We selected patients with preeclampsia who gave birth in the Obstetrics Department of Zhengzhou Central Hospital Affiliated to Zhengzhou University from January 2021 to April 2024. The gestational age was determined based on the fetus size indicated by the last menstrual period and early pregnancy color ultrasound. If there was inconsistency between the two methods, the determination was made based on the results of the early pregnancy color ultrasound.

The diagnosis of preeclampsia is based on the American College of Obstetricians and Gynecologists criteria [2]: (1) systolic blood pressure ≥140 mm Hg, diastolic blood pressure ≥90 mm Hg after 20 weeks of gestation, or both, accompanied by urine protein quantification of 300 mg/24 hours, or random urine protein (+); (2) in the absence of proteinuria, preeclampsia can also be diagnosed if newly diagnosed hypertension occurs along with thrombocytopenia (platelet count <100×10⁹/L), impaired liver function (transaminase 2 times higher than the normal upper limit), new development of renal insufficiency (serum creatinine >97 umol/L or higher than 2 times the normal upper limit), pulmonary edema, or new onset cerebral or visual disturbances. FGR diagnostic criteria (Chinese population’s birth weight chart) [7] are as follows: the estimated fetal weight below the 10th percentile compared with the population norms on growth charts. Specifically, fetuses whose birth weight is less than 10% of the standard growth value for the corresponding gestational week are defined as FGR, and the others are non-FGR.

After excluding patients with multiple pregnancies, test-tube babies, fetal chromosomal anomalies, miscarriage in the second trimester, stillbirth, and incomplete medical records, we included a total of 513 cases for analysis among 570 patients with preeclampsia, consisting of 149 patients with FGR and 364 patients without FGR. The selection of research objects is shown in Figure 1. Following the same criteria as mentioned earlier, we also collected a dataset for external validation of the model. This dataset includes records from 103 pregnant women, among whom 45 delivered infants with FGR and 58 without FGR.

We initially identified 38 features potentially related to FGR through a literature review. Recorded patient characteristics included baseline characteristics, laboratory results in late pregnancy (before delivery), and the weight of neonates. Baseline characteristics included maternal age, gravidity, parity, prepregnancy BMI, blood pressures, gestational ages, personal history (preeclampsia, repeated abortions, and missed abortion history), family history (hypertensive disorders), with or without chronic hypertension, coexistence of gestational diabetes mellitus during pregnancy, systolic-to-diastolic ratio (S/D) value, amniotic fluid index (AFI), and gestational age at birth. Relevant laboratory tests included urine protein quantitation, routine blood tests, hepatic and renal function, lipid profiles, coagulation function, and serum haptoglobin. In past studies, these features of preeclampsia were strongly associated with FGR. Subsequently, we used feature importance to select more critical features for the final model construction.

The protocol was approved by the ethics committee of Zhengzhou Central Hospital Affiliated to Zhengzhou University (no. ZXYY202470). Considering its retrospective nature, the requirement for informed consent was waived. All relevant clinical data and outcome information were collected from the hospital’s electronic medical record system.

Before model training, we applied identical preprocessing steps to the training, validation, and test sets. First, we removed variables with missing rates exceeding 10%, then imputed the remaining missing values using a random forest (RF)–based method, missForest. For continuous variables, we conducted standardization through min-max scaling. We divided the collected maternal data into a training set and a test set, with 70% used for training and 30% for validation (internal validation), in order to avoid problems with overfitting. The SHAP method was adopted to rank the feature importance. For feature selection, we added features to the model training process step by step according to the order of feature importance shown by SHAP.

For training models, we applied 7 widely used ML algorithms in assisted diagnosis prediction, including logistic regression [18], tree-based model (light gradient boosting machine [LightGBM]) [19], RF [20], and extreme gradient boosting (XGBoost) [21], multilayer perceptron [22], naive Bayes [23], and support vector machine–based methods [24]. The logistic regression model represents the linear model, which is a state-of-the-art classification model for baseline construction. RF, LightGBM, and XGBoost were representative ensemble learning models, which do well in dealing with multiple types of features. RF is a bagging ensemble algorithm containing multiple decision trees, which uses the voting method to classify samples and integrate the final voting results produced by multiple decision trees. XGBoost is an optimized distributed gradient boosting library designed to be highly efficient, flexible, and portable. It implements ML algorithms under the gradient boosting framework. LightGBM is a gradient-boosting framework that uses tree-based learning algorithms. It uses a gradient-based 1-sided sampling algorithm to reduce the sample dimension and a mutually exclusive feature bundling algorithm to reduce the feature dimension.

A total of 38 features aforementioned were used to develop the prediction models. In order to optimize the prediction model, grid search combined with manual fine-tuning was applied to obtain the final hyperparameters (Multimedia Appendix 1). Commonly used evaluation indexes, such as the area under the receiver operating characteristic curve (AUC-ROC), sensitivity (recall), specificity, accuracy, precision, and F₁-score were used to evaluate the reliability of these models. During the training of our prediction models, we used the 5-fold cross-validation method for model training. The SHAP method was able to select clinically important features that contribute significantly to model predictions. It can improve the robustness of the model to a certain extent by removing irrelevant or noisy features and avoiding the feature collinearity problem. Finally, the prediction model was further evaluated in an external dataset.

To enhance the clinical applicability, we combined the diagnostic prediction results with clinical feature importance explanations to identify a patient with preeclampsia and FGR. In this study, except for interpreting feature contributions, SHAP values were used to assess the clinical significance of predictive models [25].

For the ML model interpretation, we used both globally explainable and locally explainable methods [26]. The global explanation determines the importance of features by comparing the magnitude of the model prediction error change before and after replacing a feature. If the prediction error changes more, it indicates that the feature is more important. Meanwhile, we used SHAP values to explore individual-based decision-making processes in the view of local explanation.

Web page deployment tool based on the Streamlit framework to facilitate the use of the model in clinical settings, the final prediction model was implemented into a web application established based on the Streamlit Python-based framework (Python Software Foundation). When the values of corresponding features from the final model are provided, the application can return the probability of FGR and the force plot for the individual.

We collected information on 513 parturients and newborns, of which 149 were diagnosed with FGR. Among the information collected before delivery, 29 variables were continuous and 9 were discrete. We statistically analyzed information on the group differences between these variables between newborns with FGR and those without FGR. Demographics, pregnancy characteristics, and neonatal outcomes for eligible patients with preeclampsia are described in Table 1. Maternal demographics in regard to age, height, and prepregnancy BMI have no difference between the 2 groups (P<.05) but differed in gestational age at diagnosis of preeclampsia, weight, weight gain during pregnancy, maximum systolic or diastolic pressure (mm Hg), umbilical artery (S/D), and AFI. Previous preeclampsia, onset period of hypertension, chronic hypertension, combined with gestational diabetes, history of FGR, and family history of hypertension were strong risk factors for FGR in patients with preeclampsia in our cohort. The following laboratory parameters with statistically significant differences were screened by intergroup difference analysis (P<.05): hemoglobin, total protein, albumin, uric acid, fibrinogen, total cholesterol, and serum haptoglobin (Table 1). Besides, there was also a difference in the sex of the newborns between the 2 groups.

The data collected were used to generate 7 ML models to predict FGR developed during prenatal testing in pregnant women. The discriminative performances (including AUC, accuracy, sensitivity [recall], specificity, precision, and F₁-score) of these models are shown in Table 2 and Figure 2. Among the 7 models, the RF model (AUC 0.812) had the best predictive effect for FGR, followed by the LightGBM (AUC 0.806) and GBM (AUC 0.803) models. For the RF model, the SHAP summary plots of the top 20 features (Figure 3) for the RF model are presented.

In order to study the influence of the number of features selected during model construction on model performance, we added features to the model training process step by step according to the order of feature importance shown by SHAP. The final model was identified during the feature reduction of the RF model. As displayed in Figure 4, the 38-feature model was significantly better than the 5-feature model (△AUC=0.036; P=.004), but not significantly better than the 10-feature model (△AUC=0.011; P=.16) in predicting FGR in prenatal testing in a pregnant woman.

The 10-feature model had a good net benefit and a high threshold probability, comparable to the 38-feature model. Hence, we focused on the 10-feature RF model. Considering the adverse effect of FGR predicted by fetal sex on sociodemographic sex ratio, we retained the other 9 features—urinary protein quantification (UPQ; mg/24 h), gestational week of delivery (GWD), umbilical artery S/D (free segment), AFI (mm), triglyceride (mmol/L), D-dimer (mg/L), weight (kg), height (cm), and maximum systolic pressure (MSP; mm Hg)—as the final model for further analysis. For the model building process referred to 513 patients (Figure 5A), after 5-fold cross-validation, the final RF model achieved an AUC of 0.830 with a sensitivity of 0.886, a specificity of 0.866, and an accuracy of 0.872 for predicting FGR the prenatal testing in a pregnant woman. The external validation dataset was collected and referred to 103 patients using the same standardized protocols and inclusion or exclusion criteria as the internal dataset. Statistical tests confirmed that no significant differences (all P>.05) existed in key variables between the 2 cohorts (Table S1 in Multimedia Appendix 1). For the external validation (Figure 5B), the final model gave an AUC of 0.820, which was similar to that in the internal validation (ΔAUC=0.05; P=.48), indicating that the final model showed great performance both in internal and external validations.

Since it is difficult for clinicians to accept a prediction model that is not directly explainable and interpretable, the SHAP method is used to interpret the output of the final model by calculating the contribution of each variable to the prediction. This method could provide 2 types of explanations: a global explanation of the model at the feature level and a local explanation at the individual level. The global explanation described the overall functionality of the model. As shown in SHAP summary plots (Figure 6), the contributions of the feature to the model were evaluated using the average SHAP values and exhibited in descending order. UPQ, umbilical artery S/D, and GWD showed the top three contributions (c=0.45, 0.34, and 0.33) to the final model.

Additionally, the SHAP dependence plot can facilitate understanding how a single feature affects the output of the prediction model. The real values versus the SHAP values of these 9 features are shown in Figure 7, and SHAP values that are higher than zero correspond to a positive class prediction in the model, in other words, a higher risk of FGR. For instance, pregnant women with UPQ>2500 ml or MSP>150 mm Hg had SHAP values higher than 0, which pushed the decision toward the “FGR” class. In addition, GWD>37 weeks or AFI<100 mm pushed the decision toward the “non-FGR” class, as well as triglyceride (mmol/L) ≤5 or a high actual value ≥3 of umbilical artery S/D (free segment).

Furthermore, local explanation analyzed how a certain prediction was made for a specific individual by incorporating the individualized input data. The actual measured values of features were also displayed in the waterfall plot (Figure 8). Figure 8A shows a patient whose child did not develop FGR during the pregnancy stay. As observed, the values of GWD (weeks), umbilical artery S/D (free segment), MSP (mm Hg), UPQ (mg/24 h), D-dimer (mg/L), and body weight pushed the decision toward the “non-FGR” class, but triglyceride did not. Figure 8B shows another patient whose child developed FGR during pregnancy. At present, multiple indicators such as abnormal UPQ (mg/24 h) and umbilical artery S/D (free segment) can increase the risk of FGR, pushing the decision toward the “FGR” class.

The final prediction model was implemented into the web application to facilitate its use in clinical scenarios. When the actual values of the 9 features required for the model are entered, this application will automatically predict the risk of FGR for an individual child. The web application is available in Multimedia Appendix 1.

Among the 7 ML models used to predict FGR from patients with preeclampsia, the RF model had the best performance with a good net benefit and a high threshold probability in feature reduction. Thus, we used the RF algorithm to develop the final model with 10 features. These features can be easily obtained or evaluated during pregnancy, so the model is promising as an early identification tool for the occurrence of FGR in pregnant women with preeclampsia.

In our study, training and testing data were collected from real clinical environments, ensuring that the model is exposed to a diverse and representative dataset. The random sampling of data helps to minimize bias and ensures that the model is robust to variations in input data. The model’s performance on real-world data validates its ability to generalize well and maintain high predictive accuracy under different input assumptions.

The ML technique has been described as a “black box” with little explanation about how predictions are derived. This may result in clinicians refusing to use it because they are hesitant to make medical decisions based on opaque information. This brought up another advantage of this study: we used the SHAP approach to explain the “black box” of ML models. The SHAP method could provide a global explanation that describes the overall functionality of a model and a local explanation that details how a certain prediction is made for an individual by inputting individualized data. Moreover, with a convenient tool based on the Streamlit framework, this prediction model can be used on a web page and shared with more clinicians.

Our final model performed well in both internal and external validations, with an AUC of 0.83 and 0.82, respectively. Except for fetal sex, there are 9 main characteristics for predicting FGR in our model, including UPQ, MSP, triglyceride, GWD, D-dimer, weight and height of pregnant woman, umbilical artery S/D, and AFI. Urinary protein quantity and blood pressure are important indicators to reflect the severity of preeclampsia. Our prediction model found a positive correlation between urinary protein quantity, MSP, and FGR, especially when the 24-hour urinary protein quantity was greater than 2500 mg and the MSP was greater than 160 mm Hg.

A Mendelian randomization study has revealed a strong correlation between triglyceride levels in pregnant women’s blood and the occurrence of preeclampsia [27]. Previous research has found that patients with preeclampsia have higher serum triglyceride levels than normal pregnant women [28,29], and those with FGR have significantly higher levels than those without FGR [30]. Our study further discovered that when triglyceride levels in patients with preeclampsia exceed 5 mmol/L, the occurrence of FGR increases. The mechanism may be due to the presence of dyslipidemia, which may result in shallow chorionic villus implantation, ultimately leading to reduced blood flow to the placenta and uterus. This, in turn, can hinder nutrient supply and fetal development, thereby increasing the likelihood of FGR occurrence in patients with preeclampsia [14,31,32]. Bozkurt et al [33] discovered that compared with normal pregnant women, patients with preeclampsia had higher D-dimer levels, and other studies have also found a correlation between D-dimer levels and the severity of preeclampsia [34]. These findings provide strong support for our study that the risk of FGR increases when D-dimer levels in patients with preeclampsia reach 2.5 mg/L. Among ultrasound indicators of fetal growth, AFI and S/D values can reflect the occurrence of FGR. The AFI was negatively correlated with the occurrence of FGR, especially when it was less than 90 mm, and the probability of FGR is significantly increased when the S/D value is greater than 3.

Class imbalance can indeed affect the performance of ML models in multiple aspects [35]. Sampling methods such as Synthetic Minority Oversampling Technique (SMOTE) and resampling (including undersampling and oversampling) can provide solutions to this problem [17]. In this study, the number of positive (preeclampsia complicated with FGR) and negative (preeclampsia without FGR) samples are 149 and 364 in the model-building process, as well as 45 and 58 in the external dataset. The ratios between the classes are 1:2.5 and 1:1.3, respectively, which are much lower than the commonly recognized class imbalance threshold of 1:10. Therefore, we believe that the class imbalance in this study has an extremely negligible impact on the construction of the model, and thus, no special treatment is carried out.

For future FGR prediction, integrating current models based on maternal physiological, ultrasound, and genetic data into real-time clinical decision support systems is vital. These systems, connected to hospital electronic health records, can provide instant risk assessments during prenatal checkups, guiding pregnancy management. Additionally, exploring new features like pregnant women's metabolomic profiles, maternal-fetal interface microbiomes, and longitudinal placental data can enhance model accuracy. Incorporating these into existing models will likely lead to more precise FGR prediction and better maternal-fetal outcomes.

We propose integrating the predictive model into the hospital information system. During routine prenatal checkups, health care professionals can collect data through the information system, after which the model will automatically generate predictions. Health care providers can then perform interventions or follow-ups based on these results. Furthermore, using model interpretation techniques to explain the predictions will provide crucial decision-making support for clinicians. For high-risk pregnancies, doctors can develop individualized management plans in advance, such as enhanced monitoring, nutritional guidance, and medication, to improve neonatal growth and prognosis.

The study has several limitations. First, clinical data of patients with preeclampsia were retrospectively collected from electronic records; therefore, it is necessary to collect some prospective cases and observe the follow-up results to verify the accuracy of the model. Second, due to the lack of guidelines or consensus for selecting features for the prediction model, how many features should be included in the model remains elusive. Although more features may provide more information for the prediction model, including a large number of features may limit the clinical use of the model, and including noncausal features may reduce the accuracy of the prediction. Nevertheless, further studies are needed to simplify and validate the application of the model in clinical conditions.

The prediction of FGR in patients with preeclampsia offers a valuable guide for improving perinatal outcomes. Combined with basic medical history and peripheral blood laboratory test results, this machine prediction model exhibits a reliable predictive value in predicting the occurrence of preeclampsia complicated by FGR. The model could be used as a decision-making tool to support clinicians.