Massive Online Open Validation and Evaluation of Large Language Models for Resource Constrained Settings - a Study of Contextual Awareness, Safety, and Accuracy (SilentMOOVE)
Healthcare settings in low and middle income countries, clinicians are frequently required to manage a wide range of acute conditions under the constraints of time pressure, limited diagnostic resources, and variable availability of up-to-date clinical guidelines. Large language models (LLMs), which can synthesize and generate medical information in a natural language, have been proposed as potential clinical decision support tools. Most evaluations of LLMs in medicine have relied on standardized benchmark examinations or hypothetical case vignettes, often derived from high-income country contexts. MOOVE (Massive online open validation and evaluation) framework was developed to address the need for rigorous and context-sensitive evaluation of LLMs through a staged approach.
Study is a non-interventional, prospective study designed to test LLMs safely in real clinical settings, evaluate their performance against human care, and explore how feasible their future implementation in healthcare might be.
研究概览
详细说明
SilentMOOVE (Silent massive online open validation and evaluation) phase of the MOOVE project, which represents the intermediate step between previous hypothetical testing study and future interventional trial. In silentMOOVE, patients are prospectively recruited during routine clinical care in participating outpatient health facilities. Clinical and demographic data collected during consultations are used to prompt LLMs retrospectively without influencing patient care. The routine diagnosis and care plan of the treating clinicians (human care plan) are also captured. LLM outputs are subsequently evaluated by a panel of independent clinical experts against the human care plan as a comparator. Patient clinical outcomes are collected on 7 and 28 days after enrolment to establish a baseline and to be used in the evaluation of LLMs.
The rationale for this study is to generate robust, prospective evidence on the safety and performance of LLMs in real-world clinical contexts, while minimizing risk to participants by fully separating model evaluation from clinical decision-making.
Study also provides the opportunity to assess the usability and gather early insights on implementation readiness of the LLM prompting interface. For this reason, besides evaluating the performance and safety of various LLMs, this study includes an implementation feasibility component. These activities are needed to better understand the context and any amelioration needed before future implementation in healthcare settings. Learnings from the implementation feasibility research will be shared among countries, especially applicable within a given use case. Implementation feasibility research will be grounded in artificial intelligence in healthcare implementation frameworks and other Implementation science frameworks Primary objectives
- To evaluate the overall performance profile of LLMs on real-world clinical scenarios using the human care plan as a comparator
- To assess whether each LLM being evaluated is not inferior to the human comparator with respect to safety on real-world clinical scenarios, based on structured retrospective assessment of model- and human-generated care plans.
Implementation feasibility research objectives
- Assess clinician experience with the LLM prompting interface: ease of use, time burden, and alignment with clinical workflow.
- Gather clinician impressions of LLM output disclosed and reviewed post-consultation: clarity, relevance, perceived safety, alignment with local clinical guidelines, and trust in LLM outputs.
- Facilitators and Barriers to Future Implementation: Identify site-level infrastructure, staff capacity, and organizational readiness factors that may support or hinder future implementation of LLM feedback in real-time
研究类型
注册 (估计的)
联系人和位置
学习联系方式
- 姓名：Mary-Anne Hartley
- 电话号码：+41216930828
- 邮箱：mary-anne.hartley@epfl.ch
研究联系人备份
- 姓名：Kristina Keitel
- 电话号码：+41216932039
- 邮箱：Kristina.keitel@epfl.ch
学习地点
-
-
Dar es Salaam Region
-
Dar es Salaam、Dar es Salaam Region、坦桑尼亚
- 尚未招聘
- Muhimbili University of Health and Allied Sciences (MUHAS)
-
接触：
- Hendry R Sawe
- 电话号码：+255754885658
- 邮箱：hendry_sawe@yahoo.com
-
首席研究员：
- Hendry R Sawe
-
-
Dodoma
-
Dodoma、Dodoma、坦桑尼亚
- 招聘中
- Ifakara Health Institute (IHI)
-
接触：
- Grace Mhalu
- 电话号码：+255655888891
- 邮箱：gmhalu@ihi.or.tz
-
首席研究员：
- Grace Mhalu
-
-
-
-
Nairobi County
-
Nairobi、Nairobi County、肯尼亚、19676
- 尚未招聘
- University of Nairobi College of health sciences
-
接触：
- Paul Macharia
- 邮箱：pmachariah@gmail.com
-
首席研究员：
- Paul Macharia
-
-
-
-
Blantyre
-
Blantyre、Blantyre、马拉维、303
- 招聘中
- Malawi University of Business and Applied Sciences (MUBAS)
-
接触：
- Chiyembekezo Kachimanga
- 电话号码：+265 887247777
- 邮箱：chembekachimanga@yahoo.co.uk
-
首席研究员：
- Chiyembekezo Kachimanga
-
-
参与标准
资格标准
适合学习的年龄
接受健康志愿者
有资格学习的性别
取样方法
研究人群
The study population consists of patients of any age presenting to participating health facilities for a distinct care-seeking episode related to a clinical condition within the defined use case (e.g. primary care, emergency care).
Both adult and pediatric patients are eligible, reflecting the real-world case mix of the participating facilities.
描述
Inclusion Criteria:
- Presentation at a participating facility for a distinct care seeking episode related to a clinical condition.
- Use case specific inclusion criteria
- Willing and able to provide written informed consent.
Exclusion Criteria:
- Prolonged care-seeking episode.
- Previously enrolled in the same study.
- Use case specific exclusion criteria
- Immediate life-threatening condition requiring emergency intervention precluding study procedures.
学习计划
研究是如何设计的？
设计细节
- 观测模型：COHORT
- 时间观点：PROSPECTIVE
队列和干预
团体/队列 |
干预/治疗 |
|---|---|
|
Non interventional study
Non interventional study |
Non interventional study |
研究衡量的是什么？
主要结果指标
结果测量 |
措施说明 |
大体时间 |
|---|---|---|
|
Performance Scores
大体时间：Day 0
|
Scoring is between -2 to +2 where '-2' means minimum, and '-2' means maximum. Score to be >0. Higher score means better outcome. |
Day 0
|
|
Safety performance threshold achievement
大体时间：'Day 7' and 'Day 28'
|
The mean safety score (minimum is -2 and maximum is +2) is formally compared between each LLM and the human comparator to assess non-inferiority within a UCA (use case appendix) dependent pre-specified margin. |
'Day 7' and 'Day 28'
|
次要结果测量
结果测量 |
措施说明 |
大体时间 |
|---|---|---|
|
Implementation Feasibility Research
大体时间：week 4 until study completion
|
Usability: Quantitative scoring rubric for clinicians; semi-structured interviews with clinicians; evaluation methods will be used |
week 4 until study completion
|
|
Implementation Feasibility Research
大体时间：week 4 until study completion
|
Workflow integration feasibility: This will be evaluated using semi structured interviews and questionnaires |
week 4 until study completion
|
|
Implementation Feasibility Research
大体时间：week 4 until study completion
|
Acceptability and perceived trust: This will be evaluated with Quantitative scoring rubric for clinicians; semi-structured interviews with clinicians |
week 4 until study completion
|
|
Implementation Feasibility Research
大体时间：week 4 until study completion
|
Implementation facilitators and barriers: This will be evaluated using semi-structured interviews and questionnaires |
week 4 until study completion
|
合作者和调查者
出版物和有用的链接
一般刊物
- 1. Arora RK, Wei J, Hicks RS, Bowman P, et al. HealthBench: Evaluating Large Language Models Towards Improved Human Health. Preprint [Under Review]. 2025. 2. Burnell R, Schellaert W, Burden J, Ullman TD, Martinez-Plumed F, Tenenbaum JB, et al. Rethink reporting of evaluation results in AI. Science. 2023;380: 136-138. doi:10.1126/SCIENCE.ADF6369. 3. Hager P, Jungmann F, Holland R, Bhagat K, Hubrecht I, Knauer M, et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nature Medicine 2024 30:9. 2024;30: 2613-2622. doi:10.1038/s41591-024-03097-1. 4. Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., et al. (2023). Large language models encode clinical knowledge. Nature, 620(7972), 172-180. https://doi.org/10.1038/s41586-023-06291-2 5. WHO. Towards the Development of guidance on ethics and governance of artificial intelligence for health. 2019. 6. Azad TD, Krumholz HM, Saria S. Principles to guide clinical AI readiness and move from benchmarks to real-world evaluation. Nature Medicine. 2026;32:802-804. doi:10.1038/s41591-025-04198-1 7. Declaration of Helsinki https://www.wma.net/policies-post/wma-declaration-of-helsinki-ethical-principles-for-medical-research-involving-human-subjects 8. European Parliament and Council of the European Union, 2016. Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal of the European Union, L119, pp.1-88. 9. International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH), 2016. Integrated Addendum to ICH E6(R1): Guideline for Good Clinical Practice E6(R2). Step 4, 9 November 2016. 10. U.S. Food and Drug Administration. 21 CFR Part 11-Electronic Records; Electronic Signatures. Code of Federal Regulations. Title 21, Part 11. Revised 2024.
研究记录日期
研究主要日期
学习开始 (实际的)
初级完成 (估计的)
研究完成 (估计的)
研究注册日期
首次提交
首先提交符合 QC 标准的
首次发布 (实际的)
研究记录更新
最后更新发布 (实际的)
上次提交的符合 QC 标准的更新
最后验证
更多信息
与本研究相关的术语
关键字
其他研究编号
- LiGHT-26-02
- INV-076674 (其他赠款/资助编号：BILL and Melinda gates foundation)
计划个人参与者数据 (IPD)
计划共享个人参与者数据 (IPD)？
IPD 计划说明
We will share IPD within 12 months of study completion.
IPD 共享时间框架
IPD information will be shared within 12 months of study completion.
IPD 共享访问标准
Will be confirmed within 12 months of study completion
IPD 共享支持信息类型
- 研究方案
药物和器械信息、研究文件
研究美国 FDA 监管的药品
研究美国 FDA 监管的设备产品
此信息直接从 clinicaltrials.gov 网站检索，没有任何更改。如果您有任何更改、删除或更新研究详细信息的请求，请联系 register@clinicaltrials.gov. clinicaltrials.gov 上实施更改，我们的网站上也会自动更新.