Empowering Healthcare with
Knowledge-Augmented Large Language Models

Introduction

In the rapidly evolving landscape of healthcare technology, the integration of Large Language Models (LLMs) stands at the forefront of innovation, offering unprecedented opportunities to transform patient care, medical research, and healthcare operations. However, current mainstream generative LLMs, such as OpenAI’s ChatGPT and Meta’s LlaMa 2, may generate inaccurate, misleading, or even harmful content due to their inherent uncertainly, incompleteness, or bias from training data, among others. These issues can have serious consequences for patient safety when these models are directly used for health applications. Our collaborative workshop, jointly organized by AMIA KDDM, NLP, and KRS Working Groups, aims to delve into the cutting-edge advancements that augment LLM capabilities, focusing on Retrieval-Augmented Generation (RAG), in-context learning, contrastive learning, LLM finetuning (e.g., low-rank adaptation), prompt tuning, and zero-short/few-shot learning. These techniques can be broadly categorized into prompting strategies, retrieval-based strategies, and finetuning strategies, and their optimal use is still in debate. These sophisticated techniques are pivotal in enhancing the quality, accuracy, and reliability of LLM-based applications across a broad spectrum of healthcare domains. One prime example is the development of advanced medical Question Answering (QA) systems that leverage RAG to provide precise, evidence-based answers drawn from vast medical literature and patient data. Similarly, the use of contrastive learning methods enables the creation of more robust diagnostic tools by effectively distinguishing between complex cases with subtle differences in symptoms or imaging. Few-shot learning techniques empower LLMs to adapt to specialized medical fields quickly, using minimal example data to generate knowledgeable responses or identify rare conditions. RAG combines the benefits of retrieval-based methods, which search through a database to find relevant information, with the generative capabilities of LLMs, which can synthesize and predict text. This approach allows LLMs to pull in the most current information from medical databases, research papers, or clinical guidelines, ensuring the generated responses are based on the latest medical knowledge. Beyond these, the workshop will explore explainability and interpretability of LLMs, as well as the ethical considerations in deploying these advanced AI tools in sensitive healthcare environments. By convening experts from healthcare, AI research, and ethics, this workshop seeks to explore the potential of these technologies to provide up-to-date medical information, facilitate better decision-making, and improve patient outcomes. We invite the attendees to navigate the intersection of AI innovation and medical expertise, charting a course toward a future where technology and healthcare work hand in hand to achieve optimal health outcomes, all while ensuring ethical standards and patient privacy are upheld.

By convening experts from healthcare, AI research, and ethics, this workshop seeks to explore the potential of these technologies to provide up-to-date medical information, facilitate better decision-making, and improve patient outcomes. We invite the attendees to navigate the intersection of AI innovation and medical expertise, charting a course toward a future where technology and healthcare work hand in hand to achieve optimal health outcomes, all while ensuring ethical standards and patient privacy are upheld.

Objective

Our intended audience includes individual (e.g., clinicians, researchers, policymakers, and trainees) who are interested in practical applications of AI and LLM in healthcare and have working knowledge of LLMs. Researchers and clinicians in the clinical and healthcare community mainly focusing on traditional knowledge-driven approaches are welcome to join and converse with researchers from the data science and computer science fields. In terms of the level of expected backgrounds, 25% is basic, 50% is intermediate, and 25% is advanced.

At the conclusion of this workshop, a participant will be able to:

• Understand the promises and limitations of LLMs for healthcare applications.
• Summarize the state-of-the-art research and techniques to improve LLM quality and accuracy.
• Identify research and development opportunities in this space and potential collaborators for deeper engagement.
• Understand the clinical relevance of this work and how it can be capitalized to improve patient outcomes and biomedical research.

Accepted Posters