Evaluating Medical Agent Pipelines for Diabetes Prediction

CS 6604 Course Project(Github)

Background and Challenge:

The “black box” nature of many LLM agents obscures their decision-making process, raising concerns for critical healthcare applications where transparency is essential. Additionally, existing agent systems struggle to effectively interpret individual patient data within the broader context of population-level trends, and there is limited research on how different types of agents (traditional ML and LLM-based) can collaborate effectively for comprehensive medical tasks.

Our Method:

• This project addresses these challenges by designing and implementing three distinct LLM agent architectures for diabetes classification and personalized advice generation.
• This research aims to identify the most effective pipeline for integrating LLMs into medical advisory systems while advancing the field of autonomous healthcare agents.
• Three pipelines:

Conclusion:

• A3’s ML method demonstrated consistent predictions across multiple runs.
• Benefits of Combining multi-agent framework with traditional ML methods: transparent reasoning, structured and detailed analysis, evidence-based personalization.