Speakers
Mark Sendak, MD, MPP
Population Health & Data Science lead
Duke Institute for Health Innovation (DIHI)
Suresh Balu, MD, MBA
Director, Duke Institute for Health Innovation (DIHI)
Associate Dean, Innovation and Partnership
Duke School of Medicine
Keywords
AI, ML, health equity
Key Points
- The Duke Institute for Health Innovation’s (DIHI) mission is to catalyze transformative innovation in health and healthcare through high-impact research, leadership development and workforce training and the cultivation of a community of entrepreneurship.
- DIHI approaches this work through four pillars of innovation: implementation and health delivery science, health technology innovation, leadership and workforce development, and best practices development and dissemination.
- In 2021, DIHI started a Health AI Partnership to empower healthcare professionals to use AI effectively, safely, and equitably through community-informed up-to-date standards. We have continued to see the digital divide, where a small number of teams have expertise. Through this partnership, we are trying to build those skills in low resource settings and build a community of practice.
- The Health AI Partnership started with 7 organization partners and has expanded to about 20 organizations. The main deliverables for the first phase of the project were developing standard key decision points for the AI product lifecycle and developing the Health Equity Across the AI Lifecycle (HEAAL) Framework.
- There are 8 key decision points in the AI product lifecycle: identify and prioritize a problem; evaluate AI as a viable component of the solution; develop measures of outcomes and success of the AI product; design a new optimal workflow to facilitate integration; evaluate pre-integration safety and effectiveness of the AI product; execute change management, workflow integration, and scaling strategy; monitor and maintain the AI product; and update or decommission the AI product.
- The first key decision point is procurement, which begins with identifying a problem and ends with allocation of resources to either build or buy an AI product or solution.
- Key decision point 1 is procurement. Frontline staff have to have buy-in to submit a proposal so organizational resources can be used to sustain the innovation. Align frontline staff and organizational leaders – create alignment throughout project selection.
- The second key decision point is development and adaptation, which is either building or adapting an external solution for internal clinical use. The first step is to develop measures of success, the second step is designing the workflow, and the third step is evaluating pre-integration safety and effectiveness before a solution is put into clinical use.
- The next key decision point is clinical integration. DIHI built a modular infrastructure to support many projects via a flexible data pipeline technology infrastructure that started with one model and now includes dozens of models. The final key decision point is lifecycle management, which includes monitoring and maintaining the AI product and updating or decommissioning the product.
- The Health AI Partnership held a workshop that examined how to assess the potential future impact of a new AI solution on health inequities. The discussion resulted in five assessment domains to be evaluated across the span of the AI adoption process: accountability, fairness, fitness for purpose, reliability and validity, and transparency.
Learn more
Duke Institute for Health Innovation
Discussion Themes
-Can you talk about how DIHI/Health AI Partnership go about with the challenges in medical AI adoption outside of academic medical centers? There is a massive inequity and teams like DIHI are very rare. The next phase of work is building out a practice network in building AI capabilities. It will be a massive undertaking. We have a small philanthropic gift to get started. There is a need for support of infrastructure to bring AI partnerships to low resource settings.
–How do you get reliable information and data for care that is received outside of Duke? How do you mitigate bias because of incomplete data? For the CKD example, we combined data from Duke and claims data.
Tags
#pctGR, @Collaboratory1