Evaluating artificial intelligence solutions from vendors is one of the biggest challenges informatics leaders face today. The Health AI Partnership (HAIP), a multi-stakeholder collaborative, has published in NEJM AI a description of its AI Vendor Disclosure Framework, a tool designed to support responsible AI system procurement. 

First some background on HAIP: It seeks to be a resource for guidance for healthcare professionals using AI and related emerging technologies, and a platform for community-generated, expert-curated guidance, resources, and standards for responsible AI adoption in healthcare. The group has developed a network that creates a safe space for peer advice and collaboration to address issues health system leaders face while adopting AI in healthcare settings. Its Coordinating Center team, located at the Duke Institute for Health Innovation (DIHI), manages and coordinates the partnership’s activities. HAIP received initial funding in 2022 from the Gordon and Betty Moore Foundation to establish this community resource.

According to HAIP,  the AI Vendor Disclosure Framework, which is publicly available and free to use, identifies essential information across five core domains that health systems should request — and vendors should disclose — to effectively evaluate vendor-developed AI systems: 

1. System Capabilities and Intended Use establishes foundational knowledge about the AI system’s functionalities, use, and affected stakeholders.
2. System Performance and Compliance establishes the AI system’s operational metrics, potential biases, associated risks, and regulatory standing.
3. Data Stewardship outlines the approach to data governance, including security measures, quality assurance processes, secondary use, and retention policies.
4. Integration Requirements evaluate the total cost of ownership, including technical prerequisites, resource requirements, and implementation timelines.
5. Lifecycle Management defines vendor responsibilities for ongoing support, monitoring, and maintenance after implementation. 

By standardizing expectations for the information needed in procurement decision-making, the framework aims to enhance transparency and promote safer health care AI adoption. It serves both as a best-practice guide and a customizable resource to support healthcare delivery organizations in procuring vendor-developed AI systems.

Vega Health Co-founder and CEO Mark Sendak, M.D., M.P.P, was part of the framework's development team. On LinkedIn, he explained the significance of the new resource. He wrote that “Model facts labels/model cards are great resources for front-line clinicians who need a high-level synthesis of what an AI solution is, how it was built, and how it should be used.” But he added that the model facts label/model card is not sufficient information to guide procurement and implementation decisions. “The information needed for those stakeholders is much more comprehensive across domains and much more detailed,” Sendak wrote. “Most folks don't appreciate the difference. Hence, through Health AI Partnership we pulled together a group of leaders across multiple institutions who were already building out these vendor assessments precisely because the information provided by vendors was insufficient for procurement decisions.”

Before founding Vega, which seeks to curate a marketplace of healthcare AI solutions proven safe and effective in real-world settings, Sendak was a population health and data science lead at the Duke Institute for Health Innovation.