Development and Prospective Validation of an AI-Based Diagnostic Model for Hepato-Pancreato-Biliary Diseases

Overview

The rapid advancement of artificial intelligence (AI) has expanded its applications in healthcare, particularly in diagnostic assistance, intelligent triage, and patient interaction. Hepatobiliary and pancreatic diseases (such as liver cancer, pancreatic cancer, cirrhosis) are characterized by insidious onset, rapid progression, low early-diagnosis rates, and poor prognosis. However, grassroots medical institutions in China face challenges including physician shortages, variable patient health literacy, and incomplete initial information collection, leading to high misdiagnosis/missed diagnosis risks. Recent breakthroughs in large language models (LLMs) and multi-agent systems (MAS) offer new solutions. LLMs enable advanced natural language processing, while MAS coordinates specialized agents for complex decision-making. Integrating MAS with medical LLMs could create intelligent pre-consultation systems that systematically collect patient symptoms, risk factors, family history, and lifestyle data to enhance diagnostic efficiency. This study aims to develop a MAS-based pre-consultation system for hepatobiliary-pancreatic diseases featuring four specialized agents ("guidance agent," "medical history agent," "risk assessment agent," and "summary generation agent"). The system will simulate clinical reasoning to generate structured diagnostic reports for physicians. Research Objectives: Develop a specialized multi-agent framework combining LLMs to simulate clinical diagnostic logic and standardize symptom collection Enhance pre-consultation data integrity through intelligent dialogue focusing on key disease indicators Generate structured diagnostic summaries highlighting critical symptoms and risk factors Establish foundation for clinical validation and application through expert evaluation and user feedback This pre-diagnostic tool will assist physicians rather than replace clinical judgment, promoting safe, effective AI applications in early disease screening and tiered healthcare systems.

Detailed description

With the rapid development of artificial intelligence (AI) technologies, their applications in the healthcare field have become increasingly widespread, particularly demonstrating significant potential in diagnostic assistance, intelligent triage, and patient interaction. Hepatobiliary and pancreatic diseases (e.g., liver cancer, pancreatic cancer, bile duct cancer, cirrhosis) are characterized by insidious onset, rapid progression, low early diagnostic rates, and poor prognosis. Clinical diagnosis and treatment demand high requirements for early recognition and precise evaluation. However, grassroots medical institutions in China commonly face challenges such as insufficient specialist physicians, heterogeneous patient health literacy, and inadequate initial information collection during first visits, leading to a high risk of missed or incorrect diagnoses and delays in optimal intervention timing.

In recent years, large language models (LLMs) have achieved breakthrough advancements in natural language understanding and generation, providing a technical foundation for constructing intelligent, personalized medical dialogue systems. Multi-agent systems (MAS) simulate collaborative workflows among multiple specialized roles, enabling more complex and structured decision-making processes with notable advantages in task decomposition, knowledge integration, and dynamic reasoning. Integrating multi-agent architectures with medical LLMs offers the potential to develop intelligent pre-consultation systems with domain-specific expertise, enabling systematic collection and preliminary analysis of critical patient information-including symptoms, risk factors, family history, and lifestyle-to enhance the efficiency and completeness of medical consultations.

This study aims to develop an intelligent pre-consultation LLM system for hepatobiliary and pancreatic diseases based on a multi-agent architecture. By establishing clearly defined "guidance agents," "medical history collection agents," "risk assessment agents," and "physician summary generation agents," the system will simulate clinical diagnostic logic, proactively guide patients through structured disease presentations, and automatically generate initial diagnostic reports compliant with clinical standards for physician reference. This system seeks to assist grassroots physicians in improving early recognition of hepatobiliary and pancreatic diseases, optimize outpatient resource allocation, and enhance patient healthcare experiences.

Research Objectives:

This study aims to design, develop, and preliminarily validate an intelligent pre-consultation LLM system for hepatobiliary and pancreatic diseases based on a multi-agent architecture, exploring its feasibility and potential value in clinical pre-consultation scenarios. Specific objectives include:

Establishing a specialized multi-agent collaborative framework: Leveraging LLM technology, design a multi-agent system capable of division of labor and collaboration, incorporating modules such as guidance, medical history collection, preliminary risk screening, and report generation. By simulating the diagnostic logic and reasoning pathways of clinical physicians, the system will achieve systematic, structured information collection for hepatobiliary and pancreatic disease-related symptoms.

Enhancing the completeness and accuracy of pre-consultation information: Through intelligent dialogue, proactively guide patients to review and articulate their medical history, prioritizing key risk factors for hepatobiliary and pancreatic diseases-including jaundice, abdominal pain, weight loss, alcohol consumption history, history of viral hepatitis, and family history of cancer-to improve patient self-reporting completeness and reduce information omissions caused by memory biases or unclear expression.

Generating structured initial diagnostic assistance reports: Automatically integrate patient-reported symptoms and medical history into pre-consultation summaries adhering to clinical documentation standards, highlighting critical physical signs and risk alerts. This will assist physicians in rapidly grasping patient profiles, shortening outpatient consultation times, and improving diagnostic efficiency.

Laying the foundation for subsequent clinical validation and application: Through small-scale simulation testing and expert evaluation, preliminarily verify the system's usability and clinical alignment, collect feedback from physicians and patients, and optimize interactive workflows. This will provide scientific evidence and ethical compliance support for future prospective clinical studies and product development.

This research is not intended for direct clinical diagnosis or treatment decisions but rather as an information-assistance tool for physicians prior to consultations. It aims to promote safe, effective, and equitable applications of AI in the early screening of hepatobiliary and pancreatic diseases and tiered healthcare delivery systems.

Interventions

• Other AI Integration type 1
  Patients first complete a full interaction with the multi-agent system until the Arbiter confirms the medical record is error-free. The system then generates a structured "Case Characteristics" (CC) summary and preliminary diagnostic recommendations via the Oracle Agent. However, the complete AI output is not displayed to the subsequent attending physician. The physician conducts an independent consultation following standard clinical protocols, and their medical records are solely used to maint
• Other AI Integration type 2
  After patients finalize the CC through interaction with the multi-agent system, the structured "Case Characteristics" (excluding Oracle-generated diagnostic advice to avoid over-guidance) are pushed in a standardized format to the corresponding physician's electronic workstation. Physicians may reference this summary before formal consultation to adjust their questioning focus, verify information accuracy, or supplement missing details. The physician's final written medical record serves as the

Primary outcome measures

• Concordance Rate Between AI-Generated Medical Records and Gold Standard [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]
• Diagnostic Accuracy of Physician Documentation With AI Summary Reference [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]
• Proportion of physician medical records meeting predefined quality criteria without AI support [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]

Secondary outcome measures (6)

• Average physician consultation duration [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]
• Average pre-consultation preparation time for physicians [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]
• Patient satisfaction measured by validated questionnaires [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]
• Physician satisfaction with AI workflow [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]
• Frequency of AI-related adverse events and workflow disruptions [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]
• System Usability Scale (SUS) scores for the AI interface [Time frame: From enrollment through completion of the index outpatient visit and chart assessment, assessed up to 7 days]

Eligibility criteria

Inclusion criteria

• Age and Gender: Patients aged 18 to 75 years, of either sex.
• Clinical Diagnosis Requirements: Suspected or confirmed hepatobiliary or pancreatic diseases (e.g., liver cancer, pancreatic cancer, cholangiocarcinoma, cirrhosis) based on preliminary clinical evaluation.
• Ability to provide a complete medical history and symptoms for AI system interaction.
• Cognitive and Physical Capacity:
• Sufficient cognitive function to complete interactions with the AI multi-agent system independently (verified by Mini-Mental State Examination \[MMSE\] score ≥24).
• Proficiency in Mandarin or English to ensure accurate communication with the system.
• Consent and Compliance: Willingness to participate and provide written informed consent.
• Ability to complete all study procedures, including physician consultations and follow-up assessments.
• Clinical Workflow Compatibility: Scheduled for outpatient consultation at participating healthcare facilities.

Exclusion criteria

• Patients with life-threatening conditions requiring immediate intervention (e.g., acute hepatic failure, severe hemorrhage).
• Presence of severe cardiovascular or cerebrovascular diseases that may interfere with study participation.
• Cognitive or Communication Barriers:
• Cognitive impairment (MMSE score <24) or language barriers preventing effective interaction with the AI system.
• Psychiatric disorders or altered mental status affecting decision-making capacity.
• Prior or Concurrent Participation:
• Enrollment in other interventional clinical trials that may confound the study outcomes.
• Current use of experimental diagnostic tools or AI systems outside the study protocol.
• Technical or Logistical Constraints:
• Inability to access or operate electronic devices required for AI system interaction (e.g., touchscreen terminals, mobile apps).
• Lack of stable internet connectivity for system access.
• Ethical or Legal Restrictions: Pregnancy or lactation (to avoid potential risks not directly related to the study).

Criteria are shown verbatim from the registry (in English). Final eligibility is always assessed by the study center.

Healthy volunteers: Yes

Study design

Observational model

Cohort

Study locations

Center list to be confirmed — check the primary protocol.

Publications

• Zhou LQ, Wang JY, Yu SY, Wu GG, Wei Q, Deng YB, Wu XL, Cui XW, Dietrich CF. Artificial intelligence in medical imaging of the liver. World J Gastroenterol. 2019 Feb 14;25(6):672-682. doi: 10.3748/wjg.v25.i6.672. PMID 30783371
• Cao LL, Peng M, Xie X, Chen GQ, Huang SY, Wang JY, Jiang F, Cui XW, Dietrich CF. Artificial intelligence in liver ultrasound. World J Gastroenterol. 2022 Jul 21;28(27):3398-3409. doi: 10.3748/wjg.v28.i27.3398. PMID 36158262
• Rompianesi G, Pegoraro F, Ceresa CD, Montalti R, Troisi RI. Artificial intelligence in the diagnosis and management of colorectal cancer liver metastases. World J Gastroenterol. 2022 Jan 7;28(1):108-122. doi: 10.3748/wjg.v28.i1.108. PMID 35125822
• Bo Z, Song J, He Q, Chen B, Chen Z, Xie X, Shu D, Chen K, Wang Y, Chen G. Application of artificial intelligence radiomics in the diagnosis, treatment, and prognosis of hepatocellular carcinoma. Comput Biol Med. 2024 May;173:108337. doi: 10.1016/j.compbiomed.2024.108337. Epub 2024 Mar 24. PMID 38547656

Identifiers

NCT: NCT07716670 · 2025-1371

Primary sources (government registries)

• ClinicalTrials.gov (US NIH)

View this study on ClinicalTrials.gov ↗