Predicting complications of diabetes with longitudinal metabolic trajectories

摘要 糖尿病的患病率在全球范围内不断增加，2019年估计有4.63亿例（3100万 美国的案例）。糖尿病患者患有许多常见和病态的并发症， 包括周围神经病变（PN）和慢性肾病（CKD）。除了糖尿病，代谢 代谢综合征（MetS）已被确定为这些并发症的危险因素。不幸的是， 改善长期代谢障碍患者代谢危险因素的干预措施 完全预防或逆转糖尿病并发症，这可能是由于目前的时间安排， 代谢危险因素后的干预措施已经存在多年。重要的是，很少有研究 评估了MetS和糖尿病并发症之间的纵向关联;因此， 这些干预措施是未知的。随着年龄的增长，代谢状况恶化， 患者代谢轨迹之间的异质性量。目前尚不清楚的是， 纵向代谢特征的特定方面（如变化率、累积效应和变化 在特定的风险期）影响这些并发症的进展。也不知道发病是否 糖尿病并发症的发生率可以用纵向代谢轨迹准确预测。机 学习算法是用于利用复杂数据进行预测的灵活且强大的分析工具;因此， 他们提供了一种理想的方法来预测糖尿病并发症， 轨迹我们最初的目标是（1）确定代谢轨迹、PN和CKD之间的关联， 和（2）开发一种机器学习算法，以预测这些并发症的详细特征， 代谢轨迹我们将利用三个互补而强大的工具来实现这些初步目标。 具有不同代谢复杂性和样本量的数据库。首先，我们将确定代谢 轨迹与糖尿病并发症相关，并可预测糖尿病并发症， 美国印第安人进行了详细的纵向代谢和糖尿病并发症表型。 然后，为了确定我们的结果是否可以在大规模的综合美国卫生系统中实施， 我们的最终目标是（3）开发一种全面的机器学习算法来预测患者的并发症 使用退伍军人事务部企业数据仓库数据库，其中包含详细的 从1999年到2021年，超过310万糖尿病退伍军人的纵向医疗信息。结果从这个 这项研究将为代谢轨迹与糖尿病之间的关系提供一种新的理解。 并发症最终，这些数据将为未来的疾病修饰干预提供信息，这些干预可以逆转代谢紊乱， 最能预测糖尿病并发症的轨迹。

ABSTRACT The prevalence of diabetes is increasing worldwide, with an estimated 463 million cases in 2019 (31 million cases in the United States). Patients with diabetes suffer from a number of common and morbid complications, including peripheral neuropathy (PN) and chronic kidney disease (CKD). In addition to diabetes, the metabolic syndrome (MetS) has been firmly established as a risk factor for these complications. Unfortunately, interventions that improve metabolic risk factors for patients with long-term metabolic impairment do not completely prevent or reverse diabetic complications, which may be the result of the current timing of interventions after metabolic risk factors have been present for many years. Importantly, very few studies have assessed the longitudinal association between the MetS and diabetic complications; therefore, the ideal timing of these interventions is unknown. Metabolic profiles worsen as age increases and there is a substantial amount of heterogeneity between patient metabolic trajectories. What remains unknown is how changes in specific aspects of longitudinal metabolic profiles (such as rate of change, cumulative effects, and changes during specific risk-periods) affect the progression of these complications. It is also unknown whether the onset of diabetic complications can be accurately predicted using longitudinal metabolic trajectories. Machine learning algorithms are flexible and powerful analytic tools for making predictions with complex data; therefore, they offer an ideal approach to predict diabetic complications using complex characteristics from metabolic trajectories. Our initial aims are to (1) determine the association between metabolic trajectories, PN, and CKD, and (2) develop a machine learning algorithm to predict these complications with detailed characteristics of metabolic trajectories. We will accomplish these initial aims using three complementary and powerful databases with differential metabolic complexity and sample size. First, we will determine if metabolic trajectories are associated with and are predictive of diabetic complications using two cohorts of Pima American Indians that have undergone detailed longitudinal metabolic and diabetic complication phenotyping. Then, to determine if our results can be implemented in a large-scale, integrated United States health system, our final aim is to (3) develop a comprehensive machine learning algorithm to predict complications for patients with diabetes using the Veterans Affairs Corporate Data Warehouse database, which contains detailed longitudinal medical information on over 3.1 million Veterans with diabetes from 1999-2021. Results from this study will provide a novel understanding of the associations between metabolic trajectories and diabetic complications. Ultimately, this data will inform future disease modifying interventions that reverse the metabolic trajectories that are most predictive of diabetic complications.