Machine Learning-Based Risk Prediction Models for Gestational Diabetes Mellitus: A Systematic Review and Meta-Analysis (Preprint)

BACKGROUND Gestational diabetes mellitus (GDM) is a common chronic disease during pregnancy, which not only increases the risk of pregnancy complications for pregnant women but also may lead to adverse pregnancy outcomes for the fetus. Currently, many studies have applied machine learning algorithms to the risk prediction of gestational diabetes. However, there are significant differences in the machine learning algorithms used in different studies, resulting in inconsistent predictive performance of various models and making it difficult to draw unified conclusions. OBJECTIVE This study systematically evaluates the application effect of machine learning algorithms in predicting the risk of gestational diabetes mellitus, analyzes the differences in predictive performance among different algorithms and the influencing factors, and explores the core risk factors of gestational diabetes mellitus. It aims to provide references for proposing a universal early risk prediction model for gestational diabetes mellitus and improving its generalization ability. METHODS By searching PubMed, Embase, Web of Science, SpringerLink, Cochrane Library, China National Knowledge Infrastructure (CNKI), Wanfang, VIP, and the core database of Chinese Biomedical Literature Database (CBM), relevant literature using machine learning methods to construct a risk prediction model for gestational diabetes was screened. The search period was from the establishment of the databases to May 2025. A systematic review was conducted according to the PRISMA guidelines. Two researchers independently completed the screening of literature, data extraction, and evaluation of model quality using the Prediction Model Bias Risk Assessment Tool (PROBAST), with a focus on the internal and external validation of the model, model efficacy and other key factors. Meta-analysis, sensitivity analysis and subgroup analysis were performed using Stata MP 18.0 software. RESULTS A total of 16 studies were included, among which 13 studies underwent internal validation, 3 studies underwent external validation, and 10 studies implemented calibration measures. 22 machine learning methods were employed to develop 141 machine learning models, all of which completed performance evaluation in the internal validation set, and 14 were evaluated in the external validation set. The overall combined total AUC for the ML models predicting GDM was 0.962 (95% CI 0.961 - 0.963, I² = 99.9%, p < 0.001), the combined AUC in the internal validation set was 0.965 (95% CI 0.964 - 0.966, I² = 99.9%, p < 0.001), and the combined AUC in the external validation set was 0.743 (95% CI 0.735 - 0.751, I² = 93.7%, p < 0.001); the sensitivity was 0.716 (95% CI 0.712 - 0.720, I² = 99.9%, p < 0.001), the specificity was 0.896 (95% CI 0.895 - 0.898, I² = 99.9%, p < 0.001), the PLR was 1.565 (95% CI 1.543 - 1.586, I² = 99.2%, p < 0.001), the NegLR was 0.321 (95% CI 0.308-0.317, I² = 99.9%,p < 0.001), and the DOR was 3.316 (95% CI 3.020 - 3.252, I² = 98.6%, p < 0.001). Subgroup analysis based on machine learning models showed that the extreme gradient boosting (XGBoost) model and the deep machine learning method recurrent neural network (RNN) performed better. The AUC of the machine learning method (0.980, 95% CI 0.979 - 0.981) was superior to that of the traditional regression method (LR) (0.751, 95% CI 0.745 - 0.756). In addition, fasting blood glucose, glycosylated hemoglobin, history of GDM, age, and pre-pregnancy BMI were the five most valuable core risk predictors for pregnant women developing GDM. The PROBASR evaluation results showed that more than half of the studies had high bias risk. CONCLUSIONS ML has excellent accuracy in predicting the early risks of gestational diabetes. The prediction model it constructs can effectively identify high-risk individuals, enabling the clinician to achieve the best intervention period. Moreover, this method does not require invasive procedures, and is more economically valuable compared to existing screening methods. It provides clinical guidance for early screening. PROSPERO Registration ID: CRD420251089143 CLINICALTRIAL CRD420251089143