TL;DR
This scoping review examines the current research on using Large Language Models (LLMs) like ChatGPT in telehealth. It highlights that while LLMs show great promise for improving healthcare delivery, there's a significant gap in research concerning their responsible and ethical implementation, particularly regarding transparency, human agency, and accountability. The paper proposes future research directions to ensure trustworthy AI in telehealth.
What Problem Does This Paper Address?
The integration of Artificial Intelligence (AI), especially Large Language Models (LLMs) and generative AI tools like ChatGPT, into healthcare holds immense potential to revolutionize digital health. Telehealth, the remote delivery of healthcare services, is a prime area where these technologies could enhance efficiency, accessibility, and patient experience. LLMs can act as sophisticated assistants for clinicians, improve e-consultations, and bridge communication gaps. However, the rapid adoption of these powerful tools outpaces our understanding of their responsible and ethical use in such a sensitive domain. Concerns about AI regulations, potential inaccuracies, privacy, and ethical implications create a critical need to assess the existing research landscape and identify knowledge gaps to guide future development and implementation.
Methods
The authors conducted a scoping review, a systematic approach to map the existing research on a topic. They followed the established framework by Arksey and O'Malley, which involves identifying research questions, searching for relevant studies, selecting those that meet inclusion criteria, charting the data, and summarizing the results. The research question focused on the current state of research regarding the responsible use of LLMs in telehealth and identified knowledge gaps. Their search strategy involved keywords related to telehealth, generative AI, ChatGPT, and LLMs across multiple databases. Inclusion criteria focused on English-language publications, encompassing empirical and non-empirical research, reviews, and perspectives. They specifically used the European Union's (EU) and Australia's AI ethics guidelines as a conceptual matrix to analyze the literature through the lens of responsible AI.
Datasets
Not discussed in the paper. This paper is a review of existing literature and does not involve the use of specific datasets for training or evaluation.
Key Insights and Ideas
- Nascent Research Landscape: The review found that research on LLMs in telehealth is very new, with all 20 included studies published in 2023. This indicates a rapidly evolving but still underdeveloped field.
- Promise vs. Practice: While the potential benefits of LLMs in telehealth (e.g., improved patient care, enhanced efficiency) are frequently discussed, the research often lacks empirical evidence or deep conceptualization of these benefits.
- Ethical Gaps are Significant: Key ethical considerations like transparency, explainability, human agency, and contestability are largely absent from the current research. This is a critical oversight given the sensitive nature of healthcare.
- Focus on "What" not "How": Much of the existing literature discusses what LLMs can do in telehealth, but there's a lack of research on how to implement them responsibly and effectively.
- Telehealth as a Crucial Testbed: The unique context of telehealth provides a compelling environment to explore and operationalize responsible AI practices, especially concerning fairness and mitigating digital divides.
- Environmental Impact is Overlooked: The potential environmental benefits of telehealth, amplified by LLMs, are a promising but under-researched area.
Key Results
- 20 Studies Included: The review identified 20 relevant studies, with 5 being empirical (qualitative or quantitative) and 15 non-empirical (reviews, perspectives, letters).
- Dominance of "Beneficial Outcomes": "Beneficial outcomes" was the most frequently discussed concept (15 papers), though often peripherally.
- "Reliability and Safety" Second Most Explored: Nine papers discussed reliability and safety, with four providing empirical evidence.
- Peripheral Discussion of Key Ethical Concepts: Concepts like fairness, inclusiveness, privacy, security, and accountability were discussed in a limited number of studies, often peripherally.
- Complete Absence of Empirical/Conceptual Exploration: Transparency, explainability, human agency, and contestability were not empirically or conceptually explored in any of the included articles.
Limitations and Future Work
The authors acknowledge that they omitted the "expert consultation" step due to the emerging nature of the topic, suggesting that future research could benefit from expert input, especially on sociotechnical structural assurance.
They propose three key research directions for responsible LLM use in telehealth:
- Beneficial Outcomes: Focus on personalized care, adapting to individual patient needs, understanding the impact on telehealth providers' skills and decision-making, and exploring effective use theories to maximize benefits for all stakeholders. This also includes investigating the environmental impact of LLMs in telehealth.
- Human-Centered Values and Rights: Research should delve into human agency, ensuring LLMs empower rather than diminish autonomy. Transparency and explainability need to be explored in practical telehealth contexts, along with developing methods to mitigate bias and promote fairness. Privacy protection, especially concerning sensitive health data and regulations like HIPAA, requires in-depth investigation. Contestability, the right to challenge AI decisions, also needs institutional mechanisms for resolution.
- Sociotechnical Structural Assurance: This involves ensuring reliability, safety, security, and human oversight in LLM-enabled telehealth. Research should focus on developing robust security measures, effective human-in-the-loop approaches, and clear lines of accountability, considering the complex interplay between providers, developers, and regulators.
The paper emphasizes that future research should move beyond theoretical discussions to empirical investigations that detail the direct impact of LLMs on clinical practice, workflows, patient outcomes, and clinician well-being, ultimately ensuring ethical practices, safety, and patient privacy.
Summary updated