Machine Learning Prediction Models for Gestational Diabetes Mellitus: A meta-analysis (Preprint)

BACKGROUND Gestational diabetes mellitus (GDM) is a kind of common endocrine metabolic diseases, including carbohydrate intolerance of variable severity during pregnancy. The incidence rates of GDM related complications and adverse pregnancy outcomes will decline partly due to early screening. Nowadays, machine learning (ML) models have found an increasingly wide utilization, whether for risk factors selection or early prediction of GDM. OBJECTIVE Though many models for pregnancy women have been proposed and verified through experimental studies, few of them have been clinically recognized. Since seldom publication has evaluated the performance of ML prediction models for GDM, this meta-analysis was conducted and put forward some suggestions for model providers, users and policy makers basing on the findings. METHODS Four reliable electronic databases were searched for studies that developing ML prediction models for GDM in the general population, instead of the high-risk groups. The Prediction model Risk of Bias Assessment Tool (PROBAST) was used as a novel tool assessing the risk of bias of ML models. The software program Meta-Disc 1.4 was utilized to perform the Meta-analysis and determination of heterogeneity. To limit the influence of heterogeneity, results of sensitivity analysis, meta-regression and subgroups analysis were provided. RESULTS Twenty-five studies were analyzed which included women older than 18 years without a history of vital disease. The pooled area under receiver operating characteristic curve (AUC) and the pooled sensitivity and specificity for ML to predict GDM was 0.8492, 0.69 (95%CI: 0.68–0.69, P < .001, I2 = 99.6%)and 0.75 (95%CI:0.75–0.75, P < .001, I2 = 100%) respectively. As one of the most employed ML methods, logistic regression (LR) achieved an overall pooled AUC at 0.8151 while non-LR models performed better with an overall polled AUC at 0.8891. Additionally, maternal age, family history of diabetes, BMI and fasting blood glucose were the four mostly used features of models established by various feature selection methods. CONCLUSIONS ML methods could be cost-effective screening methods for GDM. The importance of quality assessment and unified diagnostic criteria should be further emphasized.