ChatASCPT: Use, Promise, Biases of Large Language Models

Session Description

Artificial Intelligence (AI) enables new discoveries and improved processes in the entire healthcare continuum. AI chatbots based on powerful pre-trained large language models (LLMs) are already deployed in the drug development process. These LLMs might be able to identify and validate new drug targets, design new drugs, optimize drug properties, and assess toxicity. They can summarize and analyze data from clinical trials and can perform market research. They can assist in the planning and execution of clinical trials by generating protocols and recruiting patients. LLMs may be applied to guide prescribing, and to assisting in real-world data (RWD) analysis.

However, the existence of meaningful risk for biased/unethical recommendations/conclusions must be assessed. Examples of this might include recommendations of drugs to develop, clinical trials to conduct, or a summary of RWD, leading to unequal access to medical treatment. LLMs, like other machine learning models, suffer from limitations in interpretability, in handling uncertainty, and from lack of transparency. A false response (“hallucination”), which is a noted issue of LLM-based chatbot outputs, can be particularly dangerous in medical scenarios.

Reproducible and ethical machine learning in healthcare is important, along with improved understanding of the bias that may be present in the models (e.g., demographic, cultural, and linguistic). More work should be done to establish how advice from biased AI can be mitigated by delivery method, for instance by presenting it descriptively rather than prescriptively. Because of concerns about both utility and safety, new applications will generally have to adhere to the same standards applied to other medical technologies. Ethical, governance, and regulatory considerations are critical in the design, implementation, and integration of every component of the AI applications and systems. Just before the fascinating LLMs are in widespread use in the pharmaceutical realm, it is our duty to discuss limitations, ethical consideration, and regulatory requirements.

Learning Objectives

1. Upon completion participants will be able to discuss LLMs use (such as chatbots) in context of the pharmaceutical environment and RWD analyses.
2. Participants will understand social biases in LLMs.
3. Participants will be able to discuss ways to mitigate social biases when using LLMs.

Session Schedule

11:00 AM - 11:05 AM MT
Introduction
Suraj Bhansali, MBA, PhD and Brian Cicali, PhD

11:05 AM - 11:15 AM MT
Leveraging LLMs to Aid in Analysis of Clinical Information from RWD
Wesley Anderson, PhD

11:15 AM - 11:25 AM MT
Towards Understanding and Mitigating Social Biases in Language Models
Paul P. Liang, MS

11:25 AM - 11:35 AM MT
Promoting Data Transparency, Documentation and Responsible AI Governance and Collaboration
Liang Zhao, PhD, MBA

11:35 AM - 12:00 PM MT
Q&A and Closing Remarks