Comparison of Prediction Models for Gestational Diabetes Mellitus: Logistic Model and Random Forest Model

This study aimed to develop and evaluate gestational diabetes mellitus (GDM) prediction models based on the logistic regression approach and Random Forest algorithm. GDM, defined as intolerance for glucose when first diagnosed during pregnancy, influences 5-12% of pregnancies worldwide and is a serious threat to the mothers' and infants’ health. Therefore, doctors and mothers must predict and intervene in GDM as early as possible. Data from 1012 women who were pregnant from laboratories in the Kurdistan Region was collected for this study, including age, number of deliveries, weight, height, body mass index (BMI), genetic information, and GDM diagnosis. Among them, 21.4% of pregnant women were diagnosed with GDM. By data preprocessing, division of training and test sets (80% training set, 20% test set), and using the SMOTE-Tomek Links sampling method to deal with data imbalance, it constructed logistic regression (LR) and random forest (RF) models, respectively. A grid search algorithm optimized the random forest model with parameters (n_estimators=300, max_depth=20, min_samples_split=10) and performed five-fold cross-validation. These two models were evaluated on the test set and the outcomes showed that the RF model had a good performance in terms of accuracy (0.875, 95% confidence interval (CI) [0.8261, 0.9239]), precision (0.874, 95% CI [0.7199, 1.0000]), specificity (0.979, 95% CI [0.9536, 1.0000]), and area under curve (AUC) value (0.740, 95% CI [ 0.6611, 0.8153]), which were superior to the LR model. According to the study, the random forest algorithm shows better performance in GDM prediction.