Construction of a prediction model for adverse perinatal outcomes in fetal growth restriction based on a machine learning algorithm

Objective The objective of this study was to create and validate a machine learning (ML)-based model for predicting the adverse perinatal outcome (APO) in fetuses with a perinatal diagnosis of fetal growth restriction (FGR). Method This was a retrospective study of singleton gestations meeting the ISUOG-endorsed criteria for FGR from January 2021 and November 2023 at Beijing Obstetrics and Gynecology Hospital. The APO comprised one or more of: perinatal demise (stillbirth, immediate neonatal demise, or death before neonatal intensive care unit) discharge, cord arterial pH ≤ 7.10, and/or base excess ≥ 12, bronchopulmonary dysplasia, hypoxic ischemic encephalopathy, grade III-IV intraventricular hemorrhage, periventricular leukomalacia, seizures, necrotizing enterocolitis, and sepsis. Feature screening was performed using the random forest (RF), the Least Absolute Shrinkage and Selection Operator (LASSO) and logistic regression (LR). Subsequently, six ML methods, including Stacking, were used to construct models to predict the APO of FGR. Model’s performance was evaluated through indicators such as the area under the receiver operating characteristic curve (AUROC). The Shapley Additive exPlanation analysis was used to rank each model feature and explain the final model. Finally, we constructed a nomogram to make the predictive model results more readable. Results In total, a cohort of 411 non-anomalous singleton pregnancies with FGR were divided into a training set and a test set at a ratio of approximately 7:3. Among 16 candidate predictors (including maternal characteristics, maternal comorbidities, obstetric characteristics, and ultrasound parameters), the integration of RF, LASSO, and LR methodologies identified maternal pre-pregnancy body mass index, hypertensive disorders of pregnancy, gestational age at diagnosis of FGR, estimated fetal weight (EFW) z-score, EFW growth velocity, and abnormal umbilical artery Doppler as significant predictors. The Stacking model demonstrated a good performance in the test set (AUROC: 0.861,95% confidence interval (0.838–0.896)). The calibration curve and Hosmer-Lemeshow test demonstrated good calibration. Conclusions The ML algorithm was developed to possess the promising capacity of predicting APO in FGR at time of diagnosis. This approach may potentially improve early detection at high risk of APO in FGR.