Predictive Models Using Machine Learning to Identify Fetal Growth Restriction in Patients With Preeclampsia: Development and Evaluation Study.

BACKGROUND

Fetal growth restriction (FGR) is a common complication of preeclampsia. FGR in patients with preeclampsia increases the risk of neonatal-perinatal mortality and morbidity. However, previous prediction methods for FGR are class-biased or clinically unexplainable, which makes it difficult to apply to clinical practice, leading to a relative delay in intervention and a lack of effective treatments.

OBJECTIVE

The study aims to develop an auxiliary diagnostic model based on machine learning (ML) to predict the occurrence of FGR in patients with preeclampsia.

METHODS

This study used a retrospective case-control approach to analyze 38 features, including the basic medical history and peripheral blood laboratory test results of pregnant patients with preeclampsia, either complicated or not complicated by FGR. ML models were constructed to evaluate the predictive value of maternal parameter changes on preeclampsia combined with FGR. Multiple algorithms were tested, including logistic regression, light gradient boosting, random forest (RF), extreme gradient boosting, multilayer perceptron, naive Bayes, and support vector machine. The model performance was identified by the area under the curve (AUC) and other evaluation indexes. The Shapley additive explanations (SHAP) method was adopted to rank the feature importance and explain the final model for clinical application.

RESULTS

The RF model performed best in discriminative ability among the 7 ML models. After reducing features according to importance rank, an explainable final RF model was established with 9 features, including urinary protein quantification, gestational week of delivery, umbilical artery systolic-to-diastolic ratio, amniotic fluid index, triglyceride, D-dimer, weight, height, and maximum systolic pressure. The model could accurately predict FGR for 513 patients with preeclampsia (149 with FGR and 364 without FGR) in the training and testing dataset (AUC 0.83, SD 0.03) using 5-fold cross-validation, which was closely validated for 103 patients with preeclampsia (n=45 with FGR and n=58 without FGR) in an external dataset (AUC 0.82, SD 0.048). On the whole, urinary protein quantification, umbilical artery systolic-to-diastolic ratio, and gestational week of delivery exhibited the highest contributions to the model performance (c=0.45, 0.34, and 0.33) based on SHAP analysis. For specific individual patients, SHAP results reveal the protective and risk factors to develop FGR for interpreting the model's clinical significance. Finally, the model has been translated into a convenient web page tool to facilitate its use in clinical settings.

CONCLUSIONS

The study successfully developed a model that accurately predicts FGR development in patients with preeclampsia. The SHAP method captures highly relevant risk factors for model interpretation, alleviating concerns about the "black box" problem of ML techniques.