Abstract：
OBJECTIVE: To evaluate the performance of an artificial intelligence (AI) and machine learning (ML) model for first-trimester screening for pre-eclampsia in a large Asian population.
METHODS: This was a secondary analysis of a multicenter prospective cohort study in 10 935 participants with singleton pregnancies attending for routine pregnancy care at 11-13+6 weeks of gestation in seven regions in Asia between December 2016 and June 2018. We applied the AI+ML model for the first-trimester prediction of preterm pre-eclampsia (<37 weeks), term pre-eclampsia (≥37 weeks), and any pre-eclampsia, which was derived and tested in a cohort of pregnant participants in the UK (Model 1). This model comprises maternal factors with measurements of mean arterial pressure, uterine artery pulsatility index, and serum placental growth factor (PlGF). The model was further retrained with adjustments for analyzers used for biochemical testing (Model 2). Discrimination was assessed by area under the receiver operating characteristic curve (AUC). The Delong test was used to compare the AUC of Model 1, Model 2, and the Fetal Medicine Foundation (FMF) competing risk model.
RESULTS: The predictive performance of Model 1 was significantly lower than that of the FMF competing risk model in the prediction of preterm pre-eclampsia (0.82, 95% confidence interval [CI] 0.77-0.87 vs. 0.86, 95% CI 0.811-0.91, P = 0.019), term pre-eclampsia (0.75, 95% CI 0.71-0.80 vs. 0.79, 95% CI 0.75-0.83, P = 0.006), and any pre-eclampsia (0.78, 95% CI 0.74-0.81 vs. 0.82, 95% CI 0.79-0.84, P < 0.001). Following the retraining of the data with adjustments for the PlGF analyzers, the performance of Model 2 for predicting preterm pre-eclampsia, term pre-eclampsia, and any pre-eclampsia was improved with the AUC values increased to 0.84 (95% CI 0.80-0.89), 0.77 (95% CI 0.73-0.81), and 0.80 (95% CI 0.76-0.83), respectively. There were no differences in AUCs between Model 2 and the FMF competing risk model in the prediction of preterm pre-eclampsia (P = 0.135) and term pre-eclampsia (P = 0.084). However, Model 2 was inferior to the FMF competing risk model in predicting any pre-eclampsia (P = 0.024).
CONCLUSIONS: This study has demonstrated that following adjustment for the biochemical marker analyzers, the predictive performance of the AI+ML prediction model for pre-eclampsia in the first trimester was comparable to that of the FMF competing risk model in an Asian population.
METHODS: This was a secondary analysis of a multicenter prospective cohort study in 10 935 participants with singleton pregnancies attending for routine pregnancy care at 11-13+6 weeks of gestation in seven regions in Asia between December 2016 and June 2018. We applied the AI+ML model for the first-trimester prediction of preterm pre-eclampsia (<37 weeks), term pre-eclampsia (≥37 weeks), and any pre-eclampsia, which was derived and tested in a cohort of pregnant participants in the UK (Model 1). This model comprises maternal factors with measurements of mean arterial pressure, uterine artery pulsatility index, and serum placental growth factor (PlGF). The model was further retrained with adjustments for analyzers used for biochemical testing (Model 2). Discrimination was assessed by area under the receiver operating characteristic curve (AUC). The Delong test was used to compare the AUC of Model 1, Model 2, and the Fetal Medicine Foundation (FMF) competing risk model.
RESULTS: The predictive performance of Model 1 was significantly lower than that of the FMF competing risk model in the prediction of preterm pre-eclampsia (0.82, 95% confidence interval [CI] 0.77-0.87 vs. 0.86, 95% CI 0.811-0.91, P = 0.019), term pre-eclampsia (0.75, 95% CI 0.71-0.80 vs. 0.79, 95% CI 0.75-0.83, P = 0.006), and any pre-eclampsia (0.78, 95% CI 0.74-0.81 vs. 0.82, 95% CI 0.79-0.84, P < 0.001). Following the retraining of the data with adjustments for the PlGF analyzers, the performance of Model 2 for predicting preterm pre-eclampsia, term pre-eclampsia, and any pre-eclampsia was improved with the AUC values increased to 0.84 (95% CI 0.80-0.89), 0.77 (95% CI 0.73-0.81), and 0.80 (95% CI 0.76-0.83), respectively. There were no differences in AUCs between Model 2 and the FMF competing risk model in the prediction of preterm pre-eclampsia (P = 0.135) and term pre-eclampsia (P = 0.084). However, Model 2 was inferior to the FMF competing risk model in predicting any pre-eclampsia (P = 0.024).
CONCLUSIONS: This study has demonstrated that following adjustment for the biochemical marker analyzers, the predictive performance of the AI+ML prediction model for pre-eclampsia in the first trimester was comparable to that of the FMF competing risk model in an Asian population.
摘要：
目的：评估人工智能（AI）和机器学习（ML）模型在大量亚洲人群中筛查先兆子痫的早期筛查的性能。
方法：这是对2016年12月至2018年6月亚洲7个地区妊娠11-13+6周接受常规妊娠护理的10935名单胎妊娠参与者的多中心前瞻性队列研究的二次分析。我们应用AI+ML模型预测早产先兆子痫（<37周），足月子痫前期（≥37周），和任何先兆子痫，这是在英国的一组怀孕参与者(模型1)中得出和测试的。该模型包括测量平均动脉压的母体因素，子宫动脉搏动指数,和血清胎盘生长因子(PlGF)。通过对用于生化测试的分析仪(模型2)的调整来进一步重新训练模型。通过受试者工作特征曲线下面积(AUC)评估辨别。Delong检验用于比较模型1、模型2和胎儿医学基金会(FMF)竞争风险模型的AUC。
结果：在预测早产先兆子痫方面，模型1的预测性能明显低于FMF竞争风险模型（0.82，95％置信区间[CI]0.77-0.87vs.0.86，95%CI0.811-0.91，P=0.019），足月子痫前期（0.75，95%CI0.71-0.80vs.0.79，95%CI0.75-0.83，P=0.006），和任何先兆子痫（0.78，95%CI0.74-0.81vs.0.82，95%CI0.79-0.84，P<0.001）。在重新训练数据并调整PlGF分析仪后，模型2预测早产先兆子痫的性能，足月子痫前期,和任何先兆子痫改善AUC值增加到0.84（95％CI0.80-0.89），0.77（95%CI0.73-0.81），和0.80（95%CI0.76-0.83），分别。模型2和FMF竞争风险模型在预测子痫前期(P=0.135)和足月子痫前期(P=0.084)方面的AUC无差异。然而,模型2在预测先兆子痫方面劣于FMF竞争风险模型(P=0.024)。
结论：这项研究表明，在对生化标志物分析仪进行调整后，AI+ML预测模型对妊娠早期子痫前期的预测性能与亚洲人群中FMF竞争风险模型的预测性能相当.
方法：这是对2016年12月至2018年6月亚洲7个地区妊娠11-13+6周接受常规妊娠护理的10935名单胎妊娠参与者的多中心前瞻性队列研究的二次分析。我们应用AI+ML模型预测早产先兆子痫（<37周），足月子痫前期（≥37周），和任何先兆子痫，这是在英国的一组怀孕参与者(模型1)中得出和测试的。该模型包括测量平均动脉压的母体因素，子宫动脉搏动指数,和血清胎盘生长因子(PlGF)。通过对用于生化测试的分析仪(模型2)的调整来进一步重新训练模型。通过受试者工作特征曲线下面积(AUC)评估辨别。Delong检验用于比较模型1、模型2和胎儿医学基金会(FMF)竞争风险模型的AUC。
结果：在预测早产先兆子痫方面，模型1的预测性能明显低于FMF竞争风险模型（0.82，95％置信区间[CI]0.77-0.87vs.0.86，95%CI0.811-0.91，P=0.019），足月子痫前期（0.75，95%CI0.71-0.80vs.0.79，95%CI0.75-0.83，P=0.006），和任何先兆子痫（0.78，95%CI0.74-0.81vs.0.82，95%CI0.79-0.84，P<0.001）。在重新训练数据并调整PlGF分析仪后，模型2预测早产先兆子痫的性能，足月子痫前期,和任何先兆子痫改善AUC值增加到0.84（95％CI0.80-0.89），0.77（95%CI0.73-0.81），和0.80（95%CI0.76-0.83），分别。模型2和FMF竞争风险模型在预测子痫前期(P=0.135)和足月子痫前期(P=0.084)方面的AUC无差异。然而,模型2在预测先兆子痫方面劣于FMF竞争风险模型(P=0.024)。
结论：这项研究表明，在对生化标志物分析仪进行调整后，AI+ML预测模型对妊娠早期子痫前期的预测性能与亚洲人群中FMF竞争风险模型的预测性能相当.
