Click to View Presentation

Presentation Time: 01:00 PM - 04:00 PM

Abstract Keywords: Large Language Models (LLMs), Artificial Intelligence, Natural Language Processing
Primary Track: Applications
Programmatic Theme: Clinical Informatics

Large Language Models (LLMs) like GPT-4 and Llama-3 have demonstrated remarkable capabilities in natural language understanding and generation. However, deploying these models in clinical settings requires adapting them to domain-specific data and aligning their behavior with medical knowledge and ethical standards. Fine-tuning on clinical texts (e.g., MIMIC-IV ICU notes) can greatly improve an LLM’s utility for tasks such as summarizing patient records or answering medical questions. Yet, full fine-tuning of multi-billion-parameter models is often impractical due to computational cost. Moreover, alignment with human preferences is critical: even after supervised fine-tuning, models may generate inappropriate or incorrect outputs without further tuning. This tutorial addresses these challenges by introducing parameter-efficient fine-tuning (PEFT) techniques and reinforcement learning (RL) methods for alignment, enabling the medical informatics community to adapt LLMs effectively and safely to clinical applications. We propose a half-day hands-on tutorial split into two parts. In the first half, participants will learn how to perform supervised fine-tuning of LLMs on medical text using PEFT methods. In the second half, we will explore reinforcement learning approaches for model alignment (often termed RLHF – Reinforcement Learning from Human Feedback and newer variants) to ensure the tuned LLM’s outputs adhere to clinical norms and user intentions. Throughout, we will use examples from the MIMIC-IV clinical note dataset to illustrate concepts, with an emphasis on practical implementation using open-source tools. The website for this tutorial is available at: https://sites.google.com/view/amia-2025-llm-sft-rl/home, and the GitHub repository can be accessed here: https://github.com/PittNAIL/amia25-llm-sftrl-tutorial.

Speakers:
Yanshan Wang, PhD
University of Pittsburgh

Satya Sahoo, PhD, FAMIA
Case Western Reserve University

Authors:
Yanshan Wang, PhD - University of Pittsburgh; Satya Sahoo, PhD, FAMIA - Case Western Reserve University; Hang Zhang, MS - University of Pittsburgh;

---