A Machine Learning Framework to Quantify Postprandial Glucose Responses in Gestational Diabetes

Area of Science:

• Endocrinology
• Medical Informatics
• Machine Learning

Background:

• Gestational diabetes mellitus (GDM) requires careful glucose monitoring during pregnancy.
• Continuous glucose monitoring (CGM) provides valuable data, but analysis can be time-consuming.
• Automated methods are needed to efficiently interpret CGM data for GDM management.

Purpose of the Study:

• To develop and validate a machine learning (ML) framework for automatic identification of postprandial glucose responses (PPGR) from CGM data.
• To assess the accuracy of ML-identified PPGRs compared to those derived from self-reported mealtimes in pregnant adults with GDM.

Main Methods:

• A random forest ML algorithm was employed to analyze CGM data from pregnant adults with GDM or impaired glucose tolerance (IGT).
• Participants wore blinded CGMs and logged mealtimes.
• The ML algorithm's performance was evaluated by comparing its predicted mealtimes and subsequent PPGRs against those based on self-reported meal data.

Main Results:

• The ML algorithm demonstrated a median absolute error of 30 minutes in predicting mealtimes compared to self-reports.
• Differences in 1-hour and 2-hour postprandial glucose responses between ML-predicted and self-reported mealtimes were minimal (median differences of 8.7 mg/dL and 3.3 mg/dL, respectively).

Conclusions:

• A random forest ML algorithm effectively identifies PPGRs from CGM data in individuals with GDM.
• This automated approach provides a convenient and accurate method for monitoring postprandial dysglycemia in this population.