床获益率分别为69%、76%。结论： PE患者的不良妊娠结局与入院孕周、临床症状数目、24 h蛋白尿定量、血小板计数、D-二聚体、miR-21表达量相关，以此构建风险预测列线图模型具有较高的预测效能。

Abstract:

Objective: To explore the influencing factors of adverse pregnancy outcomes in preeclampsia (PE) and build a risk prediction model. Methods: The retrospective selection of PE patients admitted to the Ninth Affiliated Hospital of Suzhou University and the Affiliated Hospital of Nantong University from June 2018 to December 2020 as the modeling set (model development). According to whether the patients had adverse pregnancy outcomes within 48 hours of admission, they were divided into adverse group and nonadverse group. Single factor screening was performed for each index of the two groups. Multivariate logistic regression analysis was performed to analyze the influencing factors of adverse pregnancy outcomes in PE. Based on the filtered results, R language was used to build a risk prediction nomogram model. Receiver operating characteristic curve analysis and goodness-of-fit bias tests were used to evaluate model performance. Bootstrap method (self-sampling method) was used to verify and make calibration chart. A decision curve was used to evaluate the clinical benefit rate of the model. The PE patients admitted to Suzhou Ninth Hospital affiliated to Suzhou University from January 2021 to March 2022 were taken as the validation set. Results: A total of 381 PE patients were included as the modeling set. Among them, 126 cases had adverse pregnancy outcomes, and 255 cases had no adverse pregnancy outcomes. A total of 102 PE patients were included as the validation set, among which, 34 had adverse pregnancy outcomes, and 68 had no adverse pregnancy outcomes. Logistic regression analysis showed: that the smaller the gestational age at admission (OR=2.672, 95%CI=1.495-5.153) , the more the number of clinical symptoms (OR=2.643, 95%CI=1.394-4.917) , and the higher the 24-hour proteinuria (OR=3.662, 95%CI=1.982-7.604) , lower platelet count (OR=2.396, 95%CI=1.307-4.653) , higher D-dimer (OR=2.929, 95%CI=1.728-5.843) , higher miR-21 expression levels (OR=4.302, 95%CI=2.426-9.185) were all influencing factors of adverse pregnancy outcomes in PE patients (P<0.05). Based on the above six factors, a risk prediction nomogram model was constructed. The area under the curve (AUC) of the model was 0.912 (95%CI=0.864-0.956) , the optimal cutoff value (the threshold probability) was 0.32, and the sensitivity and specificity at this time were 0.889 and 0.845, respectively; goodness-of-fit bias test ( χ²=4.214, P=0.837) ; the mean absolute error of the bias calibration curve was 0.013. The AUC of the validation set was 0.904 (95%CI=0.842-0.936) ; the sensitivity and specificity were 0.874 and 0.823, respectively; goodness-of-fit bias test (χ²=3.720, P=0.729) ; The mean absolute error of the bias calibration curve was 0.021. When the threshold probability value in the decision curve was set to 32.0%, clinical benefit rates for the modeling set and validation set were 69% and 76%, respectively. Conclusion: Adverse pregnancy outcomes in PE patients are related to gestational age at admission, number of clinical symptoms, 24-h proteinuria, platelet count, D-dimer, and miR-21 expression. Risk prediction nomogram model constructed in this way has high prediction performance.