Using Machine Learning to Improve Readmission Prediction in Alzheimer's Disease and Related Dementia

项目摘要/摘要 到2060年，预计大约有1400万成年人患有阿尔茨海默病和相关的痴呆症 (ADRD)。虽然ADRD患者占一般老年人口的10%，但他们占37% 医疗保健的直接支出。与其他老年人相比，ADRD患者的发病率显著高于其他老年人 住院风险和非计划的30天再次住院(以下简称“再次入院”)。重新入院是 费用高昂，使ADRD患者面临认知功能减退、过早住院和死亡的风险。 出院后护理人员的可用性对于ADRD患者确保遵守饮食至关重要， 药物治疗和后续预约。在这些人中，重新入院的证据很少 ADRD人口。大多数风险评估工具(如蕾丝指数)辨别能力差，缺乏 包括有影响力的医疗和社会特征，以及ADRD患者特有的照顾者可用性。一个 潜在的解决方案是使用医院的电子健康记录(EHR)开发一种风险工具，因为它们 包含显著的临床和社会人口学特征以及来自医生的丰富信息， 护士和社会工作者笔记(非结构化EHR数据)。针对这一建议进行的具体研究 目的是(1)开发和验证用于预测ADRD患者重新入院的风险评估工具 患者；(2)检查再入院风险的可行性/可接受性和临床/经济效用- 评估工具；以及(3)开发自然语言处理(NLP)算法来提取信息 关于非结构化电子病历中护理人员的可用性(探索性)。我们假设预测能力 我们的风险工具的价值将至少比Lace Index(密歇根州目前使用的风险工具)高出20% 医学医院)。为了完成这个项目，我和我的导师们定义了一套有针对性的职业目标 和教育培训。我的培训目标包括(1)熟悉ADRD的临床方面 (与研究目标1相关)；(2)掌握机器学习和预测的方法技能 建模(与研究目标1相联系)；(3)发展对ADRD的物流的理解 病人出院和护理过渡过程(与研究目标2相关)；和(4)熟练掌握 在NLP和算法验证中(与研究目标3相关联)。到完成这个奖项时，我将使用 EHR和数据科学将为住院ADRD的再次住院开发有效的风险评估工具 病人。结果将使有效和有针对性的出院计划，以减少重新入院和浪费 花销。它还将提供申请R01所需的试验数据，以检查排放的优化 住院ADRD患者的流程/地点。这个职业发展奖将为我的职业发展奠定基础 成为一位专门研究ADRD患者有效护理过渡的独特健康经济学家。

Project Summary/Abstract By 2060, approximately 14 million adults are expected to live with Alzheimer’s disease and related dementia (ADRD). Although ADRD patients represent 10% of the general geriatric population, they account for 37% of the direct healthcare expenditures. Compared to other older adults, ADRD patients are at a significantly higher risk of hospitalization and unplanned 30-day hospital readmission (hereafter “readmission”). Readmissions are costly and expose ADRD patients to expedited cognitive decline, premature institutionalization, and death. Availability of a caregiver after hospital discharge is critical for ADRD patients to ensure adherence to diet, medications, and follow-up appointments. There is a paucity of evidence examining readmission among the ADRD population. Most risk-assessment tools (e.g. LACE Index) have poor discrimination power and lack inclusion of influential medical and social features, and caregiver availability particular to ADRD patients. A potential solution is to develop a risk tool using hospitals’ electronic health records (EHRs) because they contain salient clinical and sociodemographic features as well as a wealth of information from physicians’, nurses’ and social workers’ notes (unstructured EHRs data). The specific research aims for this proposal are to (1) develop and validate a risk-assessment tool for predicting readmission among ADRD patients; (2) examine the feasibility/acceptability and clinical/economic utility of the readmission risk- assessment tool; and (3) develop a natural language processing (NLP) algorithm to extract information on caregiver availability from unstructured EHRs (exploratory). We hypothesize that the predictive power of our risk tool will be at least 20% higher than that of LACE Index (the current risk tool used in the Michigan Medicine hospitals). To accomplish this project, my mentors and I have defined a set of targeted career goals and educational training. My training aims include (1) gain familiarity with the clinical aspects of ADRD (linked with Research Aim 1); (2) acquire methodological skills in machine learning and predictive modeling (linked with Research Aim 1); (3) develop an understanding of the logistics of the ADRD patient discharge and care transition processes (linked with Research Aim 2); and (4) gain proficiency in NLP and algorithm validation (linked with Research Aim 3). By completion of this award, I will have used EHRs and data science to develop a validated risk-assessment tool for readmission for hospitalized ADRD patients. The results will enable efficient and targeted discharge planning to reduce readmission and wasteful spending. It will also provide pilot data needed to apply for an R01 examining the optimization of discharge process/location for hospitalized ADRD patients. This career development award will lay the foundation for me to become a unique health economist specialized in efficient care transitions for ADRD patients.