Hu Mengyao, Qin Tian, Gonzalez Richard, Freedman Vicki, Zahodne Laura, Melipillan Edmundo, Murphey Yi
The University of Texas Health Science Center at Houston.
University of Michigan-Dearborn.
Res Sq. 2024 Oct 15:rs.3.rs-4909790. doi: 10.21203/rs.3.rs-4909790/v1.
Alzheimer's disease and related dementias (ADRD) is a growing public health concern. The clock-drawing test (CDT), where subjects draw a clock, typically with hands showing 11:10, has been widely used for ADRD-screening. A limitation of including CDT in large-scale studies is that the CDT requires manual coding, which could result in biases if coders interpret and implement coding rules differently. This study created and evaluated an intelligent CDT Clock Scoring system built with Deep Learning Neural Networks (DLNN) to automatically code CDT images. We used a large, publicly available repository of CDT images from the 2011-2019 National Health and Aging Trends Study (NHATS) and compared three advanced DLNN methods - ResNet101, EfficientNet and Vision Transformers (ViT) in coding CDT into binary and ordinal (0 to 5) scores. We extended beyond the traditional nominal classification approach (which does not recognize order) by introducing structured ordering into the coding system and compared DLNN-coded CDT images with manual coding. Results suggest that ViT outperforms ResNet101 and EfficientNet, as well as manual coding. The ordinal coding system has the ability to allow researchers to minimize either under- or over-estimation errors. Starting in 2022, our developed ViT-coding system has been used in NHATS' annual CDT-coding.
阿尔茨海默病及相关痴呆症（ADRD）是一个日益引起公众健康关注的问题。画钟测试（CDT）要求受试者画一个时钟，通常指针显示为11:10，该测试已被广泛用于ADRD筛查。在大规模研究中纳入CDT的一个局限性在于，CDT需要人工编码，如果编码人员对编码规则的解释和执行方式不同，可能会导致偏差。本研究创建并评估了一个基于深度学习神经网络（DLNN）构建的智能CDT时钟评分系统，用于对CDT图像进行自动编码。我们使用了一个来自2011 - 2019年国家健康与老龄化趋势研究（NHATS）的大型公开可用的CDT图像库，并比较了三种先进的DLNN方法——ResNet101、EfficientNet和视觉Transformer（ViT），将CDT编码为二进制和序数（0到5）分数。我们通过在编码系统中引入结构化排序，超越了传统的名义分类方法（不识别顺序），并将DLNN编码的CDT图像与人工编码进行了比较。结果表明，ViT的表现优于ResNet101、EfficientNet以及人工编码。序数编码系统能够让研究人员将低估或高估误差降至最低。从2022年开始，我们开发的ViT编码系统已被用于NHATS的年度CDT编码。
