Multi-agent LLMs for Decision Support in Cervical Cancer During Pregnancy

Overview

The aim of this study is to develop an AI-assisted decision-making system based on multi-agent large language models and to evaluate its effectiveness and accuracy in the diagnosis and treatment of cervical cancer during pregnancy.

This project intends to construct China's first artificial intelligence model for the vertical field of "multidisciplinary team (MDT) consultation" for cervical cancer during pregnancy. Centered on the massive case data of gynecological oncology from Obstetrics and Gynecology Hospital of Fudan University, combined with cervical cancer during pregnancy guidelines to formulate multi-oncology department judgment standards, it will focus on overcoming key technical bottlenecks such as model reliability, model result evaluation, and multi-agent collaborative scheduling. With this model as the engine, a trinity AI hub will be built, driven by agent collaboration and combined with a guideline-based evaluation system to realize intelligent support for " cervical cancer during pregnancy MDT consultation + guideline evaluation". By constructing a vertical model for MDT consultation for cervical cancer during pregnancy, MDT AI agent consultation, and a guideline-based evaluation system, the project will comprehensively improve the standardization, homogenization level and efficiency of diagnosis and treatment, promote the upgrading of the diagnosis and treatment capabilities for cervical cancer during pregnancy diagnosis and treatment capabilities, and provide more accurate and high-quality medical service guarantees for patients. 1. Vertical Large Language Model Based on the guidelines for cervical cancer during pregnancy and other relevant authoritative guidelines and expert consensus, it provides high-quality structured knowledge support and reliable decision-making explanation basis for the gynecological oncology vertical model. Construct a high-quality gynecological oncology vertical corpus: Address the difficulty of manual corpus construction, build controllable data generation based on the existing full tumor process, establish a gynecological oncology corpus with high accuracy and strong generalization ability, and enhance the quality of fine-tuning data. Construct a vertical gynecological oncology large language model: Improve the basic professional capabilities of the model through methods such as maximizing internal coherence and measuring mutual predictability; introduce a length penalty mechanism and neighborhood-adaptive reinforcement learning based on large language models to enable the language model to discriminate gynecological oncology logic. Through a reference reward mechanism based on standard answers, the gynecological guideline reward model supports the reinforcement learning of the gynecological model and solves the key problem of model interpretability. Improve core clinical tasks such as early screening, accurate staging diagnosis, personalized treatment plan recommendation, risk stratification assessment and intelligent follow-up of gynecological tumors, achieve or exceed the level of international advanced similar models in key performance indicators, and provide scientific and reliable intelligent support for clinical diagnosis and treatment decisions. 2. Agent Construction Construct a collaborative intelligent platform driven by exclusive diagnosis and treatment agents for each department: With hierarchical modular design as the core and agent cluster as the driving force, it covers the full-cycle diagnosis and treatment process of cervical cancer during pregnancy. The underlying data layer integrates clinical business data and specialized knowledge bases to support the diverse AI capabilities of the model layer; the Agent framework layer, as the technical hub, realizes multi-model scheduling, long-term and short-term memory management, and automatic arrangement of diagnosis and treatment processes through modules such as model gateway, memory enhancement, and strategy orchestration; the platform management layer provides visual scenario configuration, rule engine and workflow management to ensure medical compliance; the application layer includes gynecological oncology, neonatology, chemotherapy, and obstetrics agents to form a collaborative network. Design task routing and time-sharing classification scheduling strategies, assign tasks based on Agent capabilities and load, prioritize quality control and auxiliary diagnosis and treatment tasks, and improve computing power efficiency; realize the full life cycle management of Agents, including registration, release, and optimization, configure vertical model binding, prompts, etc., optimize corpus quality through user feedback, and coordinate the division of labor among multiple Agents through workflow engine to solve problems such as dynamic branches of complex diagnosis and treatment paths and knowledge tool collaboration, ensuring efficient collaboration. Construct the platform's intelligent scheduling and management capabilities for Agents: Design task routing to assign tasks based on Agent capabilities and current load; statistically analyze Agent usage efficiency and frequency to design execution priorities, and improve the efficiency of computing resource utilization. Design time-sharing classification scheduling strategies based on the confidence interval of AI in clinical business and the feasibility priority of non-core process substitution. Provide independent deployment capabilities for core businesses to avoid computing resource contention, ensure scheduling executability, and improve business resilience. 3. Guideline-Based MDT Decision Evaluation System Analyze and restructure knowledge based on the Guidelines for cervical cancer during pregnancy, extract key diagnosis and treatment points applicable to MDT, and establish an exclusive evaluation index system covering gynecological oncology, obstetrics, chemotherapy, and neonatology to quantitatively evaluate and compare diagnosis and treatment plans from different sources. In real case datasets, we will retrospectively enrolled patients diagnosed as cervical cancer during in obstetrics and gynecology hospital from January 2007 to December 2025. The inclusion criteria is as follows: 1) Pathologically confirmed diagnosis of cervical cancer; 2) Confirmed intrauterine pregnancy status via ultrasound. 3) Patients receiving initial treatment. 4)Agreement to participate in the study with signed informed consent. The exclusion criteria is as follows: 1) Previous treatment received for cervical cancer during pregnancy. 2) Pathological pregnancy states (e.g., ectopic pregnancy). 3) Inability or unwillingness to provide signed informed consent. For each case, we will collect diagnosis and treatment opinions from multi-disciplinary agents, junior doctors, and junior doctors after referring to agent results respectively. Then we calculate and compare accuracy and consistency scores according to evaluation indicators of each discipline for the three parties' results. In virtual case datasets, similar to criteria for the retrospective study section, 100 virtual cases of cervical cancer during pregnancy were generated. These cases were randomly divided in a 1:1 ratio into MDT-agents group and a real MDT team group. We will calculate accuracy and compare results according to the same evaluation indicators. Besides, we will also compare the consuming time for each case by MDT-agents and real MDT team respectively.

Conditions

Interventions

Eligibility

Outcomes

Central Contacts