Imputation of Missing Continuous Glucose Monitor Data.

Area of Science:

• Endocrinology
• Biomedical Engineering
• Data Science

Background:

• Continuous glucose monitors (CGMs) are crucial for assessing glycemic profiles in research and clinical settings.
• Missing data from CGM devices is an unavoidable challenge in data analysis.
• The Glycemic Observation and Metabolic Outcomes in Mothers and Offspring (GO MOMs) study provided data to evaluate imputation techniques.

Purpose of the Study:

• To investigate the impact of missing data on continuous glucose monitoring (CGM) summary metrics.
• To evaluate the performance of various imputation techniques for handling missing CGM data.
• To compare the accuracy of imputed data against true CGM data using mean relative bias (MRB).

Main Methods:

• Utilized 105 CGM profiles with complete nine-day glucose measurements.
• Introduced missing data strings at 2%, 5%, 10%, and 20% using a zero-inflated negative binomial hurdle model.
• Evaluated imputation methods including single, multiple, machine learning, hot-deck imputation, and complete case analysis.

Main Results:

• Mean relative bias (MRB) was minimal at 2% missing data across most imputation methods.
• MRB increased with higher frequencies of missing data, varying by metric and imputation technique.
• Hot-deck imputation and complete case analysis demonstrated consistently low MRB, indicating good performance.

Conclusions:

• Missing CGM data is an expected occurrence in real-world applications.
• Hot-deck imputation and complete case analysis are acceptable for up to 20% missing data if missingness is random.
• While imputation techniques are robust, their inherent limitations must be considered during implementation.