Natural Language Processing and Automated Speech Recognition to Identify Older Adults with Cognitive Impairment

项目摘要 该提案的目的是开发两种策略，自然语言处理（NLP）和自动化 语音分析（ASA），能够自动识别患有认知障碍（CI）的患者，从轻度到轻度 认知障碍（MCI）与阿尔茨海默病相关性痴呆（ADRD）的关系。数量 在美国，MCI和ADRD的老年人正在增加，但临床医生的能力和 近几十年来，研究人员在大规模识别它们方面进展甚微， 评估不一致。使用现有数据的替代策略，如诊断分析 临床记录或保险索赔中的代码的敏感性非常低。与机器一起使用的NLP和ASA 学习是可以大大提高在临床环境中检测MCI和ADRD的能力的技术。NLP 自动将电子健康记录（EHR）中的文本转换为适合分析的结构化概念。 因此，临床医生的体征和症状的文件或测试和服务的顺序，反映或解决 认知限制可以有效地捕获，可能早在临床医生使用ADRD相关的 诊断代码ASA通过识别不同的认知特征来直接测量认知， 演讲因此，通过NLP和ASA提取特征可以提供一种独特的认知测量方法， 它对个人及其照顾者的影响。 MCI和ADRD的早期检测可以帮助研究人员确定合适的患者进行研究和帮助 临床医生和卫生系统以患者为目标进行预防性护理和护理协调。由于这些原因， 需要更有效的、高度可扩展的策略来识别患有MCI和ADRD的人。具体目标 （1）使用从EHR中提取的特征开发和验证ML算法， NLP识别CI患者，（2）使用从ASA中提取的特征开发和验证ML算法， 在常规初级保健访问期间患者-提供者会面的录音，以识别CI患者， (3)使用NLP和ASA提取的特征开发并验证ML算法，以创建集成CI 诊断算法我们将使用NLP和ASA提取的特征开发机器学习算法 根据纽约市800名初级保健患者的神经认知评估数据进行培训， 他们在芝加哥的200名患者中进行了独立的抽样调查。在二次分析中，我们将训练ML算法 以识别MCI及其亚型。这个项目将是NLP、ASA和ML最严格的发展 第一个在初级保健环境中测试ASA，第一个测试NLP和ASA 特征提取策略的组合。由临床医生、卫生服务部门、 研究人员，神经认知和数据科学家将应用机器学习来开发这些高度 可扩展的自动化技术，用于识别MCI和ADRD。 1

Project Summary The purpose of this proposal is to develop two strategies, natural language processing (NLP) and automated speech analysis (ASA), to enable automated identification of patients with cognitive impairment (CI), from mild cognitive impairment (MCI) to Alzheimer’s Disease Related Dementias (ADRD) in clinical settings. The number of older adults in the United States with MCI and ADRD is increasing and yet the ability of clinicians and researchers to identify them at scale has advanced little over recent decades and screening with clinical assessments is done inconsistently. Alternative strategies using available data, like analysis of diagnostic codes in the clinical record or insurance claims, have very low sensitivity. NLP and ASA used with machine learning are technologies that could greatly increase ability to detect MCI and ADRD in clinical contexts. NLP automatically converts text in the electronic health record (EHR) into structured concepts suitable for analysis. Thus, clinicians’ documentation of signs and symptoms or orders of tests and services that reflect or address cognitive limitations can be efficiently captured, possibly long before the clinician uses an ADRD-related diagnostic code. ASA directly measures cognition by recognizing different features of cognition captured in speech. Extracting features through both NLP and ASA could thus provide a unique measure of cognition and its impact on the individual and their caregivers. Early detection of MCI and ADRD can help researchers identify appropriate patients for research and help clinicians and health systems target patients for preventive care and care coordination. For these reasons, more efficient, highly scalable strategies are needed to identify people with MCI and ADRD. The Specific Aims of this proposal are to (1) Develop and validate a ML algorithm using features extracted from the EHR with NLP to identify patients with CI, (2) Develop and validate a ML algorithm using features extracted from ASA of audio recordings of patient-provider encounters during routine primary care visits to identify patients with CI, (3) Develop and validate a ML algorithm using both NLP and ASA extracted features to create an integrated CI diagnostic algorithm. We will develop machine learning algorithms using NLP and ASA extracted features trained against neurocognitive assessment data on 800 primary care patients in New York City and validate them in an independent sample of 200 patients in Chicago. In secondary analyses we will train ML algorithms to identify MCI and its subtypes. This project will be the most rigorous development of NLP, ASA, and ML algorithms for CI yet performed, the first to test ASA in primary care settings, and the first to test NLP and ASA feature extraction strategies in combination. The multi-disciplinary team of clinicians, health services researchers, and neurocognitive and data scientists will apply machine learning to develop these highly scalable, automated technologies for identification of MCI and ADRD. 1