AI Chatbots in Clinical Laboratory Medicine: Foundations and Trends

He Sarina Yang, PhD, MBBS, DABCC (CC, TC)

Associate Professor of Clinical Pathology and Laboratory Medicine Weill Cornell Medical College, Cornell University

Director, Clinical Chemistry service and Toxicology and Therapeutic Drug Monitoring service

Artificial intelligence (AI) conversational agents, also known as chatbots or intelligent virtual assistants, are computer programs designed to simulate human conversation using natural language processing (NLP) to interact with users. In an article published in Clinical Chemistry, Yang et al. briefly explained the history and major development approaches of chatbots, reviewed the opportunities and limitations of using chatbots in healthcare, particularly in the field of Clinical Laboratory Medicine, and discussed future trends and research directions for improving their clinical utility1.

One of the main methods for developing chatbots is the end-to-end approach. This approach uses unified deep learning models, such as large language models (LLMs), which are trained to generate responses from scratch without the need for manually crafted features or domain-specific knowledge. The abilities of LLMs to understand words in context, accurately predict the next word, and mimic human speech patterns can be attributed to their access to extensive amounts of books, articles, Wikipedia, the wider internet, and even computer source code. LLMs encode text into feature embeddings, estimate the probability of words or phrases based on the preceding context within a given window size, and then decode this information to generate response text. Leveraging the transformer model architecture, LLMs demonstrate exceptional learning capabilities and can produce human-like responses to text inputs in conversational settings. However, the training methods of LLMs also determine their limitations, such as limited understanding and context retention, misinformation, bias, and security concerns. Concerns have been raised that chatbots may generate responses that appear convincing but contain inaccurate and fabricated information, often referred to as “hallucinations”.

AI chatbots have become an integral part of daily life, with companies employing them for tasks such as customer support and marketing. In healthcare, AI chatbots provide a wide range of functions and applications, including disseminating health information, assisting with self-triage and personal risk assessments, and offering remote patient support and monitoring. In the realm of Laboratory Medicine, chatbots hold great potential for improving the automation of responses to routine inquiries concerning laboratory tests, facilitating the interpretation of laboratory results, and providing “curbside consultation” to physicians. A study by Yang et al. revealed that Chat Generative Pre-Trained Transformer (ChatGPT)-4.0, when prompted with specific laboratory inquiries, demonstrated significant improvements in accuracy, consistency, and comprehensiveness compared to its predecessors, ChatGPT-3.5 and ChatGPT-3.0 2. Additionally, when paired with retrieval-augmented generation (RAG), LLM output can be optimized by referencing an authoritative knowledge base outside its training data source, greatly reducing the occurrence of AI hallucinations.

The application of AI in healthcare, specifically in laboratory medicine, holds tremendous potential for advancement and improvement. One significant advancement would be the development of hybrid models that combine LLMs with domain-specific modules extensively trained on subject matter. These models should be trained using reliable data sources, including specialized medical language and terminology. They should be rigorously evaluated and validated before being deployed to users. Additionally, the models should be trained to adhere to and ground their responses in human assumptions, such as consistency with previous facts and alignment with common sense. Moreover, models that could provide personalized medical assistance by integrating patients' medical history and laboratory test results, while complying with privacy regulations to protect patient privacy, would be highly valuable.

AI chatbots, such as ChatGPT, are rapidly advancing AI applications at a remarkable pace, with exciting potential to enhance communication among patients, clinicians, and laboratorians. Therefore, it is imperative for clinicians and laboratorians to understand the benefits and limitations of chatbots in order to use them effectively in their clinical practice.

临床检验医学中的人工智能聊天机器人：基础与趋势

人工智能会话工具，也称为聊天机器人或智能虚拟助手，是设计用来模拟人类对话并使用自然语言处理与用户互动的计算机程序。在临床化学杂志的这篇文章中，杨博士和她的团队简要介绍了聊天机器人的历史和主要建造方法，回顾了聊天机器人在医疗行业，尤其是在临床检验中的发展前景和局限性，并讨论了提高其临床实用性的趋势和研究方向1。

开发人工智能聊天机器人的主要方法之一是端到端的方法。该方法使用统一的深度学习模型，如大语言模型。这种模型的训练模式是以对问题提示的编码和分析而算出相应的回应。这些模型的功能没有经过特殊的调试，也没有在特定知识领域经行培训。大语言模型可以根据上下文中的单词而准确预测下一个单词，同时它也可以模仿人类的语言模式。这些都归功于它学习了大量书籍，文献，维基百科，互联网，甚至电脑源代码。大语言模型将文本根据其特征进行分解和编码，在指定文本窗口内根据上下文而评估单词或短语出现的概率从而解码计算出回应。利用变换器模型架构，大语言模型展现了出色的学习能力并且能够在对话环境中根据文本输入生成类似人类的回应。但是，大语言模型的训练方法也决定了其局限性。比如，取决于文本窗口的大小，它对充分理解上下文有一定的局限性。同时，因为训练材料的局限性，它有产生错误信息，偏见和泄露隐私的问题。有担忧认为，人工智能聊天机器人可能会生成看似可信实则错误的信息，通常称为“幻觉”。

人工智能聊天机器人已经融入了日常生活，特别是在客户服务和产品营销领域有着充分的利用。在医疗行业中，人工智能聊天机器人在传播健康信息和知识，协助自我隔离和个人风险评估，提供远程患者支持和监测方面有着广泛的应用。在临床检验中，人工智能聊天机器人在改善常规问题自动查询，辅助实验结果分析，为临床医生提供“路边咨询”有着巨大潜力。根据杨博士和她的团队的另外一篇发表在临床化学杂志的研究表明，当被要求回答特定的临床检验问题时，生成型预训练变换器（ChatGPT）-4.0 在准确性、一致性和全面性方面相比其前身 ChatGPT-3.5 和 ChatGPT-3.0 有着显著的提高2。此外，当与检索增强生成器配对时，大语言模型可以参考其训练数据源之外的权威知识库，优化输出，从而大大减少了人工智能幻觉的发生。

人工智能在医疗行业中，特别是在临床检验中的应用，具有巨大的进步和改进潜力。其中一个重要方向是对大型语言模型与领域特定模块相结合的混合模型的开发。这些模型应使用包括专业医学语言和术语在内的可靠数据源进行训练，并在用户使用之前进行严格评估和验证。此外，这些模型应遵循和模仿人类思维模式，例如利用常识以及对之前事实的认知而产生回应。此外，能够通过整合患者的病史和实验测试结果提供个性化医疗帮助，同时遵守隐私法规以保护患者隐私的模型将非常有价值。

人工智能聊天机器人，如ChatGPT，以令人瞩目的速度迅速推进AI应用，具有增强患者、临床医生和实验人员之间沟通的巨大潜力。因此，临床医生和实验人员有必要了解聊天机器人的优缺点，以便在临床实践中有效的使用它们。

Reference

1He S Yang, Fei Wang, Matthew B Greenblatt, Sharon X Huang, Yi Zhang, AI Chatbots in Clinical Laboratory Medicine: Foundations and Trends, Clinical Chemistry, Volume 69, Issue 11, November 2023, Pages 1238–1246, https://doi.org/10.1093/clinchem/hvad106

2Munoz-Zuluaga C, Zhao Z, Wang F, Greenblatt MB, Yang HS*. Assessing the accuracy and clinical utility of ChatGPT in laboratory medicine. Clinical Chemistry, volume 69, issue 8, 2023, page 930-940.