Enhancing Early Prediction of Gestational Diabetes Mellitus Through Data Augmentation and Feature Guidance: Model...

Area of Science:

• Computational biology and bioinformatics
• Medical informatics
• Machine learning in healthcare

Background:

• Early prediction of gestational diabetes mellitus (GDM) is crucial for maternal health.
• Existing predictive models face challenges with limited early-pregnancy data, class imbalance, and complex feature interactions.

Purpose of the Study:

• To develop and evaluate a unified dual-dimensional enhancement framework for improved early GDM prediction.
• To address data imbalance and leverage medical prior knowledge for better predictive performance.

Main Methods:

• A framework combining Generative Adversarial Network (GAN)-based data augmentation and large language model-inspired feature engineering was proposed.
• GANs generated synthetic minority class samples to mitigate data imbalance.
• LLMs organized features and generated higher-order composite features, integrating medical knowledge.

Main Results:

• The Tabular Variational Autoencoder (TVA)-enhanced random forest model achieved the best performance with recall 0.7559, accuracy 0.8444, and AUROC 0.8873.
• TVA significantly outperformed baseline and Conditional GAN methods in recall enhancement (Cohen d=2.894; P<.001).
• Key predictors included fasting blood glucose, a composite feature, activated partial thromboplastin time, leukocyte count, and neutrophil count.

Conclusions:

• The dual-dimensional enhancement framework effectively addresses data limitations and complex feature interactions for early GDM prediction.
• This strategy improves model performance, particularly recall, and offers interpretable biological insights.
• The findings support rapid clinical screening, stratified management, and early intervention in pregnancy.