Artificial Intelligence (AI), Clinical Decision Support Systems, Diagnostic Reasoning
Conditions
Keywords
AI agent, Large Language Model, Clinical Decision Support, Admission Diagnosis, Diagnostic Accuracy, Diagnostic Reasoning, Randomized Controlled Trial, Management Planning
Brief summary
The goal of this clinical trial is to evaluate whether AI-assisted workflows improve physicians' admission diagnosis and management planning performance on standardized simulated inpatient cases, among practicing internal medicine and surgery physicians across all seniority levels and across three tiers of the Chinese healthcare system. The main questions it aims to answer are: * Does the Agent-assisted workflow yield better structured admission diagnosis and management planning scores than standalone LLM assistance? * Does the Agent-assisted workflow outperform the traditional workflow without AI tools? Researchers will compare three parallel groups (traditional workflow group, LLM-assisted group, Agent-assisted group) to determine whether the Agent tool can improve diagnostic accuracy and efficiency. Participants will: * Be recruited from 15 hospitals in China and participate remotely under video proctoring * Be randomly assigned to one of the three fixed workflows, with randomization stratified by hospital tier, specialty and seniority * Complete 6 anonymized simulated HIS admission cases within one hour * Submit structured answers for each case covering principal diagnosis, secondary diagnoses, differential diagnoses, diagnostic justification, next diagnostic or therapeutic steps, consultation and referral decisions, and diagnostic confidence * Have their operation logs and time consumption recorded automatically by the study platform
Interventions
OTHERAgent-assisted workflow
Conventional resources (pre-admission clinical record, search engines) plus an in-system Agent entry that automatically reads the full record and report images, produces a structured summary with source-text tracing, and supports multi-turn Q\&A and one-click editable drafts.
OTHERLLM-assisted workflow
Conventional resources plus an in-system multi-turn AI dialogue entry. The AI does not automatically read the record; participants paste text or send partial screenshots.
OTHERTraditional Workflow
Conventional resources only: the pre-admission clinical record, standard search engines. No AI assistance.
Sponsors
Second Affiliated Hospital, Zhejiang University, School of Medicine
Shangrao People's Hospital
Study design
Allocation
RANDOMIZED
Intervention model
PARALLEL
Primary purpose
HEALTH_SERVICES_RESEARCH
Masking
SINGLE (Outcomes Assessor)
Masking description
Participants and investigators cannot be masked, as the intervention is the workflow itself. Outcome assessors scoring the structured answers are blinded to group allocation; submitted answers are anonymized and stripped of workflow-identifying information before scoring.
Intervention model description
Three-arm parallel-group design with 1:1:1 allocation. All arms share the same conventional resources (pre-admission clinical record, search engines); the arms differ only in the AI entry added on top (none / LLM / Agent). Randomization is stratified by hospital tier, specialty and seniority.
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
Yes
Inclusion criteria
* Hold a Medical Practitioner Qualification Certificate and/or Medical License, or be a recognized standardized resident physician; able to independently read electronic medical records, laboratory and imaging reports on an HIS. * Currently engaged in clinical work in internal medicine or surgery at one of the 15 participating hospitals. * Able to complete the case assessment in one continuous hour without breaks. * Able to participate remotely under video proctoring, with a stable internet connection and a working camera. * Voluntarily agree to participate and sign the informed consent form, including the declaration not to use unauthorized AI tools during the assessment. * Have not participated in case drafting, review, rubric development, or any activity that may leak the reference standard.
Exclusion criteria
* Have previously accessed the official test cases or reference standard of this study. * Unable to complete the training module, qualification test, or all experimental tasks. * Have conflicts of interest, e.g. participation in developing core algorithms of the tested system. * Unwilling to comply with remote proctoring, including keeping the camera on throughout. * Judged unsuitable by the investigators.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Mean Normalized Structured Score | Within one-hour study | Mean of the rescaled case scores (each case rescaled to 100), divided by the number of cases completed; range 0 to 100. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Active Response Time per Case | Within one-hour study | Active response time in seconds for each case, recorded automatically by the platform.Time is counted per case while that case's response page is active. |
| Degree of Adherence to AI-Generated Recommendations | Within one-hour study | Degree to which the submitted answer incorporates AI output. Assessed in the Agent and LLM arms only. |
Countries
China
Contacts
CONTACTYixin Zhang
STUDY_CHAIRYuan Ding
Second Affiliated Hospital, Zhejiang University, School of Medicine
Outcome results
None listed