Abstract
Elevated postprandial glucose levels present a global epidemic and a major challenge in type-2 diabetes (T2D) management. A key barrier to developing effective dietary interventions for T2D management is the wide inter-individual variation in glycemic and behavioral responses, which limits the impact of one-size-fits-all recommendations. To enable personalized dietary prompts for glycemic control, it is critical to first predict an individual's susceptibility to elevated postprandial (PPG) levels-or state of momentary vulnerability to PPG excursions. We examined the feasibility of personalized models to predict PPG excursions,and their associated vulnerability states, in the daily lives of 67 Chinese adults with T2D (Mage = 61.39; median = 63.00; 35 women; 2,463 glucose observations). We developed machine learning models trained on past individual observations to predict the next-in-time PPG excursion, using continuous glucose monitoring (CGM) data or CGM data combined with manually-logged meals and glucose-lowering agent intake. On average, personalized models predicted PPG excursions (F1-score: M = 75.88%; median = 78.26%), with substantial variation in predictability across individuals. Notably, no two individuals shared the same dietary and temporal predictors of PPG excursions. This study is the first to predict individual vulnerability states to glucose responses among adults with T2D in China. Findings can help personalize just-in-time adaptive interventions by tailoring dietary prompts based on individuals' unique vulnerability states to PPG excursions. This approach can inform the development of digital dietary interventions in mHealth apps and clinical decision support tools, thereby helping optimize glycemic control and patient-centered T2D lifestyle management..