What's Happening?
Sharon Davis, PhD, Assistant Professor of Biomedical Informatics at Vanderbilt University Medical Center, has been awarded a four-year, $1.6 million research grant from the National Library of Medicine, part of the National Institutes of Health (NIH).
The grant will fund her research into addressing weaknesses in artificial intelligence (AI) and machine learning predictive models used in medicine. These models often suffer from 'model drift,' where their performance degrades over time due to changes in clinical practice, information systems, and patient populations. Another challenge is poor generalizability across diverse patient groups, which can lead to disparities in accuracy. Dr. Davis and her team will investigate how accuracy gaps between patient groups evolve and develop methods to monitor and update these models to restore overall and subgroup accuracy. Their research will utilize data from the Department of Veterans Affairs (VA) and Vanderbilt Health, focusing on prediction models for surgical complications, hospital-acquired acute kidney injury, and postpartum hemorrhage. The project also aims to account for how effective AI tools can alter the outcomes they predict, a phenomenon that can be mistaken for model drift.
Why It's Important?
The degradation and lack of generalizability in AI predictive models pose significant risks to patient outcomes and equitable access to care in the U.S. healthcare system. If AI models become less accurate or perform unevenly across different patient demographics, it could lead to misdiagnoses, ineffective treatments, and widening health disparities. This research is crucial for ensuring that AI tools, which are increasingly integrated into clinical decision-making, remain reliable and fair for all patients. By developing methods to detect and correct model drift and improve generalizability, Dr. Davis's work can enhance patient safety and the quality of care. The findings could lead to more robust and trustworthy AI applications, fostering greater confidence among healthcare providers and patients in these advanced technologies. This initiative also highlights the ongoing effort to refine AI's role in medicine, moving towards more precise and personalized healthcare solutions that benefit a broader population.
What's Next?
The research project, supported by the NIH grant, will span four years. During this period, Dr. Davis and her team will track performance gaps in traditional machine learning and large language models, develop new monitoring and updating methods, and use synthetic patient data to support monitoring for smaller patient groups, such as those in rural hospitals. The postpartum hemorrhage prediction tool currently in use at Vanderbilt Health will serve as a testing ground for understanding how AI tools influence the outcomes they predict. The goal is to provide health systems with practical strategies to maintain the accuracy and dependability of AI tools for all patients. The outcomes of this research are expected to inform best practices for AI model deployment and maintenance in clinical settings, potentially leading to new guidelines or standards for the responsible use of AI in healthcare across the nation.
Beyond the Headlines
This research delves into the ethical and practical complexities of deploying AI in healthcare. The issue of 'model drift' and 'generalizability' touches upon fundamental questions of fairness and equity in AI. If AI models are not continuously monitored and adapted, they can perpetuate or even amplify existing biases within healthcare data, leading to unequal care for different patient groups. The project's focus on using synthetic data for small groups, like rural hospital patients, underscores the challenge of ensuring AI benefits underserved populations. Furthermore, understanding how an AI tool's effectiveness can change the very outcomes it predicts highlights a sophisticated feedback loop that requires careful consideration. This work is not just about technical improvements; it's about building a framework for ethical AI governance in medicine, ensuring that technological advancements genuinely serve to improve health outcomes for everyone, rather than inadvertently creating new forms of disparity.
