MCP-Driven Multi-Agent RAG-Enhanced LLM System for Clinical Decision-Making in Prostate Cancer

A clinical AI system co-developed by the University of Chicago Medical Center, integrating the MCP governance framework, multi-agent RAG retrieval, and survival analysis models, provides evidence-based intelligent decision support for prostate cancer diagnosis and treatment. This project is maintained by hyunji0618, sourced from the GitHub project "MCP-Driven-RAG-Enhanced-LLM-for-Oncology". It aims to address core challenges faced by current oncology AI systems, such as insufficient time-series reasoning capabilities, hallucination issues, and lack of auditability, providing doctors with intelligent decision support backed by traceable evidence.

Current oncology AI systems face several core challenges: insufficient time-series reasoning capabilities, hallucination issues in generative summaries, lack of auditability in clinical AI outputs, and the disconnect between machine learning survival models and narrative reasoning. These issues severely restrict the safe deployment of AI in the medical field.

Prostate cancer diagnosis and treatment, in particular, require the integration of multi-dimensional information—from dynamic PSA changes, Gleason scores to TNM staging, as well as treatment history and metastasis indicators. Doctors need not only prediction results but also an understanding of the model's reasoning process and the ability to trace the evidence source for each recommendation.

System Architecture Flow

After entering the patient ID, secure context retrieval is first performed via the MCP server to obtain longitudinal medical record data. Then, the LangGraph orchestrator coordinates the collaborative work of multiple agents: the retrieval tool constructs PubMed phenotype-aware queries, the summary agent generates structured clinical reports, and the verification agent performs hallucination detection and missing data checks. The system also integrates structured model APIs: the XGBoost classifier for treatment plan recommendations, and the integrated survival analysis model (Cox proportional hazards + Weibull regression + random survival forest) for life expectancy prediction. The final output is a complete clinical report audited by MCP.

Core Modules of the Tech Stack

• MCP Governance Layer: Provides versioned patient context retrieval, tool mediation, auditable endpoint calls, and metadata tracking (model name, version, timestamp) to ensure transparency and reproducibility.
• RAG Retrieval Mechanism: Uses the PubMed XML API, equipped with a phenotype-aware query builder. Evidence scoring considers clinical alignment, timeliness (≥2016), etc., and eliminates hallucinations through deterministic reference embedding.
• Treatment Recommendation Model: Based on the XGBoost classifier, inputs features such as TNM staging and Gleason score, outputs sorted treatment plans, probabilities, and recommendation reasons.
• Integrated Survival Analysis Framework: Cox proportional hazards model (interpretable hazard ratio), Weibull regression (stage-stratified baseline curve), and random survival forest (nonlinear interactions) work collaboratively to output 5/10/15-year survival probabilities and expected time.

Hallucination Detection and Quality Control

The system sets up a verification agent to detect hallucinatory content and missing patient data, using an iterative retry loop to ensure no unresolved hallucinations enter the final report. Additionally, survival order checks (M1 < N1 < localized stage), probability boundary constraints, and curve monotonicity verification are performed.

Clinical Validation and Academic Recognition

The system has been validated at the University of Chicago Medical Center using 500 synthetic longitudinal prostate cancer records (each with 5-7 visit records), and the data has been verified for medical rationality by practicing radiologists. The project results have been published at SIIM-CAIMI 2025 and featured in a special report by the University of Chicago Data Science Institute.

Key Innovations

1. MCP-governed agent-based clinical AI framework: Provides a reproducible architectural model for the safe deployment of medical AI.
2. Anti-hallucination RAG integration: Reduces hallucination risks through deterministic literature references and verification loops.
3. Survival analysis integration: Seamlessly integrates traditional survival models into the LLM workflow.
4. Modular API design: Decouples predictive models from narrative reasoning for independent iteration.
5. Fully auditable report generation: Complete traceability chain from input to output.

Limitations and Outlook

The system is trained and evaluated based on sample data and is for research demonstration only; synthetic data cannot fully replace the complexity of real clinical data. Future directions include expanding to more cancer types, integrating real-world evidence, and validating effects through prospective clinical trials.

This project demonstrates a feasible path for the safe deployment of large language models in the medical field. Through MCP governance, multi-agent collaboration, RAG evidence retrieval, and strict verification mechanisms, the system ensures clinical relevance while minimizing AI hallucination risks. For researchers and practitioners concerned with the safety and interpretability of medical AI, this is a reference implementation worth in-depth study.