Integrating Sleep and Circadian Factors into Machine Learning Models for Personalized Glycemic Response Prediction

Colorado State University, Fort Collins CO

Investigators

Linked publications & trials

Abstract

PROJECT SUMMARY/ABSTRACT There is increasing consumer interest in using continuous glucose monitors (CGMs) for personalized nutrition and health optimization. However, while CGMs can identify postprandial glucose spikes that contribute to type 2 diabetes (T2D) and cardiovascular disease (CVD), the ability to predict abnormal postprandial glucose responses (PPGRs) remains limited—undermining efforts to offer tailored dietary advice. A key challenge is that even when the same meal is consumed, PPGRs can vary dramatically within individuals due to the timing of behavioral factors like sleep, food intake, and physical activity. Despite this, existing machine learning (ML) models do not account for circadian biology, even though evidence from our group and others shows that eating at the “wrong” circadian time (e.g., at night) leads to significantly worse glycemic control. Moreover, current models are trained on datasets with minimal repeated-meal trials, limiting their ability to separate physiological signal from noise. To address this gap, we will leverage a funded R01 study in adult night shift workers undergoing 8 weeks of tightly controlled feeding and CGM monitoring—producing thousands of PPGR data points linked with detailed sleep, activity, and dietary timing data. Our overall objective is to improve the accuracy and generalizability of ML-based PPGR prediction models by accounting for real-world circadian and behavioral variability. In Aim 1, we will quantify within-subject PPGR variability and determine the relative impact of circadian factors using linear mixed-effects models and actigraphy-derived sleep metrics. In Aim 2, we will develop ML models that incorporate both traditional features and novel circadian predictors, benchmark them against existing models, and validate generalizability in five external NIH-funded datasets. Successful completion of this project will make progress towards scalable personalized glucose response prediction tools that can inform CGM-based dietary guidance and help reduce cardiometabolic disease risk in populations experiencing sleep and circadian disruption. This Re-Entry Supplement will also support the return of Dr. Corey Rynders to independent research following a family-related career interruption, equipping him with critical skills in data science and ML-based approaches to nutrition research.