Generative AI in healthcare: A trustworthy approach
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Generative AI, healthcare, trustworthy, Transformer Architecture, guardrails
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: In this paper, we focus on the method to achieve trustworthiness in an LLM (Large Language Model) based decision support system for physicians.
Abstract: Generative AI in healthcare: A trustworthy approach
Abstract: The recent advancements in self-supervised algorithms like Transformer Architecture and Diffusion models have expanded the means of applying AI in healthcare and life sciences. To achieve real world adoption, it is important to measure and audit the trustworthiness of the AI system as per the legal and compliance requirements for privacy, security, fairness, and safety. In this paper, we focus on the method to achieve trustworthiness in an LLM (Large Language Model) based decision support system for physicians. The stakeholders for this decision support system are patients, physicians, regulators, and external auditors. We focus on the limitations of large or foundation models and the method to overcome these limitations, with the aim of accelerating the adoption of this far-reaching technology in the healthcare sector. It also explores possible guardrails for safety and the methods for aligning AI systems to guardrails.
Our Solution Approach:
We explore an approach to an AI system which can enhance decision capabilities by using the data and EHRs (Electronic Health Record) collected over many years for a vast volume of patients. The longitudinal data consists of biomarkers, disease progression indicators, treatment administered, and patient outcome. The goal of the system is to assist physicians in identifying the best treatment option for a given patient context. The LLM-based system will be able to predict optimal options based on hundreds of similar cases on which it was trained. The paper addresses the transparency, data integrity, model development, and performance validation of the system. In the sections below, we explore the various stages of development and deployment of such a system, the challenges, and the methods to overcome the challenges.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 6887
Loading