Exploring the Use of Deep Learning Neural Networks to Improve Dementia Detection: Automating Coding of the Clock-Drawing Test

项目摘要 阿尔茨海默病和相关痴呆症（ADRD）是老年人残疾的主要原因， 成为一个严重的公共卫生问题。时钟绘制测试（CDT），测量多个方面的 认知功能包括理解、视觉空间能力、执行功能和记忆， 广泛用作临床研究、流行病学研究和小组中检测痴呆症的筛查工具 调查。CDT要求受试者画一个时钟，通常是11点后10点的指针，然后分配 二进制（例如正常与异常）或序数（例如0到5）评分。大规模的一个重要限制 CDT需要手动编码，如果编码人员解释和实现， 以不同的方式编码规则。 一些小规模的研究探索了使用机器学习方法来自动化CDT编码。 这类研究在序数编码方面取得了有限的成功， 特别是对于复杂的图像分类，并且不如深度学习神经网络（DLNN）有效， 这是机器学习的一个新的、有前途的领域。最近，机器学习方法已被应用于 数字CDT（dCDT），一种使用数字笔和平板电脑的CDT形式。尽管在小的方面取得了一些有希望的结果- 规模的数据，到目前为止，dCDT研究只试图编码二进制类别。 该研究将开发先进的DLNN模型，以创建和评估智能CDT时钟 评分系统- CloSco -将自动编码CDT图像。我们将使用一个大型的，公开的 来自2011-2019年国家健康和老龄化趋势研究（NHATS）的CDT图像库， 由国家老龄化研究所资助的对65岁及以上的医疗保险受益人的研究。我们特别 将：1）为顺序和连续评分开发一个自动化的CDT编码系统; 2）评估 CloSco系统的性能，并研究连续CDT评分对痴呆的价值 分类和纵向CDT模型;以及3）编制和传播NHATS公共使用文件， 使用CloSco沿着CloSco DLNN分配有序和连续CDT代码的文档 程序.如果成功的话，DLNN程序可以为其他广泛可用的自动编码提供一个模型。 用于评估各种认知功能的绘画测试。

Project Summary Alzheimer's disease and related dementias (ADRD), a leading cause of disability among older adults, has become a critical public health concern. The clock-drawing test (CDT), which measures multiple aspects of cognitive function including comprehension, visual spatial abilities, executive function and memory, has been widely used as a screening tool to detect dementia in clinical research, epidemiologic studies, and panel surveys. The CDT asks subjects to draw a clock, typically with hands showing ten after 11, and then assigns either a binary (e.g. normal vs. abnormal) or ordinal (e.g. 0 to 5) score. An important limitation in large-scale studies is that the CDT requires manual coding, which could result in biases if coders interpret and implement coding rules in different ways. Several small-scale studies have explored the use of machine learning methods to automate CDT coding. Such studies, which have had limited success with ordinal coding, have used methods that are not designed specifically for complex image classification and are less effective than deep learning neural networks (DLNN), a new and promising area of machine learning. More recently, machine learning methods have been applied to digital CDT (dCDT), a form of CDT that uses a digital pen and tablet. Despite some promising results on small- scale data, thus far dCDT studies have only attempted to code binary categories. The proposed study will develop advanced DLNN models to create and evaluate an intelligent CDT Clock Scoring system – CloSco – that will automatically code CDT images. We will use a large, publicly available repository of CDT images from the 2011-2019 National Health and Aging Trends Study (NHATS), a panel study of Medicare beneficiaries ages 65 and older funded by the National Institute on Aging. Specifically, we will: 1) Develop an automated CDT-coding system for both ordinal and continuous scores; 2) Evaluate the performance of the CloSco system and investigate the value of continuous CDT scoring for dementia classification and longitudinal CDT models; and 3) Prepare and disseminate NHATS public-use files and documentation with ordinal and continuous CDT codes assigned using CloSco along with the CloSco DLNN program. If successful, the DLNN programs may offer a model for automating coding of other widely available drawing tests used to evaluate a variety of cognitive functions.