AD-AutoGPT: An Autonomous GPT for Alzheimer’s Disease Info demi ology

AD-AutoGPT: 用于阿尔茨海默病信息学的自主GPT

Haixing $\mathbf{Dai}^{\mathbf{1}}$ , Yiwei $\mathbf{Li^{1}}$ , Zhengliang $\mathbf{Liu^{1}}$ , Lin Zhao1, Zihao $\mathbf{W}\mathbf{u}^{\mathbf{1}}$ , Suhang $\mathbf{Song^{2}}$ , Ye $\mathbf{Shen^{3}}$ , Dajiang $\mathbf{Zhu^{4}}$ , Xiang $\mathbf{Li^{5,6}}$ , Sheng $\mathbf{Li^{7}}$ , Xiaobai $\mathbf{Yao^{8}}$ , Lu $\mathbf{Shi^{9}}$ , Quanzheng $\mathbf{Li^{5,6}}$ , Zhuo Chen2, Donglan Zhang10, Gengchen $\mathbf{Mai^{8,1^{* }}}$ , Tianming Liu1* *

Haixing $\mathbf{Dai}^{\mathbf{1}}$，Yiwei $\mathbf{Li^{1}}$，Zhengliang $\mathbf{Liu^{1}}$，Lin Zhao1，Zihao $\mathbf{W}\mathbf{u}^{\mathbf{1}}$，Suhang $\mathbf{Song^{2}}$，Ye $\mathbf{Shen^{3}}$，Dajiang $\mathbf{Zhu^{4}}$，Xiang $\mathbf{Li^{5,6}}$，Sheng $\mathbf{Li^{7}}$，Xiaobai $\mathbf{Yao^{8}}$，Lu $\mathbf{Shi^{9}}$，Quanzheng $\mathbf{Li^{5,6}}$，Zhuo Chen2，Donglan Zhang10，Gengchen $\mathbf{Mai^{8,1^{* }}}$，Tianming Liu1* *

Received: date / Accepted: date

收稿日期：date / 录用日期：date

Haixing Dai E-mail: hd54134@uga.edu

Haixing Dai 电子邮箱：hd54134@uga.edu

Yiwei Li E-mail: yl80817@uga.edu

Yiwei Li 电子邮箱: yl80817@uga.edu

Zhengliang Liu E-mail: zl18864@uga.edu

Lin Zhao E-mail: lin.zhao@uga.edu

Lin Zhao 电子邮箱: lin.zhao@uga.edu

Zihao Wu E-mail: zw63397@uga.edu

吴子豪电子邮箱：zw63397@uga.edu

Suhang Song E-mail: suhang.song@uga.edu

Suhang Song 电子邮箱: suhang.song@uga.edu

Ye Shen E-mail: yeshen@uga.edu

叶申电子邮箱: yeshen@uga.edu

Dajiang Zhu E-mail: dajiang.zhu@uta.edu

Xiang Li E-mail: xli60@mgh.harvard.edu

项立 E-mail: xli60@mgh.harvard.edu

Sheng Li E-mail: li sheng 1989@gmail.com

Sheng Li 邮箱：li sheng 1989@gmail.com

Xiaobai Yao E-mail: xyao@uga.edu

Xiaobai Yao 电子邮箱: xyao@uga.edu

Lu Shi E-mail: LUS@clemson.edu

Quanzheng Li

Abstract In this pioneering study, inspired by AutoGPT, the state-of-the-art open-source application based on the GPT-4 large language model, we develop a novel tool called AD-AutoGPT which can conduct data collection, processing, and analysis about complex health narratives of Alzheimer’s Disease in an autonomous manner via users’ textual prompts. We collated comprehensive data from a variety of news sources, including the Alzheimer’s Association, BBC, Mayo Clinic, and the National Institute on Aging since June 2022, leading to the autonomous execution of robust trend analyses, intertopic distance maps visualization, and identification of salient terms pertinent to Alzheimer’s Disease. This approach has yielded not only a quantifiable metric of relevant discourse but also valuable insights into public focus on Alzheimer’s Disease. This application of AD-AutoGPT in public health signifies the transformative potential of AI in facilitating a data-rich understanding of complex health narratives like Alzheimer’s Disease in an autonomous manner, setting the groundwork for future AI-driven investigations in global health landscapes.

摘要在这项开创性研究中，我们受基于GPT-4大语言模型的最先进开源应用AutoGPT启发，开发了名为AD-AutoGPT的新型工具。该工具能通过用户文本提示，自主完成阿尔茨海默病复杂健康叙事的采集、处理与分析工作。我们整合了自2022年6月以来阿尔茨海默病协会、BBC、梅奥诊所及美国国家老龄化研究所等多渠道的全面数据，实现了趋势分析、主题间距可视化及疾病相关核心术语识别的自主化执行。该方法不仅产出了可量化的相关论述指标，更揭示了公众对阿尔茨海默病的关注焦点。AD-AutoGPT在公共卫生领域的应用，彰显了AI以自主方式促进对阿尔茨海默病等复杂健康叙事进行数据化理解的变革潜力，为未来全球卫生领域AI驱动研究奠定了基础。

Keywords AutoGPT · GPT-4 · Alzheimer’s Disease · Info demi ology

关键词 AutoGPT · GPT-4 · 阿尔茨海默病 · 信息流行病学

Mathematics Subject Classification (2020) MSC code1 · MSC code2 more

数学学科分类（2020）MSC code1 · MSC code2 更多

� Tianming Liu E-mail: tliu@uga.edu

Tianming Liu E-mail: tliu@uga.edu

1 Introduction

1 引言

Alzheimer’s Disease (AD), a progressive neuro degenerative disorder, remains one of the most pressing public health concerns globally in the 21st century [1,2]. This disease, characterized by cognitive impairments such as memory loss, predominantly affects aging populations, exerting an escalating burden on global healthcare systems as societies continue to age [3]. The significance of AD is further magnified by the increasing life expectancy globally, with the disease now recognized as a leading cause of disability and dependency among older people [4]. Consequently, AD has substantial social, economic, and health system implications, making its understanding and awareness of paramount importance [5,6].

阿尔茨海默病 (AD) 是一种进行性神经退行性疾病，仍是21世纪全球最紧迫的公共卫生问题之一 [1,2]。该疾病以记忆力减退等认知障碍为特征，主要影响老年人群，随着社会持续老龄化，其对全球医疗系统的负担日益加重 [3]。全球预期寿命的延长进一步凸显了AD的重要性，该病现已被认定为老年人致残和丧失生活自理能力的主要原因 [4]。因此，AD对社会、经济和卫生系统具有重大影响，这使得对其认知与了解变得至关重要 [5,6]。

Despite the ubiquity and severity of AD, a gap persists in comprehensive, data-driven public understanding of this complex health narrative. Traditionally, public health professionals have to rely on labor-intensive methods such as web scraping, API data collection, data post processing, and analysis/synthesis to gather insights from news media, health reports, and other textual sources [7,8,9]. However, these methods often necessitate complex pipelines for data gathering, processing, and analysis. Moreover, the sheer scale of global data presents an ever-increasing challenge, one that demands a novel, innovative approach to streamline these processes and extract valuable, actionable insights efficiently and automatically. In addition, the technical expertise required for developing data processing and analysis pipelines significantly limits the access and engagement of the broader public health community.

尽管阿尔茨海默病(AD)普遍存在且危害严重，但公众对这种复杂健康议题的全面数据驱动认知仍存在空白。传统上，公共卫生专业人员必须依赖网络爬取、API数据采集、数据后处理及分析综合等劳动密集型方法[7,8,9]，从新闻媒体、健康报告等文本来源获取洞见。然而这些方法通常需要构建复杂的数据采集-处理-分析流水线，且全球数据规模带来的挑战与日俱增，亟需创新方法来实现流程自动化、高效提取可操作的见解。此外，开发数据处理分析流水线所需的技术专长，极大限制了更广泛公共卫生群体的参与度。

AutoGPT [10] is an experimental open-source application that harnesses the capabilities of large language models (LLMs) such as GPT-4 [11] and ChatGPT [12] to automate and optimize the analytical process. With its advanced linguistic understanding and autonomous operation, AutoGPT simplifies complex data pipelines, facilitating comprehensive analyses of vast datasets with simple textual prompts. This tool transcends traditional limitations, unlocking the potential of LLMs for autonomous data collection, processing, summarization, analysis, and synthesis. In this study, we modify the AutoGPT architecture into public health applications and develop AD-AutoGPT to analyze a multitude of news sources, including the Alzheimer’s Association, BBC, Mayo Clinic, and the National Institute on Aging, focusing on discourse since June 2022. We are among the pioneers in integrating AutoGPT into public health informatics, adapting this transformative AI tool into the public health domain to elucidate the complex narrative surrounding Alzheimer’s Disease. This research underlines the enormous potential of autonomous LLMs in global health research, paving the way for future AI-assisted investigations into various health-related domains.

AutoGPT [10] 是一款实验性开源应用程序，它利用 GPT-4 [11] 和 ChatGPT [12] 等大语言模型 (LLM) 的能力来自动化和优化分析流程。凭借其先进的语言理解能力和自主操作特性，AutoGPT 简化了复杂的数据管道，仅需简单的文本提示即可实现对海量数据集的全面分析。该工具突破了传统限制，释放了大语言模型在自主数据收集、处理、汇总、分析和综合方面的潜力。在本研究中，我们将 AutoGPT 架构改造为公共卫生应用，并开发了 AD-AutoGPT 来分析包括阿尔茨海默病协会、BBC、梅奥诊所和美国国家老龄化研究所在内的多种新闻来源，重点关注自 2022 年 6 月以来的相关论述。我们是首批将 AutoGPT 整合到公共卫生信息学中的研究者之一，将这一变革性 AI 工具应用于公共卫生领域，以阐明围绕阿尔茨海默病的复杂叙事。这项研究凸显了自主大语言模型在全球健康研究中的巨大潜力，为未来 AI 辅助各类健康相关领域研究铺平了道路。

We summarize our key contributions below:

我们在下方总结了关键贡献：

– Inspired by AutoGPT, we develop a novel LLM-based tool called ADAutoGPT, which can generate data collection, processing, and analysis pipeline in an autonomous manner based on users’ textual prompts. More specifically, we adapt AD-AutoGPT to the public health domain to showcase its great potential of autonomous pipeline generation to understand the complex narrative surrounding Alzheimer’s Disease.

受AutoGPT启发，我们开发了一款基于大语言模型(LLM)的创新工具ADAutoGPT，能够根据用户文本提示(prompt)自主生成数据收集、处理及分析流程。具体而言，我们将ADAutoGPT适配至公共卫生领域，通过构建阿尔茨海默症(Alzheimer's Disease)相关复杂叙事的理解流程，展示其自主生成分析管道的强大潜力。

– While AutoGPT is an effective autonomous LLM-based tool, it has lots of limitations when applying it on AD Info demi ology during the process of public health information retrieval, text-based information extraction, text sum mari z ation, summary analysis, and visualization. To overcome AutoGPT’s limitations for the AD Info demi ology task, AD-AutoGPT provides the following improvements: 1) specific prompting mechanisms to improve the efficiency and accuracy of AD information retrieval; 2) a tailored spatio temporal information extraction functionality; 3) an improved text sum- marization ability; 4) an in-depth analysis ability on generated text summaries; and 5) an effective and dynamic visualization capability.

虽然AutoGPT是一种有效的基于大语言模型的自主工具，但在应用于AD信息流行病学领域的公共卫生信息检索、文本信息提取、文本摘要、摘要分析和可视化过程中仍存在诸多局限。为克服AutoGPT在AD信息流行病学任务中的不足，AD-AutoGPT提供了以下改进：1) 采用特定提示机制提升AD信息检索效率与准确性；2) 定制化时空信息提取功能；3) 增强的文本摘要能力；4) 对生成文本摘要的深度分析能力；5) 高效动态的可视化功能。

– We show that AD-AutoGPT transforms the traditional labor-intensive data collection, processing, and analysis paradigm into a prompt-based automated, and optimized analytical framework. This has allowed for efficient, comprehensive analysis of numerous news sources related to Alzheimer’s Disease.

我们展示了AD-AutoGPT如何将传统劳动密集型的数据收集、处理和分析范式转变为基于提示词(prompt)的自动化优化分析框架。这使得对大量阿尔茨海默病相关新闻源的高效全面分析成为可能。

– Through AD-AutoGPT, we have provided a case study for detailed trend analysis, intertopic distance mapping, and identified salient terms related to Alzheimer’s Disease from four AD-related new sources. This contributes significantly to the existing body of knowledge and facilitates a nuanced understanding of the disease’s discourse in public health.

通过AD-AutoGPT，我们提供了一个案例研究，用于详细分析阿尔茨海默病(Alzheimer’s Disease)的趋势、绘制主题间距离图，并从四个AD相关新闻来源中识别出关键术语。这显著丰富了现有知识体系，并促进了对公共卫生领域疾病讨论的细致理解。

– Our research underlines the capacity of AD-AutoGPT to facilitate datadriven public understanding of complex health narratives, such as Alzheimer’s Disease, which is of paramount importance in an aging global society.

我们的研究强调了AD-AutoGPT在促进数据驱动的公众理解复杂健康叙事（如阿尔茨海默病）方面的能力，这在全球老龄化社会中至关重要。

– The methodologies and insights from our work provide a foundation for future AI-assisted public health research. Our AD-AutoGPT pipeline is extendable to other topics in public health or even other domains. This work paves the way for comprehensive and efficient investigations into various domains.

我们的工作方法和见解为未来AI辅助的公共卫生研究奠定了基础。AD-AutoGPT流程可扩展至公共卫生其他主题甚至其他领域，为全面高效探索多领域研究铺平了道路。

2 相关工作

2.1 Large Language Models

2.1 大语言模型 (Large Language Models)

Large language models (LLMs), with their origins in Transformer-based pretrained language models (PLMs) such as BERT [13] and GPT [14], have substantially transformed the field of natural language processing (NLP). LLMs have superseded previous methods such as Recurrent Neural Network (RNN) based models, leading to their widespread adoption across various NLP tasks [12,15]. Furthermore, the emergence of very large language models such as GPT-3 [16], Bloom [17], GPT-4 [11], PaLM [18], and PaLM-2 [19] demonstrates a clear trend towards even more sophisticated language understanding capabilities.

大语言模型 (LLMs) 起源于基于Transformer的预训练语言模型 (PLMs) ，例如BERT [13]和GPT [14] ，它们彻底改变了自然语言处理 (NLP) 领域。大语言模型取代了基于循环神经网络 (RNN) 等传统方法，在各种NLP任务中得到广泛应用 [12,15] 。此外，GPT-3 [16] 、Bloom [17] 、GPT-4 [11] 、PaLM [18] 和PaLM-2 [19] 等超大规模语言模型的出现，展现了语言理解能力向更复杂方向发展的明显趋势。

These models are designed to learn accurate contextual latent feature represent at ions from input text [20], which can then be employed in a variety of applications, including question answering, information extraction, sentiment analysis, text classification, and text generation. The innovative technique of reinforcement learning from human feedback (RLHF) [21] has been used to further align LLMs with human preferences, which has found applications in Artificial General Intelligence (AGI) models such as Instruct GP T [22], Sparrow [23], and ChatGPT [12]. More recently, GPT-4 has significantly advanced the state-of-the-art of language models, opening up new opportunities for LLM applications.

这些模型旨在从输入文本中学习精确的上下文潜在特征表示 [20]，可应用于问答、信息抽取、情感分析、文本分类和文本生成等多种场景。基于人类反馈的强化学习 (RLHF) [21] 这一创新技术被用于进一步使大语言模型与人类偏好对齐，该技术已在通用人工智能 (AGI) 模型中得到应用，例如 Instruct GPT [22]、Sparrow [23] 和 ChatGPT [12]。最近，GPT-4 显著推进了语言模型的技术前沿，为大语言模型应用开辟了新机遇。

Other than the applications in NLP domain, LLMs also show promising results and significant impacts in other disciplines such as biology [24], geography [25,26], agriculture [27], education [28,29], medical and health care [30, 31], and so on.

除了在自然语言处理(NLP)领域的应用外，大语言模型(LLM)在生物学[24]、地理学[25,26]、农业[27]、教育学[28,29]、医疗健康[30,31]等其他学科中也展现出显著成效和深远影响。

2.2 Public Health Info demi ology

2.2 公共卫生信息流行病学

Info demi ology [32] is a field that studies the determinants and distribution of information on the internet or in a population, with the goal of informing public health and public policy [32,9]. The term combines "information" and "epidemiology" and is a recognized approach in public health informatics, providing insights into health-related behaviors and perceptions. It plays a crucial role in monitoring and managing the information epidemic ("infodemic") associated with major public health crises.

信息流行病学 [32] 是研究互联网或人群中信息的决定因素与分布规律的学科，旨在为公共卫生和公共政策提供依据 [32,9]。该术语由"信息"和"流行病学"组合而成，是公共卫生信息学领域的公认方法，可揭示健康相关行为与认知。它在监测和管理重大公共卫生危机伴随的信息疫情 ("infodemic") 方面发挥着关键作用。

For example, Piamonte et al. [33] analyzed global search queries for Alzheimer’s disease (AD) using Google Trends data, comparing this online interest (Search Volume Index) with measures of disease burden. The study revealed that search behavior and interest in AD were influenced by factors like news about celebrities with AD and awareness months, and also highlighted potential correlations between this online interest and socioeconomic development.

例如，Piamonte等人[33]利用Google Trends数据分析了全球关于阿尔茨海默病(AD)的搜索查询，将这种在线关注度(搜索量指数)与疾病负担指标进行了比较。该研究表明，AD相关搜索行为和关注度受到名人患AD新闻和疾病宣传月等因素的影响，并揭示了这种在线关注度与社会经济发展之间可能存在的关联性。

With the rise of the internet and digital technologies, info demi ology provides a vital lens to examine the flow of health information and misinformation, helping public health practitioners develop effective communication strategies and interventions [34,35]. In the context of Alzheimer’s disease, understanding online behaviors and interests via info demi ology can help enhance public awareness, correct misconceptions, and inform preventative and management strategies for the disease [33,36].

随着互联网和数字技术的兴起，信息流行病学 (infodemiology) 为审视健康信息和错误信息的传播提供了重要视角，帮助公共卫生从业者制定有效的传播策略和干预措施 [34,35]。在阿尔茨海默病的背景下，通过信息流行病学理解在线行为和兴趣有助于提升公众认知、纠正误解，并为该疾病的预防和管理策略提供依据 [33,36]。

2.3 AutoGPT and LLM Automation

2.3 AutoGPT 与大语言模型自动化

The development and use of AutoGPT, LangChain $^{1}$ , and many other automation techniques for LLMs represent a significant advancement in the field of

AutoGPT、LangChain$^{1}$等大语言模型(LLM)自动化技术的开发与应用，标志着该领域的重大进展

NLP and artificial intelligence. AutoGPT builds on the successes of large language models like GPT-3 and GPT-4, but takes automation a step further by providing a more user-friendly interface for non-expert users [10].

自然语言处理与人工智能。AutoGPT基于GPT-3和GPT-4等大语言模型的成功，通过为非专业用户提供更友好的界面 [10]，将自动化推向新高度。

With AutoGPT, complex tasks such as data collection, data cleaning, analysis, and even the generation of human-like text can be completed using straightforward prompts, removing the need for extensive coding or data science expertise. This has the potential to democratize access to powerful language model technology, opening up new possibilities for research and application in a wide range of fields, including public health.

借助AutoGPT，数据收集、数据清洗、分析乃至生成类人文本等复杂任务，仅需简单提示即可完成，无需大量编码或数据科学专业知识。这一技术有望降低大语言模型的使用门槛，为公共卫生等广泛领域的研究与应用开辟新可能。

Recent studies [37,38] have highlighted the potential of AutoGPT and similar tools for automating the retrieval and analysis of large datasets. For example, with a well-formulated query, AutoGPT can be directed to crawl through a wide array of online platforms, collecting and analyzing comments, discussions, and posts pertaining to vaccines. The system would subsequently generate a summarizing report, outlining major themes of public opinion and prevalent misconceptions, thereby providing valuable insights for public health officials in formulating targeted communication and intervention strategies.

近期研究[37,38]指出AutoGPT等工具在自动化检索与分析大规模数据集方面的潜力。例如，通过精心设计的查询指令，可引导AutoGPT爬取各类网络平台，收集并分析与疫苗相关的评论、讨论及帖子。该系统随后会生成总结报告，概述舆论主要观点及常见误解，从而为公共卫生官员制定精准传播与干预策略提供关键洞见。

In the context of info demi ology, AutoGPT can automate the process of analyzing online health information trends, which traditionally involves extensive manual effort. Specifically, it can efficiently scan and interpret internet data, track the spread of health information and misinformation, assess public reaction to health policies or events, and potentially predict future trends.

在信息流行病学 (infodemiology) 背景下，AutoGPT 可自动化分析在线健康信息趋势这一传统上需要大量人工干预的过程。具体而言，它能高效扫描和解析互联网数据、追踪健康信息与错误信息的传播路径、评估公众对卫生政策或事件的反应，并具备预测未来趋势的潜力。

2.4 Improving Autonomous LLM-based Tools for Public Health

2.4 提升基于大语言模型的公共卫生自主工具

While recognizing the potential of autonomous large language models (LLMs) like AutoGPT in public health research and practice, we identified certain limitations in their current state that may hinder their efficacy in particular use cases, such as info demi ology. By tailoring these tools to the specific needs of public health professionals, we aim to enhance their utility in these contexts.

在认识到诸如AutoGPT等自主大语言模型(LLM)在公共卫生研究和实践中的潜力的同时，我们也发现了它们在当前状态下存在的一些局限性，这些局限性可能会阻碍其在特定用例(如信息流行病学)中的有效性。通过针对公共卫生专业人员的具体需求定制这些工具，我们旨在提升它们在这些场景中的实用性。

Firstly, despite AutoGPT’s extensive searching capabilities, its ability to acquire specialized information quickly and precisely, for instance, about Alzheimer’s disease (AD), can be somewhat limited. In response to this, we have integrated specific prompting mechanisms in our model, AD-AutoGPT. These tailored prompts direct AD-AutoGPT to gather data from a select list of authoritative websites relevant to AD, which enhances the efficiency and relevance of information acquisition.

首先，尽管AutoGPT具备广泛的搜索能力，但其快速精准获取专业信息（例如阿尔茨海默病(AD)相关）的能力仍存在一定局限。为此，我们在AD-AutoGPT模型中整合了特定提示机制。这些定制化提示会引导AD-AutoGPT从精选的AD权威网站列表中采集数据，从而提升信息获取的效率和相关性。

Secondly, Our AD-AutoGPT model also addresses the challenge AutoGPT faces in extracting critical details such as the time and place of news events from articles accurately. AD-AutoGPT uses web-crawling scripts to extract accurate timestamps from news pieces, and employs geo-location libraries such as geopy [39] and geopandas [40] to retrieve precise location information from texts.

其次，我们的AD-AutoGPT模型还解决了AutoGPT在从文章中准确提取新闻事件时间和地点等关键细节时面临的挑战。AD-AutoGPT使用网络爬虫脚本从新闻中提取准确的时间戳，并利用geopy [39]和geopandas [40]等地理定位库从文本中检索精确的位置信息。

Thirdly, depth of analysis is another area where AutoGPT could benefit from further refinement. Owing to the token limit in models like ChatGPT,

第三，分析深度是AutoGPT可以进一步改进的另一个领域。由于ChatGPT等模型中的Token限制，

AutoGPT’s analysis is often restricted to the first 4096 tokens [12]. Consequently, it might miss core content or important details. To overcome this limitation, AD-AutoGPT segments the text, vectorizes it, and then processes these chunks independently. It creates summaries for each of these segments and then amalgamates these summaries to create a comprehensive representation of the news article.

AutoGPT的分析通常局限于前4096个token [12]，因此可能遗漏核心内容或重要细节。为突破这一限制，AD-AutoGPT采用文本分块、向量化处理后独立处理各片段的方式。它会为每个文本块生成摘要，再将这些摘要整合形成新闻文章的完整表征。

Fourthly, AutoGPT’s current capabilities, while useful, lack the capacity to conduct an in-depth analysis of the generated summaries. The synthesized data can still be redundant and may not accurately capture the most essential information. In contrast, AD-AutoGPT applies Latent Dirichlet Allocation (LDA) [41] to extract the most pertinent keywords from the text summaries, offering users a succinct understanding of the central themes in the Alzheimer’s disease domain.

第四，AutoGPT目前的功能虽然实用，但缺乏对生成摘要进行深入分析的能力。合成的数据可能仍然冗余，且未必能准确捕捉最关键的信息。相比之下，AD-AutoGPT采用潜在狄利克雷分配 (LDA) [41] 从文本摘要中提取最相关的关键词，为用户提供对阿尔茨海默病领域核心主题的简明理解。

Lastly, while AutoGPT is effective at generating text-based information, it lacks robust visualization capabilities. Addressing this limitation, AD-AutoGPT integrates dynamic visualization techniques, creating plots of news occurrences over time, highlighting locations where news events are happening, and even illustrating the evolution of research keywords over time.

最后，虽然AutoGPT在生成基于文本的信息方面很有效，但它缺乏强大的可视化能力。针对这一局限，AD-AutoGPT集成了动态可视化技术，可绘制新闻事件随时间变化的图表、突出显示新闻事件发生的地点，甚至展示研究关键词随时间的演变趋势。

AD-AutoGPT is refined through the application of domain-specific knowledge and technical adjustments to optimize its relevance and effectiveness for public health researchers and practitioners. As a result, AD-AutoGPT is faster and more efficient in its operations compared to the original AutoGPT, highlighting the advantages of tailoring autonomous LLM-based tools for specific use cases in public health.

AD-AutoGPT通过应用领域特定知识和技术调整进行优化，以提升其在公共卫生研究人员和实践者中的相关性和有效性。因此，与原始AutoGPT相比，AD-AutoGPT运行速度更快、效率更高，突显了针对公共卫生特定用例定制基于大语言模型的自主工具的优势。

3 Method

3 方法

In this section, we will introduce AD-AutoGPT, an LLM-based tool we developed to automate the process of Alzheimer’s Disease Info demi ology. ADAutoGPT uses the Langchain framework to realize the connection with GPT-4 and ChatGPT API, and establish an LLM-based autonomous framework with a chain of thinking mode for Alzheimer’s disease. This is a model that can automatically search for the latest news, extract meaningful spatio-temporal data, summarize the news, analysis news content, and visualize analysis results. The overall framework of AD-AutoGPT is shown in Figure 1. We construct an instruction library that contains a set of possible commands/tools we have developed to achieve the public health info demi ology task. A prompt shown in Figure 2a is designed to facilitate LLMs to identify usable tools from the instruction library and form a data processing pipeline that demonstrates the process of thinking. AD-AutoGPT’s ability of “translating” natural language prompts to real data processing pipeline is similar to the idea of semantic parsing used in traditional question answering literature [42,43,44], which aims at translating a natural language question into an executable query for a given database or knowledge base. However, the difference is that semantic parsing is only able to generate rather simple executable queries on a well-defined knowledge base while our AD-AutoGPT can handle much more complex realworld tasks such as searching and collecting news from Google, analyzing new contents, and visualizing topic trends and spatial-temporal distributions of news. Below we will introduce the workflow of AD-AutoGPT and the basic principles of the algorithms used in the workflow in detail.

在本节中，我们将介绍AD-AutoGPT，这是我们开发的一款基于大语言模型的工具，用于自动化阿尔茨海默病信息流行病学处理流程。AD-AutoGPT利用Langchain框架实现与GPT-4和ChatGPT API的对接，构建了一个基于大语言模型、采用思维链模式的阿尔茨海默病自主分析框架。该模型能够自动搜索最新资讯、提取有价值的时空数据、生成新闻摘要、分析内容并可视化分析结果。AD-AutoGPT的整体框架如图1所示。

我们构建了一个指令库，其中包含为完成公共卫生信息流行病学任务而开发的一系列命令/工具。如图2a所示的提示词设计用于帮助大语言模型从指令库中识别可用工具，并形成展现思维过程的数据处理流程。AD-AutoGPT将自然语言提示"翻译"为实际数据处理流程的能力，与传统问答文献[42,43,44]中使用的语义解析思想类似，后者旨在将自然语言问题转化为针对特定数据库或知识库的可执行查询。不同之处在于，语义解析只能在结构明确的知识库上生成相对简单的可执行查询，而我们的AD-AutoGPT能处理更复杂的现实任务，例如从Google搜索采集新闻、分析内容、可视化主题趋势和新闻时空分布等。下文将详细说明AD-AutoGPT的工作流程及所用算法的基本原理。

Fig. 1: The basic framework of AD-AutoGPT. The instruction library contains a set of possible commands we have developed to complete the public health info demi ology task. These commands can also be expanded in the future. In order to achieve the goal, AD-AutoGPT will access GPT-4 and divide the final goal into several smaller tasks, and then solve small tasks step-by-step by choosing the most appropriate command for the sub-task in the instruction library. After thinking and judging, if the final goal has not been achieved, AD-AutoGPT will continue to split the task and execute the command. If the final goal has been achieved, AD-AutoGPT will return the final answer.

图 1: AD-AutoGPT的基本框架。指令库包含一组我们开发的用于完成公共卫生信息流行病学任务的命令，这些命令未来还可扩展。为实现目标，AD-AutoGPT会访问GPT-4并将最终目标拆分为若干子任务，随后通过从指令库中选择最适合子任务的命令逐步解决这些小任务。经过思考判断后，若最终目标未达成，AD-AutoGPT会继续拆分任务并执行命令；若目标已达成，则返回最终答案。

3.1 Overall Framework

3.1 整体框架

Our primary goal is to learn from the chain thinking mode of AutoGPT to realize the automatic collection and summary of Alzheimer’s disease news. To achieve this goal, the power of LLMs must be used. Advanced LLMs such as ChatGPT and GPT-4 have brought earth-shaking changes to the NLP domain, and we see the potential advantages of LLMs for the public health field.

我们的主要目标是借鉴AutoGPT的链式思考模式，实现阿尔茨海默病新闻的自动采集与汇总。为实现这一目标，必须借助大语言模型的力量。ChatGPT、GPT-4等先进大语言模型已为自然语言处理领域带来颠覆性变革，我们看到了这类模型在公共卫生领域的潜在优势。

The overall framework is shown in Figure 1. For the target task, ADAutoGPT will use ChatGPT or GPT-4 to divide the target task into several small tasks and process them separately. We provide AD-AutoGPT with an instruction library which contains customized functions/tools including:

整体框架如图 1 所示。针对目标任务，ADAutoGPT 会使用 ChatGPT 或 GPT-4 将目标任务拆分为若干小任务并分别处理。我们为 AD-AutoGPT 提供了包含以下定制化功能/工具的指令库：

After operating every small task choosing from these tools, AD-AutoGPT will judge whether the overall goal has been achieved according to the running results of the function, or it needs to think again and solve the next small problem. Chain thinking is realized through such a pattern. If during the process AD-AutoGPT thinks that the system has reached the initial goal, the system will exit and return a final answer to the initial question.

在从这些工具中选择并执行每个小任务后，AD-AutoGPT会根据函数的运行结果判断整体目标是否达成，或者是否需要重新思考并解决下一个小问题。通过这种模式实现了链式思考。如果AD-AutoGPT认为系统已达到初始目标，系统将退出并向初始问题返回最终答案。

3.2 Designing Prompts to Implement Chain of Thoughts

3.2 设计提示以实现思维链

A prompt example can be seen in Figure 2a and the model thinking process of AD-AutoGPT is shown in Figure 2b. According to the input, this prompt has four parts in the task process which are question, thought, action, and action input.

提示示例如图 2a 所示，AD-AutoGPT 的模型思考过程如图 2b 所示。根据输入内容，该提示在任务过程中包含四个部分：问题 (question)、思考 (thought)、行动 (action) 和行动输入 (action input)。

For output, a prompt has three parts which are observation, thought, and final answer.

输出时，提示包含三个部分：观察、思考和最终答案。

The last part of the prompt is the question entered by the user, such as the question in Figure 2a, "Can you help me to know something new about Alzheimer’s Disease and maybe draw some plots for me? ". AI will decompose the complex target tasks proposed by users into several simple tasks, thus inspiring a chain of thoughts. And the thinking process of AI can be seen in Figure 2b

提示的最后部分是用户输入的问题，例如图 2a 中的问题 "你能帮我了解一些关于阿尔茨海默病的新知识，并为我绘制一些图表吗？"。AI 会将用户提出的复杂目标任务分解为几个简单任务，从而激发思维链。AI 的思考过程可以在图 2b 中看到。

Owing to this set of prompts, we can ensure that the thinking logic of ADAutoGPT does not deviate from the right track and make the whole chain of thoughts visible to users.

得益于这组提示词，我们可以确保 ADAutoGPT 的思维逻辑不会偏离正轨，并让整个思维链对用户可见。

Question: Can you help me to know something new about Alzheimer’s Disease and maybe draw some plots for me?

问题：你能帮我了解一些关于阿尔茨海默病的新知识吗？或许还能帮我绘制一些图表？

(a) An instance of prompt specifies the format in which the AI answers questions.

(a) 提示实例规定了AI回答问题的格式。

(b) An example of AI thinking and calling functions to solve user problems

(b) AI思考并调用函数解决用户问题的示例

Fig. 2: The prompt of AD-AutoGPT, the AI assistant will answer the question based on the given format and can use the specified functions. In the prompt, tools represent the functions that AD-AutoGPT can call, including tool_ names, tool descriptions and so on.

图 2: AD-AutoGPT的提示词(prompt)，AI助手将根据给定格式回答问题，并可调用指定功能。提示词中的tools代表AD-AutoGPT可调用的功能，包括tool_ names、tool描述等。

3.3 Text Summary

3.3 文本摘要

To achieve the purpose of extracting the most critical information from a large amount of news text, AD-AutoGPT performs new text summary and LDA topic modeling.

为了实现从大量新闻文本中提取最关键信息的目的，AD-AutoGPT执行了新闻文本摘要和LDA主题建模。

The text summary is mainly achieved by accessing ChatGPT or GPT4 API. Owing to the powerful text sum mari z ation ability of GPT-4, ADAutoGPT can make more efficient use of text than other models. AD-AutoGPT traverses the saved news URLs one by one, and then saves the text from the website by calling the web crawler scripts. Next, it uses ChatGPT or GPT-4 to summarize the news text. It is worth mentioning that because LLMs have a token limit, all the text here will be pre-processed first, and then be summarized. More specifically, since GPT-4 has a limit on the number of tokens, in order to summarize the complete news text, we use the map_ reduce method to process it [10].

文本摘要主要通过访问ChatGPT或GPT4 API实现。得益于GPT-4强大的文本摘要能力，ADAutoGPT能比其他模型更高效地利用文本内容。AD-AutoGPT会逐个遍历已保存的新闻URL，通过调用网络爬虫脚本保存网站文本，随后使用ChatGPT或GPT-4对新闻文本进行摘要。值得注意的是，由于大语言模型存在token限制，所有文本都会先进行预处理再进行摘要。具体而言，鉴于GPT-4存在token数量限制，为完整摘要新闻文本，我们采用map_ reduce方法进行处理 [10]。

3.4 S patio temporal Information Extraction

3.4 时空信息提取

Next, AD-AutoGPT will perform s patio temporal information extraction on the collected news articles. The temporal information can be easily extracted from news metadata while extracting place mentions from news articles is a kind of oral. Here, we adopt the geoparsing approach [45,46] which first recognizes place names from raw text, so-call toponym recognition [47,25] and then link the recognized place names to a specific geographic entity in an existing gazetteer or geospatial knowledge graphs [48,49], so-called toponym resolution [50], so that the spatial footprints (i.e., geographic coordinates) of these places can be obtained. More specifically, we use GeoText $^2$ , a pythonbased geoparsing tool to achieve this goal.

接下来，AD-AutoGPT将对收集的新闻文章进行时空信息提取。时间信息可以从新闻元数据中轻松提取，而从新闻文章中提取地点提及则属于口语化任务。为此，我们采用地理解析方法[45,46]，该方法首先从原始文本中识别地名（即地名识别[47,25]），然后将识别出的地名链接到现有地名录或地理空间知识图谱[48,49]中的特定地理实体（即地名解析[50]），从而获取这些地点的空间足迹（即地理坐标）。具体而言，我们使用基于Python语言的地理解析工具GeoText$^2$来实现这一目标。

3.5 LDA Analysis

3.5 LDA分析

Latent Dirichlet Allocation (LDA) [41] is a probabilistic topic model. LDA can give a probability distribution of topics of each document in the corpus. By analyzing a batch of document sets and extracting their topic distributions, topic clustering can be performed according to the topic distribution. LDA is a typical bag-of-words model, that is, a document is interpreted as a set of words, and there is no sequential relationship among words. In addition, a document can contain multiple topics, and each word in the document is assumed to be generated by one of the topics. LDA is an unsupervised learning method that does not require a manually labeled training set during training but only needs a document set and the total number of topics $K$ . In addition, another advantage of LDA is that every topic is associated with a set of most frequent keywords which can be used to interpret this topic.

潜在狄利克雷分配 (Latent Dirichlet Allocation, LDA) [41] 是一种概率主题模型。LDA能够给出语料库中每篇文档的主题概率分布，通过分析一批文档集并提取其主题分布，可根据主题分布进行主题聚类。LDA是典型的词袋模型 (bag-of-words model)，即文档被视作一组词的集合，词与词之间没有顺序关系。此外，一篇文档可以包含多个主题，且文档中的每个词都被假定由其中一个主题生成。LDA属于无监督学习方法，训练时不需要人工标注的训练集，仅需文档集和主题总数$K$。LDA的另一优势在于每个主题都关联着一组高频关键词，这些关键词可用于解释该主题。

In short, AD-AutoGPT uses LDA topic modeling to discover the topics for the summary text of each piece of collected news, For each topic, the keyword with the highest frequency of occurrence and the highest weight will be displayed to the user.

简而言之，AD-AutoGPT采用LDA主题建模(Latent Dirichlet Allocation)技术分析每条新闻摘要文本的主题分布，针对每个主题向用户展示出现频率最高且权重最大的关键词。

4 Case Study and Experimental Results

4 案例研究与实验结果

4.1 Alzheimer’s Disease News Information Retrieval

4.1 阿尔茨海默病新闻信息检索

The effectiveness of our proposed AD-AutoGPT is mainly verified on the data provided by the most authoritative websites reporting Alzheimer’s disease, which are Alzheimer’s Association, BBC, National Institute of Aging, and Mayo Clinic. By using the prompt shown in Figure 2a, we are able to instruct the LLM (e.g., ChatGPT or GPT-4) to search for the right tool in our instruction library – Search and Save News to achieve the first news data collection step.

我们提出的AD-AutoGPT的有效性主要通过在阿尔茨海默病领域最权威的网站提供的数据进行验证，这些网站包括阿尔茨海默病协会、BBC、国家老龄化研究所和梅奥诊所。通过使用图2a所示的提示，我们能够指导大语言模型（例如ChatGPT或GPT-4）在我们的指令库中搜索合适的工具——搜索并保存新闻，以实现新闻数据收集的第一步。

We have collected 277 news in total from these four websites in the period of last year. On this actual news dataset, we validate the functions of ADAutoGPT for text extraction, text sum mari z ation, spatio-temporal-data analysis, hot topics analysis, and result visualization. In this process, the time and location of the news will also be extracted and saved. Note that AD-AutoGPT automatically uses the given prompt and formalizes a data collection and processing pipeline based on the toolsets in our instruction library without any human intervention.

我们从这四个网站收集了去年全年的277条新闻。在这个真实新闻数据集上，我们验证了ADAutoGPT的文本提取、文本摘要、时空数据分析、热点话题分析和结果可视化功能。在此过程中，新闻的时间和地点信息也会被提取并保存。需要注意的是，AD-AutoGPT会自动使用给定的提示词(prompt)，并根据我们指令库中的工具集构建数据收集和处理流程，全程无需人工干预。

4.2 S patio temporal Information Extraction and Visualization

4.2 时空信息提取与可视化

(b) The number of news collected for each month from June 2022 to May 2023.

图 1:
(b) 2022年6月至2023年5月每月收集的新闻数量。

(a) Places where the latest news about Alzheimer’s diseases happened.

(a) 阿尔茨海默病最新动态发生地。

Fig. 3: The visualization of the results from the spatial and temporal information extraction. (a) shows the spatial distribution of the Alzheimer’s disease news. The news mainly happened in America and Western Europe. (b) shows the temporal change in the number of news occurrences from June 2022 to May 2023.

图 3: 时空信息提取结果的可视化。(a) 展示了阿尔茨海默病新闻的空间分布。新闻主要发生在美国和西欧。(b) 展示了2022年6月至2023年5月新闻数量的时序变化。

Based on the given prompt, AD-AutoGPT decides to use Extract Spatial Data tool and Extract Temporal Data tool in our instruction library (see Figure 1) to extract the places where these news articles mentioned and the timestamps when these news articles were posted online.

根据给定的提示，AD-AutoGPT决定使用我们指令库中的提取空间数据工具和提取时间数据工具(见图1)来提取这些新闻文章中提到的地方以及这些新闻文章在网上发布的时间戳。

The spatial locations of extracted places from all news articles are visualized in Figure 3a. Note that this map visualization is automatically generated by AD-AutoGPT based on the prompt shown in Figure 2b. It can be seen that most of the news articles about Alzheimer’s Disease in the past year mainly occurred in the United States and Western Europe. For the BBC, although it basically only reports Alzheimer’s disease news in the UK, the total number of news is not inferior to that of other websites. Similarly, websites in the United States such as NIA also pay more attention to local news, especially in the southeastern states of the United States. For the Alzheimer’s Association, the sources of news reports are relatively scattered all over the world, while the United States and Western Europe still show higher report frequencies than other regions such as South America, Africa, Australia, and so on. Finally, for Mayo Clinic, since there is less news from this news source, only a few occurrences can be seen on the map. Generally speaking, the distribution of news is worldwide, but it is concentrated in the southeastern United States and Western Europe. These might be because of the select bias of those four news media we use or the well-developed Alzheimer’s disease research in these regions.

从所有新闻文章中提取地点的空间位置如图3a所示。请注意，该地图可视化是由AD-AutoGPT根据图2b所示的提示自动生成的。可以看出，过去一年关于阿尔茨海默病的新闻文章主要集中在美国和西欧。对于BBC来说，虽然基本上只报道英国的阿尔茨海默病新闻，但新闻总数并不逊色于其他网站。同样，美国的网站如NIA也更关注本地新闻，尤其是美国东南部各州。阿尔茨海默病协会的新闻报道来源相对分散在世界各地，但美国和西欧的报道频率仍高于南美、非洲、澳大利亚等其他地区。最后，对于Mayo Clinic来说，由于该新闻来源的新闻较少，地图上只能看到少数几个地点。总的来说，新闻的分布是全球性的，但集中在美国东南部和西欧。这可能是因为我们使用的这四家新闻媒体的选择偏差，或者这些地区的阿尔茨海默病研究较为发达。

Temporal data analysis results can be seen in Figure 3b. The numbers of news reports about Alzheimer’s disease in each month of the past year (June 2022 to May 2023) are visualized. It can be seen that the overall trend of the number of news reports is declining, from 31 in a single month in June 2022 to 13 in May 2023. It can also be seen that September, October, and November 2022 are the period of high incidences of news reports. The number of news reports in each of the three months exceeded 27, and those in September 2022 reached 32, which was the highest in 2022. This might be because there was news that had a profound impact on AD-related media during this period, resulting in a sudden increase in reports, which deserves special attention from users.

时间数据分析结果如图 3b 所示。过去一年(2022年6月至2023年5月)每月关于阿尔茨海默病的新闻报道数量被可视化呈现。可以看出，新闻报道数量总体呈下降趋势，从2022年6月的单月31篇降至2023年5月的13篇。同时可见2022年9月、10月和11月是新闻报道的高发期，这三个月的报道量均超过27篇，其中2022年9月达到32篇，为全年最高值。这可能与该时期出现了对AD相关媒体影响深远的新闻事件有关，导致报道量激增，值得使用者特别关注。

Therefore, it can be seen that AD-AutoGPT can not only extract useful s patio temporal information from a wide range of news sources but can also use the visualization function to more intuitively display the spatial distribution of the AD-related news and their development through time which might be useful for users. We need to emphasize that these s patio temporal analyses was done by AD-AutoGPT without any human input. Thereby AD-AutoGPT improves the efficiency of researchers’ work, which AutoGPT cannot do because it does not design functions of information extraction from web pages.

因此可以看出，AD-AutoGPT不仅能从广泛的新闻来源中提取有效的时空信息，还能通过可视化功能更直观地展示AD相关新闻的空间分布及其时间演变趋势，这对用户可能具有实用价值。需要强调的是，这些时空分析完全由AD-AutoGPT自主完成，无需任何人工输入。因此AD-AutoGPT提升了研究人员的工作效率，而AutoGPT由于未设计网页信息提取功能，无法实现此类操作。

(a) The word count trend of each topic obtained from the LDA results.

(a) 从LDA结果中获取的各主题词频趋势。

(b) The word importance trend of each topic obtained from the LDA results.

(b) 从LDA结果中获得的各主题词重要性趋势。

Fig. 4: For each Topic, the Streamplot graph displays the occurrence times and frequency of different keywords in different time periods.

图 4: 每个主题的流线图展示了不同关键词在不同时间段内的出现时间及频率。

4.3 LDA Topic Modeling and Hot Topic Analysis

4.3 LDA主题建模与热点话题分析

Based on the LDA topic modeling, a hot topic analysis is automatically conducted by AD-AutoGPT. The results can be seen in Figure 4. AD-AutoGPT aggregates the summaries of the news reported in the past year for LDA analysis, and finally got 5 hot topics. It selects the top 5 words with the most occurrences for each of the 5 hot topics and draws stream graphs according to the number of occurrences and word weights of the words. Please refer to Figure 4a and 4b. In this way, you can see the changes in topic distributions according to time, so as to quickly understand the trend of the research topic.

基于LDA主题建模，AD-AutoGPT自动进行了热点话题分析。结果如图4所示。AD-AutoGPT汇总了过去一年新闻报道的摘要进行LDA分析，最终得到5个热点话题。它从这5个热点话题中各选取出现频率最高的5个词，并根据这些词的出现次数和词权重绘制了流图。具体请参见图4a和图4b。通过这种方式，可以观察到话题分布随时间的变化，从而快速把握研究主题的趋势。

It can be found that the keywords of the first hot topic are mainly protein, lipid, and drug, and this type of topic has occupied the largest weight in the past year, which shows that scientists are mostly concerned about seeking reliable drug treatment for Alzheimer’s disease. The keywords of the topic with the second highest proportion are individual, treatment, amyloid, and tissue. This topic is also about the drug treatment of Alzheimer’s disease, but the focus has obviously shifted from the research and development of new drugs to the current personal medication, reflecting the patients’ concerns about self-care. The keywords of the third-ranked topic include sleep, brain, blood, cell, etc. This type of news mainly focuses on the causes of Alzheimer’s disease, which is similar to popular science news. It can be seen that journalists have attached great importance to popular science in the past year. For the fourth-ranked topic, the keywords are increase, future, disorder, future, etc. This topic is mostly related to the future plan or expectation for Alzheimer’s disease research. The keywords of the last topic are mainly diagnosis, caregiver, vitamin, etc., reflecting the public’s concerns about the diagnosis, care and prevention of Alzheimer’s disease.

可以发现，第一个热门话题的关键词主要是蛋白质、脂质和药物，这类话题在过去一年中占据了最大权重，这表明科学家们主要关注为阿尔茨海默病寻求可靠的药物治疗。占比第二高的话题关键词包括个体、治疗、淀粉样蛋白和组织。该话题同样涉及阿尔茨海默病的药物治疗，但关注点显然已从新药研发转向当前的个人用药，反映出患者对自我护理的关注。排名第三的话题关键词包含睡眠、大脑、血液、细胞等，这类新闻主要聚焦阿尔茨海默病的病因，类似于科普新闻。可见过去一年中，记者们对科普工作给予了高度重视。排名第四的话题关键词为增长、未来、紊乱等，该话题多与阿尔茨海默病研究的未来计划或预期相关。最后一个话题的关键词主要是诊断、护理人员、维生素等，反映了公众对阿尔茨海默病的诊断、护理和预防方面的关注。

Therefore, we can conclude that through hot topic analysis, we can easily get the popular topics in the news during June 2022 - May 2023 period by using AD-AutoGPT’s autonomous workflow. Users no longer need to read extensively on news, but they can easily use the help of AD-AutoGPT to understand the hot topics of Alzheimer’s disease in the past period so that the efficiency of work and research on Alzheimer’s disease is greatly improved. Owing to GPT-4’s powerful summarizing ability, in the future, the work of early information collection can be completely handed over to AI. Humans only need to judge and focus on the most critical information returned by AI to quickly understand the development and changes in the public health domain, thus saving time and resources.

因此，我们可以得出结论：通过热点话题分析，利用AD-AutoGPT的自主工作流，能够轻松获取2022年6月至2023年5月期间新闻中的热门话题。用户不再需要大量阅读新闻，而是可以便捷地借助AD-AutoGPT了解过去阶段阿尔茨海默病的热点议题，从而显著提升阿尔茨海默病相关工作和研究的效率。得益于GPT-4强大的摘要生成能力，未来早期信息收集工作可完全交由AI处理。人类只需对AI返回的最关键信息进行判断和聚焦，即可快速把握公共卫生领域的发展动态，实现时间和资源的高效节约。

5 Discussion and Conclusion

5 讨论与结论

5.1 Automating Data Analytics

5.1 数据分析自动化

The success of AD-AutoGPT shows the transformative potential of LLMs in the public health domain. By harnessing the advanced linguistic understanding and autonomous operations of AD-AutoGPT, we were able to streamline the analytical process and conduct comprehensive analyses of extensive news sources related to Alzheimer’s Disease (AD). Moreover, AD-AutoGPT has the potential to go beyond the public health domain and be applied in various disciplines.

AD-AutoGPT的成功展示了大语言模型在公共卫生领域的变革潜力。通过利用AD-AutoGPT先进的语言理解能力和自主操作功能，我们得以简化分析流程，并对与阿尔茨海默病(AD)相关的大量新闻来源进行全面分析。此外，AD-AutoGPT还有潜力超越公共卫生领域，应用于多个学科。

One of the key advantages of autonomous LLM-based tools such as AutoGPT and AD-AutoGPT is their ability to automate and optimize complex data extraction and analysis tasks, as well as transcending traditional laborintensive methods. This enables researchers and professionals across different fields to access and engage with large language models, empowering them to conduct sophisticated analyses efficiently, regardless of their technical expertise.

基于大语言模型的自主工具（如AutoGPT和AD-AutoGPT）的核心优势在于能自动化并优化复杂的数据提取与分析任务，同时突破传统劳动密集型方法的局限。这使得各领域的研究人员和专业人士无需专业技术背景，即可高效访问并运用大语言模型进行复杂分析。

5.2 Prioritizing Insights and Innovation

5.2 优先考虑洞察与创新

Through the development of AD-AutoGPT, we conduct a detailed trend analysis, intertopic distance mapping, and identified salient terms relevant to AD. These findings provide valuable insights into the shifting focus and narrative surrounding AD, not only in the domain of public health but also in broader contexts. By quantifying and visualizing the discourse, we gain a nuanced understanding of the prevalent topics, concerns, and perspectives related to AD, facilitating targeted interventions, communication strategies, and decision-making across multiple fields.

通过开发AD-AutoGPT，我们进行了详细的趋势分析、主题间距离映射，并识别出与AD相关的突出术语。这些发现不仅为公共卫生领域，还在更广泛的背景下，提供了关于AD关注焦点和叙述变化的宝贵见解。通过量化和可视化讨论内容，我们对与AD相关的主流话题、关注点和观点有了细致入微的理解，从而促进跨多个领域的针对性干预、沟通策略和决策制定。

The integration of AutoGPT and other autonomous LLM-based tools into research across different disciplines represents a significant advancement. By automating data analysis tasks, researchers can dedicate more time and resources to interpreting the results and deriving actionable insights. This accelerates the research process and enhances the accuracy and reliability of the findings in diverse areas, such as social sciences, economics, technology, and more.

AutoGPT和其他基于大语言模型的自主工具融入跨学科研究，标志着重大技术进步。通过自动化数据分析任务，研究人员能将更多时间和资源投入到结果解读与可执行洞察的提炼中。这不仅加快了社会科学、经济学、技术等多元领域的研究进程，还提升了研究结果的准确性与可靠性。

5.3 Transforming Public Health

5.3 公共卫生转型

Furthermore, the insights obtained from this research have broader implications beyond public health. The automation capabilities of AD-AutoGPT can revolutionize the field of info demi ology by efficiently analyzing online information trends, tracking the dissemination of information and misinformation, and predicting future trends. This has the potential to inform evidence-based interventions, enhance communication strategies, and combat misinformation across various domains.

此外，这项研究的发现对公共卫生领域之外也具有广泛意义。AD-AutoGPT的自动化能力可以通过高效分析在线信息趋势、追踪信息与错误信息的传播路径以及预测未来趋势，彻底改变信息流行病学 (infodemiology) 领域。这种技术有望为基于证据的干预措施提供依据、优化传播策略，并在多个领域对抗错误信息。

(注：根据术语处理规则，"infodemiology"为首次出现专业术语，故在括号内保留英文原词；"AD-AutoGPT"作为系统名称保留不译；"evidence-based"采用中文医学领域通用译法"基于证据的")

While our AD-AutoGPT has made significant strides in utilizing autonomou LLM-based tools for AD analysis in the public health domain, there are still areas for further exploration and improvement. For example, based on different underlying path o logie s, AD-related dementias (ADRD) can be categorized as four major types: prion disease, AD, fronto temporal lobar degeneration (FTLD), and Lewy body diseases (LBD). In practical clinical settings, diffe rent i at ions among these subtypes of dementias are very challenging, due to both mixed path o logie s and clinical symptoms. Our proposed AD-AutoGPT is a general framework and can be easily extended and refined to adapt to other dementias and various brain disorders. Future studies could also focus on expanding the dataset to include a broader range of sources and different languages to capture a more comprehensive understanding of the global discourse on different dementias across different fields. Additionally, exploring the integration of AD-AutoGPT with other data sources, such as social media platforms and electronic records, could provide a more holistic perspective on ADRD conversations and outcomes across multiple disciplines.

虽然我们的AD-AutoGPT在利用基于大语言模型的自主工具进行公共卫生领域AD分析方面取得了重大进展，但仍有一些领域值得进一步探索和改进。例如，根据不同的潜在病理机制，AD相关痴呆(ADRD)可分为四大类型：朊病毒病、AD、额颞叶变性(FTLD)和路易体病(LBD)。在实际临床环境中，由于混合病理和临床症状的存在，这些痴呆亚型之间的鉴别极具挑战性。我们提出的AD-AutoGPT是一个通用框架，可以轻松扩展和优化以适配其他痴呆症及各类脑部疾病。未来研究还可聚焦于扩展数据集，纳入更广泛的来源和不同语言，以更全面地把握跨领域全球痴呆症讨论动态。此外，探索将AD-AutoGPT与社交媒体平台、电子记录等其他数据源整合，可为跨学科ADRD讨论及结果提供更全面的观察视角。

5.4 基于自主大语言模型工具的伦理问题

In the use of autonomous LLM-based tools, several ethical issues arise that warrant careful consideration. First, these models generate output based on their training data, which if biased or discriminatory, could result in outputs that perpetuate such biases [51,52]. Ethical considerations must therefore include the selection and handling of training data of LLMs to minimize the risk of biased or inappropriate outputs.

在使用基于大语言模型(LLM)的自主工具时，几个伦理问题值得仔细考量。首先，这些模型根据训练数据生成输出，若数据存在偏见或歧视性，可能导致输出延续此类偏见[51,52]。因此伦理考量必须包含大语言模型训练数据的选择与处理，以最小化偏见或不恰当输出的风险。

In addition, issues of privacy and consent are paramount, particularly when dealing with sensitive data such as health information [30]. Even though LLMs do not remember specific inputs or retain personal data, the potential misuse of these tools can lead to leaking private or sensitive information , which raises significant ethical and legal questions.

此外，隐私与授权问题至关重要，尤其是在处理健康信息等敏感数据时 [30]。尽管大语言模型不会记忆特定输入或保留个人数据，但这些工具的潜在滥用仍可能导致私密或敏感信息泄露，由此引发重大伦理和法律问题。

Moreover, the potential for misuse extends to the propagation of false information or misinformation [53,54], a concern that is especially salient in the context of public health. LLMs can generate plausible-sounding but factually incorrect or misleading information [12,29], which, if not properly managed, could have severe consequences.

此外，误用还可能涉及传播虚假信息或错误信息 [53,54]，这一问题在公共卫生领域尤为突出。大语言模型能够生成看似合理但实则错误或具有误导性的内容 [12,29]，若管理不当，可能造成严重后果。

Finally, the democratization of powerful technologies like AutoGPT also raises questions about responsibility and oversight. As these tools become more accessible and widespread, ensuring appropriate use and managing the potential for misuse becomes increasingly challenging.

最后，像AutoGPT这样的强大技术普及化也引发了责任与监管问题。随着这些工具变得更易获取和广泛使用，确保合理应用并管控潜在滥用风险变得愈发具有挑战性。

Addressing these ethical issues is essential for the responsible development and deployment of autonomous LLM-based tools. This includes the development of robust guidelines for data handling, the implementation of safeguards against misuse, the provision of clear user instructions and warnings about potential pitfalls, and ongoing efforts to refine and improve these tools in light of user feedback and societal needs. The goal should be to harness the potential of these technologies while mitigating risks and adverse impacts, striking a balance between technology innovation and ethical responsibility.

解决这些伦理问题对于负责任地开发和部署基于大语言模型的自主工具至关重要。这包括制定稳健的数据处理指南、实施防止滥用的保障措施、提供明确的用户说明和潜在风险的警告，以及根据用户反馈和社会需求不断完善这些工具。目标应是在发挥这些技术潜力的同时降低风险和负面影响，在技术创新与伦理责任之间取得平衡。

6 Conclusion

6 结论

In conclusion, this study proposes a transformative autonomous LLM-based tool called AD-AutoGPT which can facilitate data-driven understanding of complex narratives, not limited to public health but also applicable to various other domains. The initial success of AD-AutoGPT has paved the way for future LLM-assisted investigations in global health landscapes and beyond. By leveraging the power of large language models and automation techniques, researchers and professionals can gain valuable insights, inform evidence-based interventions, and drive positive impact across diverse domains.

综上所述，本研究提出了一种名为AD-AutoGPT的变革性自主大语言模型工具，能够促进对复杂叙事的数据驱动理解，其应用不仅限于公共卫生领域，还可扩展至其他多种领域。AD-AutoGPT的初步成功为未来在全球健康及其他领域开展大语言模型辅助研究开辟了道路。通过利用大语言模型和自动化技术的强大能力，研究人员与从业者能够获得宝贵洞见，为循证干预提供依据，并在多元领域产生积极影响。

7 Acknowledgments

7 致谢

This work received support from the National Institutes of Health (NIH) through RO1 grant R 01 MD 013886-05, partial support from grants R 01 AG 075582 and RF 1 NS 128534, as well as support from UGA Interdisciplinary Research Pre-Seed Program – “Interdisciplinary Approaches to Alzheimer’s Disease Prevention”. This work is also partially supported by the funding from National Institutes of Health, "Identification of Multi-modal Imaging Biomarkers for Early Prediction of MCI-AD Conversion via Multigraph Representation" (1 R 03 AG 078625- 01). We would like to express our gratitude to the funding agency for their financial support.

本研究获得了美国国立卫生研究院(NIH)通过RO1基金R01MD013886-05的资助，部分资金来自R01AG075582和RF1NS128534号基金，以及佐治亚大学跨学科研究预种子计划"阿尔茨海默病预防的跨学科方法"的支持。同时，美国国立卫生研究院"通过多图表示识别MCI-AD早期转化的多模态影像生物标志物"(1R03AG078625-01)项目也为本研究提供了部分资金支持。我们谨向资助机构表示衷心感谢。

References

参考文献

Dimitrios Avram o poul os. Genetics of alzheimer’s disease: recent advances. Genome medicine, 1(3):1–7, 2009.
Dimitrios Avramopoulos. 阿尔茨海默病遗传学: 最新进展. Genome medicine, 1(3):1–7, 2009.

[论文翻译]AD-AutoGPT: 用于阿尔茨海默病信息学的自主GPT

原文地址：https://arxiv.org/pdf/2306.10095