Efficient and explainable AI-driven early prediction of gestational diabetes mellitus using clinical and non-clinical data

Abstract

Gestational Diabetes Mellitus (GDM) occurs during pregnancy and can seriously affect both mother and baby. Influenced by lifestyle, genetics, and biological factors, early detection is crucial for reducing health risks. This study proposes a two-step AI-assisted approach to support doctors in efficient and cost-effective GDM screening. In the first step, non-clinical patient data is analyzed using a graphical user interface (GUI) model powered by XGBoost, achieving 88.4% accuracy, 12% false positive rate, and 87% specificity. Based on these predictions, doctors can selectively recommend further clinical testing, thereby saving time and reducing costs. In the second step, clinical test data, including glucose levels, was evaluated with XGBoost, achieving 97.63% accuracy, 98% precision and recall, 1.8% false positive rate, and 98% specificity, indicating highly reliable predictions. To enhance transparency and interpretability, Explainable AI (XAI) techniques were applied. SHAP (SHapley Additive exPlanations) identified critical features such as BMI, OGTT, HDL, Diastolic BP, and Prediabetes, while LIME provided local explanations for individual predictions. This approach not only improves prediction accuracy but also supports actionable, interpretable decision-making for medical professionals.