Large Language Model Assistance for Clinical Decision-Making Among Rural Physicians

Effect of Large Language Model Assistance on Clinical Decision-Making Among Rural Physicians: A Randomized Controlled Trial

This study will evaluate whether, relative to conventional information retrieval approaches, direct large language models (LLM) access and LLM use training can improve the overall clinical decision-making ability of rural physicians in low-resource grassroots healthcare settings.

Study Overview

• Status: Not yet recruiting
• Conditions: Clinical Decision-making
• Intervention / Treatment: Other: Conventional Non-LLM Resources; Other: LLM Second-Opinion Review; Other: Direct LLM Assistance; Behavioral: LLM-Use Training

Detailed Description

Rural physicians play an essential role in the diagnosis and management of common and frequently occurring conditions, referral decision-making, chronic disease management, and patient education. In resource-constrained primary care settings, they often face limited access to medical information and specialist support, delays in updating clinical knowledge and guidelines, and substantial pressure in clinical decision-making. These challenges are particularly relevant in northwestern China, where primary care resources are relatively limited. Improving rural physicians' abilities in diagnostic assessment, recognition of clinical warning signs, and rational prescribing is therefore an important priority for strengthening primary healthcare services.

Large language models (LLMs) can support medical information retrieval, organization of diagnostic and management approaches, differential diagnosis, medication-related decision-making, patient education, and follow-up planning, and may therefore serve as accessible tools for supporting clinical decision-making in primary care. However, general-purpose LLMs were not specifically developed for use in resource-constrained primary care settings and have not been adequately evaluated among rural physicians. Their responses may contain factual errors or fabricated evidence, overlook warning signs, provide insufficient medication safety warnings, or recommend investigations and treatments that are not feasible in local primary care settings. Without adequate verification skills, physicians may fail to benefit from LLM assistance and may even introduce new safety risks. It is therefore important to evaluate how rural physicians use LLMs and whether structured training can improve the safe and effective use of these tools before their wider implementation.

This randomized controlled trial will evaluate the effects of LLM assistance and brief training on clinical decision-making among rural physicians. Participants will complete clinical cases involving common conditions encountered in primary care, with tasks assessing diagnostic judgment, recognition of warning signs, rational treatment, and patient education. Some participants will also use the LLM as a second-opinion tool to review and revise their initial decisions. All responses will be independently evaluated by reviewers blinded to group assignment using standardized scoring criteria to assess overall clinical decision-making performance and safety.

• Study Type: Interventional
• Enrollment (Estimated): 240
• Phase: Not Applicable

Contacts and Locations

• Study Contact: Name: Liyuan Tao, MD; Phone Number: +86 82265732; Email: tendytly@163.com

Study Locations

• China
  • Xinjiang Uygur Autonomous Region
    • Karamay, Xinjiang Uygur Autonomous Region, China
      • Xinjiang Second Medical College
      • Contact: Zejiang Ma, Master of Medicine; Phone Number: +86 18811331959; Email: 18811331959@163.com

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: Yes

Description

Inclusion Criteria:

• Currently engaged in clinical practice at a rural primary healthcare institution in northwestern China.
• Has received formal medical education and holds a relevant diploma or degree.
• Able to read and understand clinical case materials in Chinese.
• Able to use a computer to complete the study tasks.
• Willing to participate and able to provide written informed consent.

Exclusion Criteria:

• Previously involved in the development of the clinical case tasks, reference answers, or scoring rubric for this study.
• Previously participated in pilot testing involving the same clinical case tasks or study procedures.

Study Plan

How is the study designed?

Design Details

• Primary Purpose: Health Services Research
• Allocation: Randomized
• Interventional Model: Factorial Assignment
• Masking: Single

Number of Arms

4

Arms and Interventions

Participant Group / Arm	Intervention / Treatment
Active Comparator: Conventional Resources Followed by LLM Second Opinion; Participants receive no LLM-use training. During the initial 60-minute assessment, they complete the clinical cases using conventional non-LLM resources only. After submitting their initial responses, they receive an additional 30 minutes to use the study-provided LLM as a second-opinion tool to review and revise their responses.	Other: Conventional Non-LLM Resources; During the initial 60-minute assessment, participants complete primary care clinical cases using conventional non-LLM resources only, including clinical guidelines, textbooks, drug labels, training materials, medical websites, and standard search engines. Participants are not permitted to use LLMs during this phase.; Other: LLM Second-Opinion Review; After completing and submitting their initial responses using conventional non-LLM resources, participants receive an additional 30 minutes to use the study-provided DeepSeek-V4 as a second-opinion tool. They may review, verify, and revise their initial clinical decisions before submitting their final responses.
Experimental: Direct LLM Assistance; Participants receive no LLM-use training and may use the study-provided LLM throughout the initial 60-minute assessment to complete the clinical cases. They submit their final responses at 60 minutes and do not participate in the additional second-opinion phase.	Other: Direct LLM Assistance; During the initial 60-minute assessment, participants may use the study-provided DeepSeek-V4 to assist with medical information retrieval, diagnostic and management reasoning, identification of warning signs, referral decisions, rational prescribing, patient education, and follow-up planning. Participants remain responsible for their final clinical decisions and responses.
Experimental: LLM Training and Conventional Resources Followed by LLM Second Opinion; Participants receive brief structured training on the safe and effective use of LLMs. During the initial 60-minute assessment, they complete the clinical cases using conventional non-LLM resources only. After submitting their initial responses, they receive an additional 30 minutes to use the study-provided LLM as a second-opinion tool to review and revise their responses.	Other: Conventional Non-LLM Resources; During the initial 60-minute assessment, participants complete primary care clinical cases using conventional non-LLM resources only, including clinical guidelines, textbooks, drug labels, training materials, medical websites, and standard search engines. Participants are not permitted to use LLMs during this phase.; Other: LLM Second-Opinion Review; After completing and submitting their initial responses using conventional non-LLM resources, participants receive an additional 30 minutes to use the study-provided DeepSeek-V4 as a second-opinion tool. They may review, verify, and revise their initial clinical decisions before submitting their final responses.; Behavioral: LLM-Use Training; Before completing the clinical cases, participants receive brief structured training on the safe and effective use of LLMs. The training covers the role and limitations of LLMs, structured prompting and follow-up questioning, identification of warning signs and referral indications, medication safety, verification of LLM-generated information, high-risk situations in which LLMs should not be relied upon, and protection of patient privacy.
Experimental: LLM Training and Direct LLM Assistance; Participants receive brief structured training on the safe and effective use of LLMs and may use the study-provided LLM throughout the initial 60-minute assessment to complete the clinical cases. They submit their final responses at 60 minutes and do not participate in the additional second-opinion phase.	Other: Direct LLM Assistance; During the initial 60-minute assessment, participants may use the study-provided DeepSeek-V4 to assist with medical information retrieval, diagnostic and management reasoning, identification of warning signs, referral decisions, rational prescribing, patient education, and follow-up planning. Participants remain responsible for their final clinical decisions and responses.; Behavioral: LLM-Use Training; Before completing the clinical cases, participants receive brief structured training on the safe and effective use of LLMs. The training covers the role and limitations of LLMs, structured prompting and follow-up questioning, identification of warning signs and referral indications, medication safety, verification of LLM-generated information, high-risk situations in which LLMs should not be relied upon, and protection of patient privacy.

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Overall Clinical Decision-Making Score	Participants' responses to primary care clinical cases will be evaluated using a prespecified scoring rubric. The overall score will reflect performance across key components of clinical decision-making. Higher scores indicate better overall clinical decision-making performance.	At the end of the initial 60-minute assessment

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Diagnostic Judgment Domain Score	Participants' diagnostic judgment in response to primary care clinical cases will be evaluated using a predefined scoring rubric. Higher scores indicate better diagnostic judgment.	At the end of the initial 60-minute assessment
Clinical Warning Sign Recognition Domain Score	Participants' ability to identify clinically important warning signs in primary care clinical cases will be evaluated using a predefined scoring rubric. Higher scores indicate better recognition of clinical warning signs.	At the end of the initial 60-minute assessment
Treatment Plan Domain Score	Participants' proposed treatment plans for primary care clinical cases will be evaluated using a predefined scoring rubric. Higher scores indicate better treatment planning performance.	At the end of the initial 60-minute assessment
Change in Overall Clinical Decision-Making Score After LLM Review	Among participants assigned to the LLM second-opinion phase, the change in overall clinical decision-making score will be calculated as the score after LLM-assisted review minus the score before LLM-assisted review. Positive values indicate improvement in overall clinical decision-making performance.	Change from 60 to 90 minutes after the start of the assessment

Collaborators and Investigators

• Sponsor: Peking University Third Hospital
• Collaborators: Xinjiang Second Medical College

Study record dates

Study Major Dates

• Study Start (Estimated): July 1, 2026
• Primary Completion (Estimated): August 1, 2026
• Study Completion (Estimated): September 1, 2026

Study Registration Dates

• First Submitted: July 14, 2026
• First Submitted That Met QC Criteria: July 14, 2026
• First Posted (Actual): July 17, 2026

Study Record Updates

• Last Update Posted (Actual): July 17, 2026
• Last Update Submitted That Met QC Criteria: July 14, 2026
• Last Verified: July 1, 2026

More Information

Terms related to this study

• Keywords: Clinical Decision Support; Large Language Models; Rural Physicians
• Other Study ID Numbers: IRB00006761-M20260614

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No