MCP-Driven Multi-Agent RAG-Enhanced LangGraph-orchestrated LLM System for Prostate Cancer Decision Support

📍 Presented at SIIM-CAIMI 2025 (Society for Imaging Informatics in Medicine)
🏥 University of Chicago – MS in Applied Data Science Capstone

🔬 Project Overview

This project presents an auditable, multi-agent Retrieval-Augmented Generation (RAG) framework designed to assist oncology workflows through:

• Longitudinal temporal summarization
• Evidence-grounded literature integration
• Treatment recommendation via supervised ML
• Lifespan estimation via survival analysis
• Hallucination detection and validation loop

The system processes 500 synthetic longitudinal prostate cancer records and produces:

• Structured timeline summaries
• Literature-verified clinical context
• Ranked treatment recommendations (with probabilities)
• Survival probabilities at 5-, 10-, and 15-year horizons
• Expected lifespan estimates (in years)

🏗 System Architecture

The architecture follows an MCP-governed modular design with strict tool mediation and validation.

High-Level Flow

Patient ID
   ↓
MCP Server (secure context retrieval)
   ↓
LangGraph Orchestration
   ├── Retrieval Tool (PubMed phenotype-aware query builder)
   ├── Summarizer Agent (structured clinical report generation)
   ├── Validator Agent (hallucination + missing data detection)
   ↓
Structured Model APIs
   ├── XGBoost Treatment Model
   ├── Survival Ensemble (Cox + Weibull + RSF)
   ↓
Final Validated Clinical Report

🧠 Model Architecture Diagram (for README)

flowchart TD

A[Patient ID] --> B[MCP Server<br/>Longitudinal Context Retrieval]

B --> C[LangGraph Orchestrator]

C --> D[Phenotype-Aware RAG Tool<br/>PubMed Query + Scoring]
C --> E[Summarizer Agent<br/>Structured Clinical Report]
C --> F[Validator Agent<br/>Hallucination & Missing Data Check]

E --> F
F -->|Retry if Needed| E

F --> G[Treatment Prediction API<br/>XGBoost Classifier]
F --> H[Lifespan Estimation API<br/>Cox + Weibull + RSF]

G --> I[Ranked Therapy + Probabilities]
H --> J[5/10/15-Year Survival + Expected Years]

I --> K[Final MCP-Audited Clinical Report]
J --> K

🛠 Technical Stack

🧩 LLM & Orchestration

• LangGraph (multi-agent workflow control)
• OpenAI GPT-4 (summarizer + validator agents)
• Prompt-constrained structured generation
• Retry routing controller with bounded iterations

🔐 Governance Layer

• Model Context Protocol (MCP) Server

  • Versioned patient context retrieval
  • Tool mediation
  • Auditable endpoint calls
  • Metadata tracking (model name, version, timestamp)

📚 Retrieval (RAG)

• PubMed XML API

• Phenotype-aware query builder

• Signal extraction via regex parsers

• Evidence scoring:

  • Clinical alignment
  • Recency filtering (≥2016)
  • Endpoint relevance
  • Novelty weighting

• Deterministic citation embedding (verbatim insertion)

📊 Treatment Recommendation Model

• XGBoost classifier

• Features:

  • TNM stage
  • Gleason grade
  • PSA trajectory
  • PSA velocity
  • Metastatic indicators
  • Treatment history

• Output:

  • Top-N ranked therapies
  • Class probabilities
  • Feature-driven rationale

• Patient-level train/test split

• Synthetic dataset accuracy: 1.00 (upper bound, not clinical claim)

📈 Lifespan Estimation Model (Ensemble Survival Framework)

Three complementary models:

1. Cox Proportional Hazards (interpretable hazard ratios)
2. Weibull Regression (stage-stratified baseline survival curves)
3. Random Survival Forest (nonlinear feature interactions)

Workflow:

• TNM-based stratification (localized / N1 / M1)
• Weibull baseline curve
• Cox-based patient-specific risk shift
• RSF nonlinear modulation
• Ensemble averaging

Outputs:

• 5-, 10-, 15-year survival probabilities
• Expected survival time (years)
• Monotonic survival validation checks

Internal QA:

• Survival ordering check (M1 < N1 < localized)
• Probability bounds enforcement
• Curve monotonicity

🏥 Conference Presentation

This work was presented at:

SIIM-CAIMI 2025
Society for Imaging Informatics in Medicine – Conference on Artificial Intelligence in Medical Imaging

• Poster: AI-in-oncology-Poster.pdf

• Featured in [UChicago Data Science Institution News]

📁 Data

• 500 synthetic longitudinal prostate cancer records

• 5–7 time-stamped visits per patient

• Variables include:

  • PSA (with kinetics)
  • Gleason grade
  • TNM stage
  • Bone lesion count
  • Visceral metastasis flag
  • ALP, LDH, albumin, hemoglobin
  • Treatment history
  • Weight trends

Validated for medical plausibility by a practicing radiologist.

🔎 Evaluation & Validation

Structured Models

• Held-out patient-level testing
• Directional consistency validation
• Internal statistical QA checks

LLM Components

• Dedicated validator agent

• Detection of:

  • Hallucinated content
  • Missing patient data

• Iterative retry loop

• Verbatim literature line insertion (no citation hallucination)

No summary finalized with unresolved hallucinations.

🎯 Key Contributions

• MCP-governed agentic clinical AI framework
• Hallucination-resistant RAG integration
• Survival ensemble integrated into LLM workflow
• Deterministic literature grounding
• Modular API-based predictive model integration
• Fully auditable report generation pipeline

📌 Research Context

This project addresses limitations in current oncology AI systems:

• Lack of temporal reasoning
• Hallucination in generative summaries
• Non-auditable clinical AI outputs
• Separation between ML survival models and narrative reasoning

The architecture demonstrates a reproducible pattern for safe LLM deployment in healthcare.

⚠ Disclaimer

This system was trained and evaluated on sample data and is intended for research demonstration only.