FAST VARIABLE SELECTION FOR DISTRIBUTIONAL REGRESSION WITH APPLICATION TO CONTINUOUS GLUCOSE MONITORING DATA

Area of Science:

• Biostatistics
• Endocrinology
• Medical Informatics

Background:

• Diabetes prevalence and its public health impact necessitate identifying modifiable factors for glycemic control.
• Continuous glucose monitors (CGMs) offer high-frequency data, but traditional analysis methods lose significant information.
• Existing Fréchet regression methods for CGM data are computationally intensive and lack rigorous inference capabilities for large datasets.

Purpose of the Study:

• To develop a computationally efficient algorithm for sparse distributional regression using CGM data.
• To enable rigorous inference on large-scale CGM datasets by overcoming computational limitations.
• To examine associations between medication, comorbidities, and glycemic control in type 2 diabetes patients with obstructive sleep apnea.

Main Methods:

• Developed a novel algorithm for sparse distributional regression by deriving explicit gradient and Hessian characterizations.
• Utilized spherical rotations for feasible computational updates, significantly improving speed.
• Integrated the algorithm with stability selection for variable selection inference on CGM data.

Main Results:

• The new algorithm is over 10,000 times faster than the original approach, enabling large-scale analysis and inference.
• Identified a significant association between sulfonylurea medication and glucose variability, but not glucose mean.
• Found overnight oxygen desaturation variability to be more strongly associated with glucose regulation than overall oxygen desaturation levels.

Conclusions:

• The enhanced sparse distributional regression algorithm makes advanced CGM data analysis feasible for large populations.
• Findings highlight specific medication (sulfonylurea) and physiological factor (overnight oxygen desaturation variability) impacts on diabetes management.
• This work paves the way for more precise and personalized diabetes care through data-driven insights.