Machine Learning and Longitudinal Analyses of Metformin Response Among Veterans

2 型糖尿病是一种慢性疾病，可能适合精准医疗方法 因为它影响到一大批不同的人口。事实上，退伍军人管理局和美国糖尿病协会 协会指南建议糖尿病管理个体化，但初始糖尿病治疗是 在目前的常规临床护理中很少进行个体化治疗。绝大多数糖尿病患者都经过初步治疗 服用二甲双胍，超过四分之一的患者对单独的二甲双胍没有反应，导致治疗延迟 实现早期血糖控制和潜在可避免的糖尿病并发症风险。有一个 缺乏经过验证的个体化初始糖尿病治疗策略。因此，精准医学方法， 试图将最佳的疾病管理策略与个体患者的特征相匹配， 非常适合解决证据差距，系统地指导个体化糖尿病护理。个性化或 精准医疗治疗在心血管疾病护理中很常见，其中临床风险预测工具 指导药物选择（例如心房颤动的抗凝治疗）和治疗强度（例如胆固醇目标） 他汀类药物治疗）。我们提出了一种基于预测的类似糖尿病治疗个体化方法 评估糖尿病并发症风险和二甲双胍血糖反应的工具。 该职业发展奖 (CDA) 的总体目标是开发和验证预测 为个体化糖尿病护理提供信息的工具。该提案的前两个目标旨在确定 治疗开始时预测糖尿病相关并发症的患者特征（目标 1）以及 对二甲双胍的血糖反应（目标 2）。在目标 3 中，我们将评估这些预测模型是否可以提供信息 改善长期糖尿病并发症的治疗方法。我们将利用大VA 数据存储库创建两个独立的 2 型糖尿病退伍军人队列，以完成以下目标： 该建议并构成未来其他观察研究的基础。这项研究的创新之处在于 利用退伍军人的真实临床数据生成证据来指导精准医疗 干预措施；使用机器学习方法和纵向方法来捕获信息 在常规临床护理中重复测量，以最大限度地利用电子健康记录数据；和 专注于基于糖尿病诊断时可用数据的预测工具，以指导初始治疗。 职业发展计划将研究目标与培训目标结合起来，以便为 申请人从事以精准医学应用为重点的研究生涯，以提高 老兵身体健康。该提案的培训目标集中在三个领域，它们的共同主题是： 最大化临床相关观察研究的纵向 VA 临床数据：1) 机器学习 应用于临床数据库的方法； 2）纵向方法，包括增长混合模型， 能够识别重复临床测量的模式； 3）使用大数据进行因果推断的方法 规模观测数据。在成功完成拟议的科学和培训目标后， 申请人将准备进行专注于使用新型糖尿病药物的精准医学研究， 将遗传数据纳入临床预测模型，并评估预测模型的使用效果 在常规糖尿病护理中。多元化的指导团队拥有糖尿病和心血管领域的内容专业知识 疾病，以及机器学习、纵向流行病学研究和因果关系方面的方法学专业知识 推论。直接指导将与课程作业和研讨会相结合，以填补申请人培训中的空白 并确保作为一名临床研究者在独立方面取得进展。最后，环境非常适合 申请人的职业发展，包括通过 VA HSR&D 丹佛-西雅图中心提供的资源 以退伍军人为中心和价值驱动的护理创新和科罗拉多大学医学院。

Type 2 diabetes mellitus is a chronic disease that may be amenable to precision medicine approaches because it affects a large, diverse segment of the population. In fact, the VA and American Diabetes Association guidelines recommend individualization of diabetes management, yet initial diabetes treatment is rarely individualized in current routine clinical care. The vast majority of diabetes patients are initially treated with metformin, and over a quarter of these patients fail to respond to metformin alone, leading to delays in achievement of early glycemic control and potentially avoidable risk of diabetes complications. There is a paucity of validated strategies to individualize initial diabetes treatment. Thus, precision medicine approaches, which attempt to match optimal disease management strategies to characteristics of an individual patient, are ideal to address the evidence gap to systematically guide individualized diabetes care. Individualized or precision medicine treatment is common in cardiovascular disease care, where clinical risk prediction tools guide drug selection (e.g., anticoagulation in atrial fibrillation) and treatment intensity (e.g., cholesterol goals on statin therapy). We propose a similar approach for diabetes treatment individualization based on prediction tools that estimate risk of diabetes complications and glycemic response to metformin. The overall goals of this career development award (CDA) are to develop and validate prediction tools that inform individualized diabetes care. The first two aims of the proposal are designed to determine patient characteristics at the onset of treatment that predict diabetes-related complications (Aim 1) and glycemic response to metformin (Aim 2). In Aim 3, we will evaluate whether these prediction models can inform treatment approaches to achieve improvements in long-term diabetes complications. We will leverage large VA data repositories to create two independent cohorts of Veterans with type 2 diabetes to complete the aims of this proposal and form the basis of additional future observational studies. This study is innovative in that it leverages real-world clinical data from Veterans to generate evidence to guide precision medicine interventions; uses machine learning approaches and longitudinal methods to capture information from repeated measurements in routine clinical care to make maximal use of electronic health record data; and focuses on prediction tools based on data available at the time of diabetes diagnosis to guide initial treatment. The career development plan aligns research aims with training aims in order to prepare the applicant to undertake a research career focused on applications of precision medicine to improve Veteran health. The training goals of the proposal are focused in three areas that share a common theme of maximizing longitudinal VA clinical data for clinically-relevant observational research: 1) machine learning approaches applied to a clinical database; 2) longitudinal methods, including growth mixture models, that enable identification of patterns in repeated clinical measures; 3) methods for causal inference using large- scale observational data. Upon successful completion of the proposed scientific and training aims, the applicant will be prepared to pursue precision medicine studies focused on the use of newer diabetes drugs, incorporation of genetic data into clinical prediction models, and evaluation of the effect of prediction model use in routine diabetes care. The diverse mentorship team has content expertise in diabetes and cardiovascular disease, and methodological expertise in machine learning, longitudinal epidemiological studies, and causal inference. Direct mentorship will be paired with coursework and seminars to fill gaps in the applicant’s training and ensure progress towards independence as a clinician-investigator. Finally, the environment is ideal for the applicant’s career development, including resources available through the VA HSR&D Denver-Seattle Center of Innovation for Veteran-Centric and Value-Driven Care and the University of Colorado School of Medicine.