MS in Applied Data Science Students Present AI-Powered Clinical Decision Support System at SIIM 2025

Three University of Chicago Master’s in Applied Data Science (MS-ADS) students, Amulya Jayanti, Amy Kim, and Zoe Calianos, recently presented their capstone research at the Society for Imaging Informatics in Medicine (SIIM) 2025 Conference on Artificial Intelligence in Medical Imaging in Irvine, California. Their capstone project, developed in collaboration with UChicago Medicine radiology and titled MCP-driven RAG-Enhanced LLM System for Prostate Cancer Care, leverages multi-agent large language models to improve the reliability, interpretability, and efficiency of clinical decision-making for prostate cancer care.

Using Generative AI to Enhance Oncology Decision Support

The team’s capstone project tackles one of healthcare’s most common challenges: fragmented patient data. Healthcare providers often spend hours piecing together scattered information from medical records, lab results, and imaging to assess disease progression. Additionally, prostate cancer care relies on longitudinal data, meaning clinicians must interpret patterns over time to understand how a patient’s condition is evolving before turning to the research literature to make treatment decisions.

The students built an AI system that does this work automatically, creating comprehensive patient timelines that integrate medical history with treatments, outcomes, and relevant research. “We aimed to make oncology care easier, more reliable, and faster for clinicians,” Calianos explained. The system outputs a clear summary for each patient, including lifespan estimation and treatment recommendations.

What makes this project stand out is the way the team brings several advanced AI components into a single, cohesive clinical system. The team used a combination of Retrieval Augmented Generation (RAG), a Model Context Protocol (MCP) server, a multi-agent framework, and clinical literature-backed machine learning models to support lifespan analysis and treatment recommendation.

RAG allows the system to pull in relevant medical research from sources like PubMed, ensuring patient summaries are grounded in current clinical literature. The MCP server maintains consistent patient context across the entire system, helping prevent information loss and reduce inconsistencies.

Within this framework, multiple AI agents work collaboratively: one generates a structured summary of the patient’s history, while another acts as a validator, reviewing the output for missing information, logical errors, or AI “hallucinations,” instances where the model produces plausible-sounding but incorrect details. The validator repeats this process until the summary meets accuracy and reliability standards, resulting in outputs that clinicians can trust and interpret with confidence.

“The multi-agent structure allows for internal reasoning and self-validation,” said Jayanti. “The model not only produces recommendations but also shows its rationale, addressing both interpretability and reliability—critical concerns in clinical decision support.”

RAG pulls from PubMed to provide clinicians with relevant, recent medical research studies tied directly to each patient case.

From Capstone to Conference

The students credit their capstone advisor, Professor Utku Pamuksuz, for encouraging them to push boundaries. “We’ve been iterating on this project since February,” said Calianos. “He constantly challenged us to try different frameworks and validation methods. For example, we briefly experimented with using CrewAI, a newer tool for coordinating AI agents, before shifting to LangGraph, which is more widely adopted.”

After refining their framework over the summer, the team submitted an abstract to SIIM 2025. Their advisor, Utku Pamuksuz, and UChicago Medicine collaborator, Dr. Bora Kalaycioglu, had shared the conference’s call for submissions. The team applied, and when they found out their abstract was accepted, they jumped into action. “Once the abstract was selected, we started to make progress much, much quicker than we did before,” said Jayanti.

At the conference, they engaged with experts in both AI and medical imaging, including Dr. Khan Siddiqui, Dr. Elliot Siegel, and Dr. Ingrid Reiser. “It was incredible to discuss our work with professionals already transforming the field,” said Jayanti. 

Calianos added, “A lot of our conversations were around new, groundbreaking tools and AI applications, and it was awesome to be able to contribute.”

Real-World Impact and Career Inspiration

For the students, the project bridged classroom learning with real-world healthcare innovation. Jayanti explained that she had to do some self-learning to work with clinical models such as Weibull function, Cox Proportional Hazards, and Random Survival Forest, but that the foundation came from the machine learning courses they had taken. 

Kim, who focused on the treatment recommendation machine learning component, said that Dr. Bousetouane’s AI course helped her apply what she’d learned to the project. The experience, she added, sparked a deeper interest in healthcare applications of AI. “I wasn’t planning to go into healthcare at first, but through this project, I realized how impactful data science can be for improving patient outcomes.”

“Our coursework in agentic AI with Professor Pamuksuz gave us the foundation for this work,” said Calianos. “I even interned at a healthcare AI startup over the summer, directly inspired by this project.”

Turning Data into Impact

The team built a transparent, explainable system that turns fragmented prostate cancer data into clear, evidence-backed summaries. By providing traceable rationale instead of black-box outputs, the tool helps clinicians spend less time gathering information and more time on patient care. Their work exemplifies how UChicago’s MS in Applied Data Science program equips students to tackle complex, high-impact challenges.