Within-person modeling of postprandial glucose using multimodal wearable data

The widespread adoption of continuous glucose monitoring (CGM) and wearable sensing technologies has enabled large-scale collection of high-resolution physiological and behavioral data in real-world settings. However, the analytical frameworks needed to translate these data into actionable, individualized insights remain limited. In particular, many existing approaches rely on population-level analysis or controlled experimental designs, which often fail to capture intra-individual variability in everyday life. To address this sensing-analysis gap, this study investigates within-person, meal-level predictors of postprandial glucose dynamics using multimodal data collected under free-living conditions. We analyzed outcomes including peak glucose, time to peak, and area under the glucose curve above 140 mg/dL. Predictors encompassed macronutrient composition, baseline glucose, meal timing, and short-term wrist movement variability derived from wearable sensors. Linear mixed-effects models were constructed with continuous predictors centered within individuals to explicitly capture within-person effects. Net carbohydrate intake showed the strongest association with postprandial glucose magnitude ( $p<0.001$ ), whereas total fat intake was associated with delayed glucose responses ( $p=0.024$ ), suggesting that glucose magnitude and response timing may reflect distinct physiological patterns. Baseline glucose was consistently associated with all outcomes ( $p<0.001$ ). In addition, greater short-term wrist movement variability was linked to reduced glucose excursions, suggesting the potential relevance of behavioral context in postprandial glucose variability. These preliminary findings demonstrate the potential utility of integrating multimodal wearable and behavioral data to characterize individualized postprandial glucose responses in real-world settings.