Validation of competing-risks model in screening for pre-eclampsia in twin pregnancy by maternal factors

Objective To examine the predictive performance of the competing-risks model in screening for pre-eclampsia (PE) by maternal demographic characteristics and medical history in twin pregnancy, in a training dataset used for development of the model and a validation dataset. Methods The data for this study were derived from two prospective non-intervention multicenter screening studies for PE in twin pregnancies at 11 + 0 to 13 + 6 weeks' gestation. The first study of 2219 women, which was reported previously, was used to develop the competing-risks model for prediction of PE and is therefore considered to be the training set. The validation study comprised 2999 women. Patient-specific risks of delivery with PE at < 34 (early), < 37 (preterm) and < 41+ 3 (all) weeks' gestation were calculated using the competing-risks model and the performance of screening for PE in the training and validation datasets was assessed. We examined the predictive performance of the model by, first, its ability to discriminate between the PE and no-PE groups using the area under the receiver-operating characteristics curve (AUC) and, second, calibration, which assesses agreement between the predicted risk and observed incidence of PE. Results The incidence of early PE, preterm PE and all PE in the training and validation datasets was similar (1.8% vs 1.4%, 5.6% vs 5.6% and 7.7% vs 7.2%, respectively) and this was substantially higher than in our previous studies in singleton pregnancies. The training and validation datasets had similar AUCs for early PE (0.670 (95% CI, 0.593-0.747) vs 0.677 (95% CI, 0.594-0.760)), preterm PE (0.666 (95% CI, (0.617-0.715) vs 0.652 (95% CI, 0.609-0.694)) and all PE (0.656 (95% CI, 0.615- 0.697) vs 0.644 (95% CI, 0.606-0.682)). Calibration plots of the predictive performance of the competing-risks model demonstrated that, in both the training and validation datasets, the observed incidence of PE was lower than the predicted one and such overestimation of risk was particularly marked for early PE. Conclusions Discrimination and calibration of the competing- risks model for PE in a validation dataset are consistent with those in the training dataset. However, the model needs to be adjusted to correct the observed overestimation of risk for early PE. Copyright (c) 2019 ISUOG. Published by John Wiley & Sons Ltd.