SCH: INT: Personalized Models of Nutrition Intake from Continuous Glucose Monitors

在美国，与心血管和代谢疾病有关的过早死亡中，有一半以上是由不良饮食造成的，其中包括2型糖尿病(T2D)。目前，患有T2D的成年人数量持续上升，有3000多万美国人患有T2D。另有8000万人目前面临从糖尿病前期进展到T2D的风险。改善食物选择仍然是现代糖尿病护理的基石，可以降低进展为T2D的风险。然而，目前，在患有或有T2D风险的成年人中实现及时和适当的生活方式改变是具有挑战性的。记录膳食选择和跟踪营养成分的传统方法可能不准确(例如，估计膳食的蛋白质含量)和繁重(即，个人必须手动将信息输入食物日记)。有趣的是，餐后血糖水平不仅取决于碳水化合物的含量，还取决于脂肪、蛋白质和纤维的含量；例如，在碳水化合物中添加脂肪和蛋白质通常会导致实现的血糖水平较小的上升和较慢的下降，从而降低风险。这表明，葡萄糖对一顿饭的反应的形状可能具有指示膳食含量的潜力。利用这些信息的一个独特机会是使用连续血糖监测仪(CGM)。CGM是一种附着在皮肤上的小型传感器，每5-15分钟连续测量一次血糖，使其能够自动记录对饮食的血糖反应。为此，研究人员将进行动态研究，让人们(健康的、患有T2D的或有T2D风险的)在自由生活条件下食用各种传统膳食，同时佩戴CGM和智能手表来评估体力活动。利用这些设备的数据，研究人员将开发能够预测一顿饭内容的机器学习算法。这个项目将有助于临床医生提供新的信息来支持积极的行为改变，以减少糖尿病前期到T2D的风险或进展，并使患者更容易被动和准确地跟踪他们饮食的营养成分，潜在地导致更健康的饮食和改善健康。该项目将开发新的混合膳食中葡萄糖反应的逆向代谢模型(IMM)，可以根据CGM和活动数据估计膳食中的常量营养素组成(碳水化合物、蛋白质、脂肪和纤维)。为了解释食物代谢的巨大个体间变异性，研究人员将开发考虑每个人的表型(例如，人体测量变量、肠道微生物区系)以及他们最近的食物摄入量和体育锻炼历史的IMM。将开发两种类型的IMM，个性化和个性化IMM。个性化模型是基于每个人自己的数据(即，标有相应常量营养素信息的CGM记录)专门为每个人开发的，但这可能需要为每个人收集大量训练集，这在临床环境中是不切实际的。出于这个原因，研究人员还将开发个性化的IMM，利用来自具有相似代谢特征和表型的其他个人的数据。完成这项工作将需要开发新的深层网络递归结构来估计膳食的大量营养素成分，并需要注意力机制和转移学习技术来探索和解释食物新陈代谢中的个体间差异。调查人员还将提供多模式去识别数据，包括CGM记录、人体测量和表型变量、体力活动和饮食条目。这类语料库是第一个公开发布的语料库，将促进对食物新陈代谢和饮食监测的计算建模的进一步研究。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

In the United States, poor diet contributes to more than half of premature deaths related to cardiovascular and metabolic disease, including type 2 diabetes (T2D). At present, the number of adults developing T2D continues to rise, with over 30 million Americans living with T2D. Another 80 million are currently at-risk of progressing from pre-diabetes to T2D. Improving food choices remains a cornerstone of modern diabetes care and can decrease the risk of progression to T2D. However, at present, achieving timely and appropriate lifestyle change in adults with or at-risk of T2D is challenging. Conventional methods to record meal choice and track nutritional composition can be inaccurate (e.g., estimating protein content of a meal) and burdensome (i.e., individuals must manually enter information into a food diary). Interestingly, the blood glucose profile after a meal depends not only on the carbohydrate content but also on the amount of fat, protein, and fiber; as an example, adding fat and protein to carbohydrates generally leads to smaller increases and slower decreases in achieved glucose levels, lowering risk. This suggests that the shape of the glucose response to a meal may have the potential to indicate meal content. A unique opportunity to exploit this information is to use a continuous glucose monitor (CGM). A CGM is a small sensor that attaches to the skin and measures glucose continuously every 5-15 minutes, making it possible to automatically record the glucose responses to meals. To this aim, the investigators will conduct ambulatory studies in which people (healthy, with T2D, or at-risk of T2D) will consume a variety of conventional meals in free-living conditions while wearing a CGM and a smartwatch to assess physical activity. With data from these devices, the investigators will develop machine-learning algorithms that can predict the content of a meal. This project would be helpful to clinicians to provide new information to support positive behavior change to reduce the risk of or progression from pre-diabetes to T2D, and would make it easier for patients to passively and accurately track nutritional components of their diet, potentially leading to healthier diets and improved health.This project will develop new inverse metabolic models (IMMs) of the glucose response to mixed meals that can estimate the meal's macronutrient composition (carbohydrates, protein, fat, and fiber) from CGM and activity data. To account for large inter-individual variability in food metabolism, the investigators will develop IMMs that consider the phenotype of each person (e.g., anthropometric variables, gut microbiota) as well as their recent history of food intake and physical exercise. Two types of models will be developed, individualized and personalized IMMs. Individualized models are developed specifically for each person based on their own data (i.e., CGM recordings labeled with the corresponding macronutrient information), but this may require collecting a large training set per person that would be impractical in clinical settings. For this reason, the investigators will also develop personalized IMMs that leverage data from other individuals who have similar metabolic characteristics and phenotype. Accomplishing this work will require development of new deep-network recurrent architectures to estimate meals' macronutrient compositions, and attention mechanisms and transfer-learning techniques to explore and explain inter-individual variability in food metabolism. The investigators will also make available the multimodal de-identified data, including CGM recordings, anthropometric and phenotype variables, physical activity and diet entries. Such corpus, the first of its kind to be publicly released, will promote further research on computational modeling of food metabolism and diet monitoring.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.