Combining information from multiple circadian activity rhythm metrics to optimally detect mild cognitive impairment using a consumer wearable

摘要：可广泛推广的早期检测阿尔茨海默病及相关疾病的方法 痴呆症(ADRD)将使早期干预成为可能，并有助于减少/推迟疾病的发生。消费者 被动收集大数据信号的可穿戴技术可以被用来检测 ADRD风险升高(见NOT-AG-20-017)，成本相对较低且可扩展。一套很有希望的套装 消费者可穿戴设备可以捕获的信号，但目前仅在研究中进行评估 设置，反映昼夜活动节律(CAR)。人类活动遵循一种可预测的24小时模式 也就是大家熟知的汽车。ADRD中各种汽车特征被破坏，反映ADRD生物标志物水平 (即使在临床前阶段)，并预测未来的认知能力下降。然而，观察性研究还没有 最终证明哪个CAR测量(S)是早期ADRD流程的最佳信号，并可以提供帮助 及早进行风险分层。以前的研究已经使用可用汽车指标的子集来建立 关联，而不是利用多个指标来改进ADRD风险预测。我们建议使用 汽车指标的综合小组可以确定对ADRD敏感的汽车指标组合 风险。此外，我们认为，将研究成果转化为临床筛查是困难的。 因为汽车测量依赖于研究人员，而不是临床/用户友好的系统。为了填补这些空白， 我们建议利用消费者可穿戴设备、现有数据、睡眠/生理科学和机器学习。我们的 首要目标是评估证据，从观察关联到临床的前进道路 使用消费者可穿戴设备进行有用的ADRD风险检测。我们的团队包括睡眠/汽车相关健康方面的专家 风险(斯马古拉博士，派)；衰老中的神经心理学和活动(古杰拉尔博士，co-I)；以及时间序列 分析/统计学习(Krafty博士，第一协办)。我们与主要群体的领导人建立了伙伴关系(见 支持)，以提供初始数据。Aim 1将在一个样本中计算一个全面的汽车指标面板 766名50岁以上的成年人；然后使用机器学习来开发算法，利用汽车测量来预测 轻度认知损害的可能性(MCI；ADRD风险升高的诊断标志物)。Aim 2将使用 新的测试样本(带MCI的n=25和不带MCI的n=25)，以验证是否将此算法应用于来自 消费者可穿戴设备可以准确检测MCI。斯马古拉博士已经开发了一种测量 使用Apple Watch的汽车被称为昼夜活动特征分析系统。这款R21可能会对 通过生产进行ADRD风险检测的领域：关于哪些汽车指标最好的信号MCI的证据；初始 一种结合有关汽车的信息以被动检测MCI可能性的算法；并通过改进 我们在一款流行的消费者可穿戴设备(Apple Watch)上收集这些信号的系统。我们还将 发展与其他队列的合作，以便如果我们发现支持潜在的临床效用的证据 在这种方法下，我们将准备使用来自多项研究的数据在R01上开发一个明确的算法。

Abstract: Widely-scalable methods for the earlier detection of elevated Alzheimer’s Disease and Related Dementia (ADRD) would enable earlier intervention and can help reduce/delay disease incidence. Consumer wearable technologies that passively gather “big data” signals could be leveraged to detect the early signs of elevated ADRD risk (see NOT-AG-20-017), in a relatively inexpensive and scalable fashion. One promising set of signals that can be captured by consumer wearable devices, but are currently only assessed in research settings, reflects the Circadian Activity Rhythm (CAR). Human activity follows a predictable 24-hour pattern known as the CAR. Various CAR characteristics are disrupted in ADRDs, reflect ADRD biomarkers levels (even in the pre-clinical stage), and predict future cognitive decline. However, observational studies have yet to conclusively demonstrate which CAR measure(s) best signal early-stage ADRD processes, and could help with early risk stratification. Previous studies have used subsets of the available CAR metrics to establish associations, rather than leveraging multiple metrics to improve ADRD risk prediction. We propose that using a comprehensive panel of CAR metrics could identify combinations of CAR metrics that are sensitive to ADRD risk. Furthermore, we propose that the translation of research findings into clinical screening has been difficult because CAR measurement relies on researcher-, rather than clinic-/user-, friendly systems. To fill these gaps, we propose leveraging consumer wearables, existing data, sleep/circadian science, and machine learning. Our overarching goal is to evaluate evidence for a path forward, from observing associations, towards clinically useful ADRD risk detection with consumer wearables. Our team includes experts in sleep/CAR-related health risks (Dr. Smagula, PI); neuropsychology and activity in aging (Dr. Gujral, co-I); and time series analytics/statistical learning (Dr. Krafty, co-I). We partnered with leaders of major cohorts (see letters of support) that provide the initial data. Aim 1 will compute a comprehensive panel of CAR measures in a sample of 766 adults aged 50+; then use machine learning to develop algorithms leveraging CAR measures to predict the likelihood of Mild Cognitive Impairment (MCI; a diagnostic marker of elevated ADRD risk). Aim 2 will use a new testing sample (n=25 with and n=25 without MCI) to validate if applying this algorithm to data from a consumer-wearable accurately detects MCI. Dr. Smagula already developed a working prototype measuring CARs using the Apple Watch called the Circadian Activity Profiling System. This R21 can have impact on the field of ADRD risk detection by producing: evidence regarding which CAR metrics best signal MCI; an initial algorithm that combines information regarding CARs to passively detect the likelihood of MCI; and by refining our system for collecting these signals on a popular consumer wearable (the Apple Watch). We will also develop collaborations with additional cohorts so that, if we find evidence supporting potential clinical utility of this approach, we will be prepared to develop a definitive algorithm in an R01 using data from multiple studies.