AI in Oncology: MCP driven RAG-enhanced LLM for Temporal Summarization of Oncology Treatment Responses

Project Status: Work in Progress

This project addresses a critical gap in oncology: the lack of AI systems that can provide explainable, clinically grounded summaries from complex, longitudinal patient data. We are developing a multi-agent, RAG-enhanced LLM pipeline to generate temporally structured, evidence-based summaries of prostate cancer patient records. Our system also aims to provide treatment recommendations and lifespan predictions to reduce clinician burden and support real-world oncology decision-making.

Problem Statement

In prostate cancer care, clinicians must synthesize complex data from various sources over time, including Prostate-Specific Antigen (PSA) scores, PI-RADS scores, weight changes, and patient-reported bone pain. Existing AI tools often fail to handle this temporal complexity, leading to fragmented care, a heavy documentation burden on clinicians, and a risk of generating inaccurate "hallucinations" in their outputs.

Methodology

Our solution uses a Multi-Agent RAG (Retrieval-Augmented Generation)-Enhanced LLM pipeline orchestrated by CrewAI. This architecture is designed to handle both structured and unstructured data through a Model Context Protocol (MCP), ensuring that all information is processed in a clinically logical, time-sequenced manner.

Data and Sampling

We used a synthetic dataset of 500 longitudinal prostate cancer cases for preliminary testing. These cases were generated via ChatGPT and refined with continuous feedback from a Radiologist and Assistant Professor of Radiology to ensure clinical realism. This allowed us to build and test our pipeline without using real Electronic Health Records (EHRs), which will be used for final validation in Summer 2025.

Core Components

Our pipeline is built on three main components, each addressing a specific clinical need:

1. Temporal Summarization

• Goal: Generate coherent, evidence-grounded summaries of a patient's clinical history.
• How it works: An agent parses dated entries for PSA, PI-RADS, weight, bone pain, and treatment history to create a structured timeline. This process ensures accurate temporal sequencing and the exclusion of irrelevant data, forming the foundation for all other reasoning tasks.

2. Treatment Prediction

• Goal: Recommend the most probable treatment based on a patient's progression.
• How it works: We trained a supervised classification model on the synthetic dataset, where each dated visit was an instance. We engineered features like PSA_Delta and weight_Delta to capture progression between visits. The model's output is the treatment administered at that point in the timeline, categorized into classes like Surgery, Radiation, or ADT.

3. Lifespan Prediction

• Goal: Provide a survival prognosis in years based on a patient’s clinical trajectory.
• How it works: This is a supervised regression task. We engineered 11 features that capture both static (e.g., max PI-RADS) and temporal (e.g., PSA_Change) dynamics. We also quantified bone pain severity and created binary flags for a patient’s treatment history. The final model predicts lifespan, with the ground truth being a synthetically generated value that rewards aggressive treatment and penalizes poor clinical indicators.

Results and Clinical Relevance

Preliminary tests on synthetic data confirm that our pipeline components are functionally integrated. The system accurately sequences temporal data and a built-in validator agent effectively flags hallucinations, ensuring that outputs are reliable and ready for human review.

This work is highly relevant to clinical practice because it can:

• Streamline Workflows: Automate the time-consuming process of creating patient summaries.
• Reduce Documentation Burden: Clinicians can focus on patient care instead of fragmented data interpretation.
• Enhance Explainability: The system provides evidence-supported reports, enabling explainable treatment planning with minimal risk of hallucination.

Final validation using real-world data from our partner hospital is scheduled before end of 2025. Key metrics for this phase will include clinical alignment, information completeness, and blinded user feedback.