An early model to predict the risk of gestational diabetes mellitus in the absence of blood examination indexes: application in primary health care centres

Abstract

Background: Gestational diabetes mellitus (GDM) is one of the critical causes of adverse perinatal outcomes. A reliable estimate of GDM in early pregnancy would facilitate intervention plans for maternal and infant health care to prevent the risk of adverse perinatal outcomes. This study aims to build an early model to predict GDM in the first trimester for the primary health care centre.

Methods: Characteristics of pregnant women in the first trimester were collected from eastern China from 2017 to 2019. The univariate analysis was performed using SPSS 23.0 statistical software. Characteristics comparison was applied with Mann-Whitney U test for continuous variables and chi-square test for categorical variables. All analyses were two-sided with p < 0.05 indicating statistical significance. The train_test_split function in Python was used to split the data set into 70% for training and 30% for test. The Random Forest model and Logistic Regression model in Python were applied to model the training data set. The 10-fold cross-validation was used to assess the model's performance by the areas under the ROC Curve, diagnostic accuracy, sensitivity, and specificity.

Results: A total of 1,139 pregnant women (186 with GDM) were included in the final data analysis. Significant differences were observed in age (Z=-2.693, p=0.007), pre-pregnancy BMI (Z=-5.502, p<0.001), abdomen circumference in the first trimester (Z=-6.069, p<0.001), gravidity (Z=-3.210, p=0.001), PCOS (χ²=101.024, p<0.001), irregular menstruation (χ²=6.578, p=0.010), and family history of diabetes (χ²=15.266, p<0.001) between participants with GDM or without GDM. The Random Forest model achieved a higher AUC than the Logistic Regression model (0.777±0.034 vs 0.755±0.032), and had a better discrimination ability of GDM from Non-GDMs (Sensitivity: 0.651±0.087 vs 0.683±0.084, Specificity: 0.813±0.075 vs 0.736±0.087).

Conclusions: This research developed a simple model to predict the risk of GDM using machine learning algorithm based on pre-pregnancy BMI, abdomen circumference in the first trimester, age, PCOS, gravidity, irregular menstruation, and family history of diabetes. The model was easy in operation, and all predictors were easily obtained in the first trimester in primary health care centres.

Keywords: Gestational diabetes mellitus; Machine Learning; Maternal and infant health care; Prediction model; Primary health care centre.

Fig. 1

Receiver operating characteristic curve for estimating the discrimination of the Random Forest Model and the Logistic Regression model with all the variables. AUC, the area under the receiver operating characteristic curve

Fig. 2

Both the Random Forest Model and the Logistic Regression model had better performance in the ROC curve and AUC after dimensionality reduction. Abbreviations:  AUC, the area under the receiver operating characteristic curve

Fig. 3

Feature importances of variables in the Random Forest Model