Penn Artificial Intelligence and Technology Collaboratory for Healthy Aging

项目摘要（摘要） 成功的居家养老可以大大提高生活质量，改善健康状况，减轻负担 在医疗系统上。Parent P30的目标是建立一个名为PennAITech的国家合作实验室， 用于开发、评估和实施人工智能软件和新技术， 促进家庭健康老龄化。人工智能的最新进展导致了高度显着的大型 语言模型（LLM），如OpenAI的ChatGPT。这些模特展示了非凡的能力， 理解和产生的文本，类似人类语言的显着程度，导致一个伟大的 有可能重塑人工智能在照顾老龄化人口方面的辅助研究。在本附录中，我们提出了一个 试点项目，使用法学硕士开发和探索强大的人工智能辅助工具，以支持老年护理 人口本研究以阿兹海默症及相关性失智症病患之家庭照顾者为研究对象 （ADRD），并检查是否可以开发LLM来回答护理人员的问题。由于大多数 LLM是在来自各个领域的文本数据上训练的，它们对特定领域的能力可能没有得到优化。因此，在本发明中， 本补编的首要目标是在我们的领域内收集高质量的数据，并利用这些数据来微调 LLM使他们更强大，以回答特定领域的问题。此外，这项工作将突出 特别是在ADRD护理的背景下LLM的研究未来的方向。为了实现这一目标，我们有两个 目标。在目标1中，我们将创建一个针对家庭行为干预的会话数据存储库 老年痴呆症患者的护理人员，以改善他们的生活质量。我们将生成，清洁和预处理 从行为干预会议的访谈记录为痴呆症患者的家庭照顾者从一个 正在进行的定性数据库，其中包括护理人员和治疗师之间的会议（截至2010年，N = 3，000）。 6/1/23）。在目标2中，我们将构建一个大型语言模型（LLM），以提供一个AI辅助的，高效的和可扩展的语言模型。 支持痴呆症照顾者行为干预的方法。我们建议使用高质量的数据 从目标1中生成的会话数据库中微调现有的强大的LLM，并构建LLM 适合回答痴呆症护理人员的问题，以帮助减少他们的焦虑和抑郁，改善 他们的精神状态和生活质量。由此产生的法学硕士预计将提供答案，密切配合 人类专家，提供人工智能辅助的，有效的和可扩展的行为干预方法， 痴呆症患者的家庭照顾者。此外，这种方法可以扩展到开发其他相关的LLM 应用，例如解决与ADRD相关的临床问题。通过这样做，这可以提供有价值的AI- 为养老服务产业赋能，为全面提升公共健康水平做出贡献。

Project Summary (Abstract) Successful aging in the home can greatly improve quality of life, improve health outcomes, and reduce burden on the healthcare system. The goal of the Parent P30 is to establish a national collaboratory, named PennAITech, for the development, evaluation, and implementation of artificial intelligence software and new technologies to facilitate health aging in the home. Recent advances in AI have led to the development of highly notable Large Language Models (LLMs) such as OpenAI’s ChatGPT. These models have showcased exceptional abilities in comprehending and producing text that resembles human language to a remarkable extent, leading to a great potential to reshape the AI assistance research in caring for aging population. In this supplement, we propose a pilot project to develop and explore powerful AI assisted tools using LLMs to support caring for the aging population. We focus our study on family caregivers of patients with Alzheimer’s Disease and Related Dementias (ADRD) and examine whether LLMs can be developed to answer questions that caregivers have. Since most LLMs are trained on text data from various domains, their ability for specific domains may not be optimized. Thus, the overarching goal of this supplement is to collect high-quality data in our domain, and use that to finetune the LLMs to make them more powerful to answer domain specific questions. Furthermore, this work will highlight future directions for research of LLM specifically in the context of ADRD care. To achieve this goal, we have two aims. In Aim 1, we will create a conversational data repository specific to behavior intervention for family caregivers of persons with dementia to improve their quality of life. We will generate, clean and preprocess the interview transcripts from behavior intervention sessions for family caregivers of persons with dementia from an ongoing qualitative data repository, which includes sessions among caregivers and therapists (N= 3,000 as of 6/1/23). In Aim 2, we will build a large language model (LLM) to provide an AI assisted, efficient and scalable approach in supporting behavior intervention for dementia caregivers. We propose to use the high-quality data from the conversational database generated in Aim 1 to finetune the existing powerful LLMs, and build an LLM suitable for answering questions from dementia caregivers to help reduce their anxiety and depression, improve their mental status and quality of life. The resulting LLM is expected to provide answers that closely align with those of human experts, offering an AI-assisted, efficient, and scalable approach to behavioral interventions for family caregivers of dementia patients. Also, this approach can be extended to develop LLMs for other relevant applications, such as addressing clinical questions related to ADRD. By doing so, this could provide valuable AI- enabled services to the aging care industry, contributing to the overall improvement of public health.