Re-envisioning Clinical Search with Retrieval Augmented Generation

Abstract

Come along for the ride.

You are a doctor doing a quick sense-check of clinical guidelines for drug dosage while visiting a patient at the ICU. Or perhaps you are at an outpatient appointment and need prompt references for drawing up a treatment plan that accounts for a specific patient's comorbidities and current medications.

What if you had a Clinical search agent that could interpret your question, and retrieve the correct answers with explanations from drug monographs and clinical overviews? Could we leverage Generative AI for clinical Q/A?

In project Apollo of Elsevier Health Markets, we are working to re-envision search with a natural language question-answer interface on top of more robust document information retrieval. However, given the stochastic nature of Large Language Model (LLM) output, and how critical safety is for the clinical use case, there are several challenges and risks we must carefully build around. Thus, evaluation plays a central role in the development process.

This talk will touch upon two themes:
1. the supporting use case and technical implementation details of how we approach RAG for clinical settings that are high stake with almost no room for error. For instance, we need to retrieve section level information, and not just abstracts.

2. how we are testing and placing checks to protect, or at the very least make transparent the impact of the chattiness and hallucinations of LLMs.

Keywords: Retrieval augmented generation, LLM, Search, Clinical search, Evaluation