Leveraging Big Data and Deep Learning to Develop Next Generation Decision Support Tools to Improve Glycemic Outcomes in Type 1 Diabetes

利用大数据和深度学习开发下一代决策支持工具以改善 1 型糖尿病的血糖结果

• 批准号: 10611369
• 负责人: Gavin Young
• 金额: $ 5.27万
• 依托单位: OREGON HEALTH & SCIENCE UNIVERSITY
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-05-01 至 2025-04-30
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10611369
• 关键词: Acute; Address; Adult; Aerobic; Aerobic Exercise; Affect; Algorithms; Area; Artificial Intelligence; Behavior; Big Data; Blood Glucose; Carbohydrates; Clinic; Clinical Medicine; Clinical Trials; Compensation; Complex; Consensus; Continuous Glucose Monitor; Control Groups; Cutaneous; Dangerousness; Data; Data Set; Decision Support Systems; Diabetes Mellitus; Disease; Dose; Eating; Event; Exercise; Fright; Functional disorder; Future; Glucose; Goals; Guidelines; Hour; Human; Hybrids; Hypoglycemia; Injections; Insulin; Insulin Infusion Systems; Insulin-Dependent Diabetes Mellitus; Intake; Jogging; Mathematics; Mediating; Modeling; Modification; Outcome; Participant; Patients; Performance; Persons; Physical Exercise; Physical activity; Physiological; Physiology; Production; Recommendation; Recording of previous events; Research; Resistance; Running; Safety; Structure of beta Cell of islet; Support System; Surveys; System; Techniques; Time; Tracer; Training; United States; Variant; Walking; Weight Lifting; Work; computer framework; deep learning; design; diabetic; empowerment; exercise regimen; experience; experimental study; glucose uptake; glycemic control; human model; human study; improved; in silico; mathematical model; medical complication; model building; next generation; novel; physiologic model; predictive modeling; predictive tools; primary outcome; recruit; resistance exercise; response; safety assessment; secondary outcome; simulation; smartphone application; strength training; support tools; tool; usability

Project Summary The hallmark of type 1 diabetes (T1D) is insufficient insulin production caused by pancreatic beta cell dysfunction. Most people treat their T1D through multiple daily injections (MDI) of insulin or use of a transcutaneous insulin pump. Several decision support smartphone apps exist to help people estimate insulin doses based on continuous glucose monitor (CGM) data and food intake. More sophisticated decision support tools employ mathematical models of human physiology to predict future glucose levels and provide generalized insulin therapy recommendations. Exercise is a crucial component of the long-term management of T1D, however many people avoid physical activity for fear of hypoglycemia (< 70 mg/dL). While consensus guidelines exist to help people manage glucose during physical activity, people still experience acute complications. Mathematical models of aerobic exercise yield promise in predicting hypoglycemia during controlled in- clinic experiments but do not perform well in the real-world or during other types of exercise. There is a critical need for a decision support system that helps people with T1D maintain safe glucose levels around exercise of varying types. The goal of this proposal is to develop a decision support tool to help people with T1D who utilize CGM better manage their glucose surrounding exercise. This tool will be called AIDES, the Artificially Intelligent Diabetic Exercise Support system. We hypothesize that use of a novel exercise- specific decision support tool, powered by predictive physiological modelling, artificial intelligence (AI), and deep learning, can provide treatment recommendations to reduce the number of hypoglycemic events experienced by people with T1D around regular physical exercise. In our first aim, we will develop a new model of resistance exercise that describes both insulin- and non-insulin mediated effects on glucose dynamics. We will then create a novel hybrid computational framework that harnesses AI to augment physiology models of aerobic and resistance exercise. This hybrid framework, called physAI, will harness real- world, free-living exercise data from the T1Dexi project (Big Data). In our second aim, we will leverage decades of research into deep learning with the Big Data provided by the T1Dexi project to train an AI-based decision support system that gives treatment recommendations to help users maintain target glucose during exercise. In our third aim, we will assess the safety and usability of our decision support engine in a small proof-of-concept study with human participants, supported by the Sponsor. This will be the first decision support system specifically designed to provide treatment recommendations that help users maintain safe glucose levels while performing aerobic and resistance exercise.

项目概要 1 型糖尿病 (T1D) 的特点是胰腺 β 细胞导致胰岛素分泌不足 功能障碍。大多数人通过每天多次注射 (MDI) 胰岛素或使用 经皮胰岛素泵。存在多种决策支持智能手机应用程序来帮助人们估计胰岛素 基于连续血糖监测仪 (CGM) 数据和食物摄入量的剂量。更复杂的决策支持 工具利用人体生理学的数学模型来预测未来的血糖水平并提供 普遍的胰岛素治疗建议。锻炼是长期管理的重要组成部分 T1D 的发病率很高，但许多人因担心低血糖 (< 70 mg/dL) 而避免体力活动。共识的同时 现有指南可以帮助人们在体力活动期间管理血糖，但人们仍然会经历急性 并发症。有氧运动的数学模型有望预测低血糖 受控的临床实验，但在现实世界或其他类型的锻炼中表现不佳。 迫切需要一个决策支持系统来帮助 T1D 患者维持安全的血糖水平 围绕不同类型的锻炼。该提案的目标是开发一种决策支持工具来帮助人们 患有 T1D 的人利用 CGM 可以更好地控制运动周围的血糖。这个工具将被称为AIDES， 人工智能糖尿病运动支持系统。我们假设使用一种新颖的练习- 特定的决策支持工具，由预测生理模型、人工智能 (AI) 和 深度学习，可以提供治疗建议，减少低血糖事件的发生次数 T1D 患者在定期体育锻炼时经历过的情况。在我们的第一个目标中，我们将开发一种新的 描述胰岛素和非胰岛素介导的葡萄糖影响的抗阻运动模型 动力学。然后，我们将创建一个新颖的混合计算框架，利用人工智能来增强 有氧运动和抗阻运动的生理学模型。这种称为 physAI 的混合框架将利用真实的 世界范围内，来自 T1Dexi 项目的自由生活锻炼数据（大数据）。在我们的第二个目标中，我们将利用 T1Dexi 项目提供的大数据对深度学习进行了数十年的研究，以训练基于人工智能的 决策支持系统，提供治疗建议，帮助用户在治疗期间维持目标血糖 锻炼。在我们的第三个目标中，我们将在一个小范围内评估我们的决策支持引擎的安全性和可用性。 在申办者的支持下，对人类参与者进行概念验证研究。这将是第一个决定 支持系统专门设计用于提供治疗建议，帮助用户维护安全 进行有氧运动和抗阻运动时的血糖水平。