Enhancing early gestational diabetes mellitus prediction with imputation-based machine learning framework: A...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Reproductive Health

Background:

• Gestational diabetes mellitus (GDM) is a common pregnancy complication.
• Electronic health records (EHRs) offer potential for GDM risk prediction.
• Missing data in EHRs hinders the development of reliable prediction models.

Purpose of the Study:

• To address missing EHR data challenges in early GDM prediction.
• To evaluate the impact of imputation methods on GDM risk prediction models.
• To identify optimal imputation strategies for clinical machine learning pipelines.

Main Methods:

• Retrospective study of 5066 women with singleton pregnancies.
• Evaluation of 6 imputation methods combined with 4 machine learning models.
• Assessment of predictive performance, robustness, data distribution restoration, and feature selection using 10-fold cross-validation.

Main Results:

• Imputation significantly improved GDM prediction model performance.
• Logistic regression with MICE achieved the highest AUC (0.6899) compared to no imputation (0.6336).
• MICE demonstrated superior average performance, data distribution restoration, and robustness to missingness.

Conclusions:

• Imputation is critical for enhancing the performance and fairness of GDM prediction models.
• Findings offer practical guidance for incorporating imputation into clinical EHR data analysis.
• Identified 18 key features for early GDM prediction in the Chinese population.