Cell Reports Medicine | Prospective RCT by LIN Haotian's Team: Large Language Models as the "Intelligent Engine" for Medical AI Research

Origin：LIN Haotian's Team; Department of Science and Technology

Editor：Liu Te

Published：2025-11-28

Pageview：

Font Size：

With the in-depth implementation of the State Council's Opinions on Further Implementing the "Artificial Intelligence +" Initiative and the comprehensive advancement of the "Healthy China" strategy, artificial intelligence (AI), as a key engine driving medical modernization, continues to empower innovations in medical research and upgrades in clinical services. Medical AI research has emerged as a crucial direction for interdisciplinary integration and the transformation of scientific research paradigms, holding strategic significance for improving disease diagnosis and treatment capabilities and optimizing the healthcare service system. However, amid the rapid development of medical AI research, many clinicians still face severe challenges: despite possessing frontline clinical experience and valuable professional insights, they often struggle to deeply participate in and lead medical AI research due to practical bottlenecks such as high technical barriers, limited engineering support, and insufficient interdisciplinary knowledge. This has restricted the breadth and vitality of medical AI innovation.

In response to the national strategic demand of promoting extensive and in-depth integration between AI and the healthcare industry and strengthening the cultivation of interdisciplinary talent, Professor LIN Haotian's team from Zhongshan Ophthalmic Center, Sun Yat-sen University, focused on the key question of "how to systematically empower clinicians to conduct medical AI research". They innovatively explored and developed a novel paradigm for medical AI research using large language models (LLMs) as the core auxiliary tool, and validated its effectiveness through a prospective, superiority-design, open-label randomized controlled trial (RCT). Recently, this research was published online in Cell Reports Medicine, a journal under Cell Press, titled "The effectiveness of large language models in medical artificial intelligence research for physicians: A randomized controlled trial".

1. Main Research Findings

01. From "Technical Gap" to "Capability Leap": LLMs Empower Physicians to Achieve Breakthroughs in AI Research

The research team recruited 64 young ophthalmologists with no prior AI research experience or programming background. Participants were randomly assigned to an intervention group using LLMs and a control group using traditional search tools, with the task of completing a medical AI research project within two weeks.

The results showed that LLMs can effectively assist young physicians in independently completing medical AI projects: the overall project completion rate in the intervention group reached 87.5%, significantly higher than the 25.0% in the control group. The proportion of participants who completed the project independently without any expert guidance was as high as 68.7% in the intervention group, compared to only 3.1% in the control group. Analysis of the research process revealed that the research plans developed by physicians in the intervention group were more feasible and the time required to complete projects was significantly shorter.

To assess the sustainability of the acquired capabilities, the team conducted a supplementary trial after a washout period, finding that 41.2% of successful participants were able to independently complete a new medical AI project after the removal of LLM support. This demonstrates that the AI research capabilities gained through LLM assistance have good transferability, allowing physicians to apply the acquired knowledge and skills to new research scenarios.

02. Coexistence of "Efficacy" and "Risks": Rational Use Awareness Needed

The study also revealed potential risks associated with LLM use. In the supplementary experiment, over 40% of successful participants were unable to independently complete a new project after LLM support was withdrawn, indicating a potential dependency tendency among participants on LLMs. Further questionnaire surveys showed that 42.6% of participants worried that using LLMs might lead to "merely mechanical replication without understanding," and 40.4% expressed concern that it might "foster lazy thinking," highlighting potential risks of over-reliance on LLMs. These findings suggest that while promoting LLM-assisted medical research, it is crucial to emphasize the comprehensive evaluation and mitigation of potential long-term risks and establish scientific usage guidelines.

03. From "Can Use" to "Know How to Use" and "Use Well": Developing the "CPGI" Prompt Guideline

Addressing issues observed in the study, such as AI hallucinations and inefficient questioning strategies, the research team summarized a set of prompt construction guidelines named "CPGI." This framework is designed to guide physicians in using LLMs more safely and efficiently, providing them with a logical and structured operational framework.

Currently, LLM-driven AI agent technology is driving profound changes in medical research paradigms – from autonomously designing experimental protocols to accelerate new drug discovery, to decoding vast genomic data, to reveal new disease mechanisms... Future "AI for Medical Science" may transcend the realm of tool assistance, moving towards a new paradigm of medical research driven by intelligent agents, characterized by autonomy and collaboration. To realize this vision, key challenges such as reliability, interpretability, and ethical alignment still need to be overcome. The team's research indicates that the starting point for advancing this process lies in cultivating interdisciplinary talent – nurturing a new generation of medical scientists who understand both clinical practice and AI application. They will become the core force leading this transformation.

As Sun Yat-sen University embarks on its new century journey and Zhongshan Ophthalmic Center enters its new 60-year cycle, this research represents a significant embodiment of the Center's continuous efforts to promote medical-engineering integration and practice the national strategy of "serving people's life and health." The project has vigorously advanced the development of a composite young talent pool integrating clinical practice and artificial intelligence, contributing "ZOC strength " to serving the nation's major strategic needs in the field of medical AI.

2. Introduction to the Research Team

Professor LIN Haotian's team at Zhongshan Ophthalmic Center, Sun Yat-sen University, collaborating with domestic and international universities and research institutions, leverages interdisciplinary strengths to establish a system for ophthalmic AI diagnosis, treatment, and clinical application: addressing key issues of data governance and security protection to lay the foundation for innovations in intelligent diagnosis and treatment technologies; driving breakthroughs in intelligent screening and diagnosis of ocular and systemic diseases through algorithm innovations based on dynamic and static ocular features; constructing a novel intelligent "three-tier diagnosis and treatment" model for eye diseases, enabling multi-scenario application of eye disease diagnosis and treatment.

The team leads major projects including Key Supported Projects of the National Natural Science Foundation of China's Major Research Plan and Intelligent Diagnosis and Treatment Projects for Major Diseases. They have published over 200 papers in top international journals including Nature, Nature Medicine, Science, The Lancet, The BMJ, Nature Biomedical Engineering, and The Lancet Digital Health.

Their technological achievements have been widely applied in representative medical institutions at all levels across China and in countries and regions along the "Belt and Road", benefiting millions of residents and patients. The team leads the innovation and development of key intelligent technologies for eye disease prevention and treatment, improving the overall level of ophthalmic healthcare.

Co-Corresponding Authors