Machine-learning-based prediction of pre-eclampsia using first-trimester maternal characteristics and biomarkers

ObjectiveTo evaluate the accuracy of predicting the risk of developing pre-eclampsia (PE) according to first-trimester maternal demographic characteristics, medical history and biomarkers using artificial-intelligence and machine-learning methods. MethodsThe data were derived from prospective non-interventional screening for PE at 11-13 weeks' gestation at two maternity hospitals in the UK. The data were divided into three subsets. The first set, including 30 437 subjects, was used to develop the training process, the second set of 10 000 subjects was utilized to optimize the machine-learning hyperparameters and the third set of 20 352 subjects was coded and used for model validation. An artificial neural network was used to predict from the demographic characteristics and medical history the prior risk that was then combined with biomarker values to determine the risk of PE and preterm PE with delivery at < 37 weeks' gestation. An additional network was trained without including race as input. Biomarkers included uterine artery pulsatility index (UtA-PI), mean arterial blood pressure (MAP), placental growth factor (PlGF) and pregnancy-associated plasma protein-A. All markers were entered using raw values without conversion into standardized multiples of the median. The prediction accuracy was estimated using the area under the receiver-operating-characteristics curve (AUC). We further computed the detection rate at 10%, 20% and 40% false-positive rates (FPR). The impact of taking aspirin was also added. Shapley values were calculated to evaluate the contribution of each parameter to the prediction of risk. We used a non-parametric test to compare the expected AUC with the one obtained when we randomly scrambled the labels and kept the predictions. For the general prediction, we performed 10 000 permutations of the labels. When the AUC was higher than the one obtained in all 10 000 permutations, we reported a P-value of < 0.0001. For the race-specific analysis, we performed 1000 permutations. When the AUC was higher than the AUC in permutations, we reported a P-value of < 0.001. ResultsThe detection rate for preterm PE vs no PE, at a 10% FPR, was 53.3% when screening by maternal factors only, and the corresponding AUC was 0.816; these increased to 75.3% and 0.909, respectively, with the addition of biomarkers into the model. Information on race was important for the prediction accuracy; when race was not used to train the model, at a 10% FPR, the detection rate of preterm PE vs no PE decreased to 34.5-45.5% (for different races) when screening by maternal factors only and to 55.0-62.1% when biomarkers were added. The major predictors of PE were high MAP and UtA-PI, and low PlGF. The accuracy of prediction of all PE cases was lower than that for preterm PE. Aspirin use was recommended for cases who were at high risk of preterm PE. The AUC of all PE vs no PE was 0.770 when screening by maternal factors and 0.817 when the biomarkers were added; the respective detection rates, at a 10% FPR, were 41.3% and 52.9%. ConclusionsScreening for PE using a non-linear machine-learning-based approach does not require a population-based normalization, and its performance is similar to that of logistic regression. Removing race information from the model reduces its prediction accuracy, especially for the non-white populations when only maternal factors are considered. (c) 2022 International Society of Ultrasound in Obstetrics and Gynecology.