The Coalition for Health AI (CHAI) unveiled initiatives aimed at creating safer and more equitable AI applications in healthcare.
Addressing AI transparency: The model card initiative
The CHAI Model Card initiative aims to create a structured and transparent framework for evaluating AI models. Brian Anderson, MD, CEO of CHAI, likened this card to a "nutrition label" for AI. It will provide information, including the identity of the AI developer, intended uses, target patient populations, performance metrics, security accreditations, potential risks and ethical considerations.
Dr. Anderson told Becker's these model cards are designed for use by health systems during procurement processes and for EHR vendors who must comply with ONC Health IT Certification Program requirements.
"Health systems need to know how AI models are built, what data they use, and whether they have any biases that could negatively affect patient care," Dr. Anderson said. "Model cards are meant to standardize that information and help hospitals make more informed decisions."
CHAI brought together AI vendors and healthcare organizations to create a standardized format for these model cards, which outline a model's inputs, limitations and independent evaluations.
According to Dr. Anderson, this initiative aims to strike a balance between protecting intellectual property for vendors, and providing health systems with enough information to ensure that AI models are safe, effective and free from unjustified biases.
Focus on transparency
CHAI also introduced the Assurance Lab Certification Framework, which will create a network of independent labs to evaluate and certify AI models used in healthcare. According to Dr. Anderson, this framework is designed to build trust in AI by ensuring that models are independently tested for safety, accuracy and fairness.
"The healthcare AI space is still like the Wild West," he said. "There's a lot of opaqueness in how these models actually perform. Our assurance lab certification framework is about creating trusted, conflict-free evaluations that hospitals can rely on."
One of the key components of the framework is ensuring that the labs remain free from conflicts of interest. The labs will be audited against ISO benchmarks to ensure their evaluations are objective. This evaluation process will be crucial in helping hospitals avoid biased AI models that could harm patients, Dr. Anderson said.
The initiative also aims to create a network of labs capable of testing AI models against datasets that are representative of diverse patient populations.
"This network is a powerful example of how we can ensure equity in AI deployment," Dr. Anderson said. "Hospitals serving smaller, underserved populations will be able to trust that the AI tools they use are evaluated on datasets that reflect their specific patient populations."
What this means for hospitals
For hospitals, these initiatives represent a leap forward in adopting AI responsibly and safely, Dr. Anderson said. As hospitals are increasingly relying on external partners to help validate and deploy new technologies, CHAI's model cards and assurance lab certification framework aim to provide tools to assist organizations in navigating AI implementation with confidence.
"We're excited about what the future holds," Dr. Anderson said. "These initiatives are just the beginning, but they are a critical step toward building a future where AI can truly transform healthcare for the better."
CHAI is actively seeking feedback on the certification process and model card templates from stakeholders, including patient advocates, under-resourced health systems and startups, according to Dr. Anderson.
CHAI has members such as Rochester, Minn.-based Mayo Clinic, Microsoft and Baltimore-based Johns Hopkins.