Identification of Mild Cognitive Impairment using Machine Learning from Language and Behavior Markers

项目摘要 最近的估计表明，阿尔茨海默病（AD）可能列为第三大死亡原因 仅次于心脏病和癌症。尽管科学家们知道 AD涉及进行性脑细胞衰竭，细胞衰竭的原因尚不清楚。了解 疾病进展的关键之一是调查轻度脑梗死患者的认知变化， 认知障碍（MCI）。即使发现了成像和临床功能等生物标志物 在区分AD患者与认知正常（NC）患者方面表现突出，研究表明 在早期MCI中，它们的辨别力相当有限。检测信号， MCI受试者与NC受试者之间的差异具有挑战性，因为MCI受试者的敏感性低，变异性高， 目前的临床措施，如每年评估的神经心理测试结果和自我报告 功能测量。此外，即使体内生物标志物如β-淀粉样蛋白和tau蛋白可以 作为AD病理进展的指标，筛选生物标志物是 昂贵得令人望而却步，无法在门诊患者中广泛使用。 我们假设MCI的渐进性认知影响已经引起了可检测的方式的变化， 人们的谈话和行为，这可以通过廉价和方便的传感器来感知， 通过机器学习（ML）算法来构建量化MCI风险的预测模型。我们 一个小队列的初步结果表明，MCI和 在半结构化对话期间，NC主题，ML算法可以使用这种差异， 区分MCI和NC，具有良好的性能。我们在行为监测方面的初步结果 还提出了使用行为信号的时间模式的高度预测性能。在这个项目中， 我们计划在初步成功的基础上，对语言和行为进行全面的研究， 大规模队列中的标记物，以构建用于筛选的高性能和可解释的ML模型 MCI。我们的三个具体目标是：（1）发现语言标记，建立预测模型 表征MCI。使用I-CONECT项目的访谈录音，我们将使用自然的 语言处理和ML算法，以提取语言和声学标记，并开发多个 模态学习算法来融合这两种类型的信息。(2)发现行为标记， 开发表征MCI的预测模型。使用ORCATECH的家庭监测数据， 我们将提取短期和长期行为模式，并整合多粒度行为 用于区分MCI和NC的标记物。(3)将语言和行为标记与信息联系起来 框架.我们将使用人口统计学和常见的临床信息来描述患者并匹配 这两个群体通过某些相似性度量，创建互补的功能，以改善预测。

Project Summary Recent estimates indicate that Alzheimer’s disease (AD) may rank as the third leading cause of death for older people in the United States, just behind heart disease and cancer. While scientists know that AD involves a progressive brain cell failure, the reason why cells fail is still not clear. To understand the progression of the disease, one of the keys is to investigate the cognitive changes in patients with mild cognitive impairment (MCI). Even though biomarkers such as imaging and clinical functions are found to be outstanding in differentiating AD patients from those with normal cognition (NC), studies suggest that their discriminative power in early-stage MCI are rather limited. Detecting signals which distinguish subjects with MCI from those with NC is challenging due to the low sensitivity and high variability of current clinical measures such as annually assessed neuropsychological test results and self-reported functional measurements. Moreover, even though in-vivo biomarkers such as beta-amyloid and tau can be used as indicators of pathological progression towards AD, the screening of biomarkers are prohibitively expensive to be widely used among pre-symptomatic individuals in the outpatient setting. We hypothesize that progressive cognitive impact from MCI has elicited detectable changes in the way people talk and behave, which can be sensed by inexpensive and accessible sensors and leveraged by machine learning (ML) algorithms to build predictive models for quantifying the risk of MCI. Our preliminary results on a small cohort indicated that there are significant differences between MCI and NC subjects during a semi-structured conversation, and ML algorithms can use such differences for differentiating MCI and NC with promising performance. Our preliminary results in behavior monitoring also suggest highly predictive performance using temporal patterns of behavior signals. In this project, we plan to build upon our initial success and conduct comprehensive studies on language and behavior markers in larger-scale cohorts to build high-performance and interpretable ML models for screening MCI. Our three Specific Aims are: (1) Discover language markers and develop predictive models characterizing MCI. Using interview recordings from the I-CONECT project, we will use natural language processing and ML algorithms to extract linguistic and acoustic markers and develop multi- modal learning algorithms to fuse the two types of information. (2) Discover behavior markers and develop predictive models characterizing MCI. Using the in-home monitoring data from ORCATECH, we will extract short-term and long-term behavior patterns and integrate multi-granularity behavior markers to differentiate MCI and NC. (3) Linking language and behavior markers with an information framework. We will use demographics and common clinical information to profile the patients and match the two cohorts via certain similarity metrics, creating complementary features for improved prediction.