Coalition for Health AI Unveils Model Card Registry to Standardize AI Transparency in Healthcare

AI model cards, often described as “nutrition labels,” provide structured documentation of how an AI model was trained, tested, and should be used.

The Coalition for Health AI (CHAI) has announced details of its long-awaited model card registry, a centralized repository designed to standardize AI model cards across the healthcare industry.

This initiative aims to provide AI purchasers, particularly health systems, with a transparent view of a model’s training data, fairness metrics, and intended use.

By establishing a standardized format, CHAI seeks to offer an industry-wide “stamp of approval” for AI vendors that accurately fill out and submit their CHAI-compliant model cards.

However, while the registry improves transparency, it does not validate AI model performance. Health systems will still need to conduct independent evaluations to assess how a model performs in their specific clinical setting.

Improving AI Procurement with Standardized Model Cards

AI model cards, often described as “nutrition labels,” provide structured documentation of how an AI model was trained, tested, and should be used.

CHAI’s model card registry aims to streamline AI procurement by offering healthcare providers a clear and uniform way to evaluate AI models beyond sales pitches and marketing materials.

“The goal here is to just make sure that the content reported in the nutrition label is correctly formatted and follows the instructions of the model card,” said Merage Ghane, Director of Responsible AI at CHAI.

Using the registry allows healthcare providers to make more informed purchasing decisions while reducing ambiguity in early conversations with AI vendors.

“[For pilot users] it led to a productive first conversation rather than one that feels like a sales pitch, leaving them with more questions than answers,” Ghane added.

Building a Reliable AI Marketplace

For AI developers, inclusion in the CHAI model card registry offers a credible platform to showcase their models and connect with potential healthcare customers.

The repository is expected to serve as a trusted industry resource for responsible AI vendors.

CHAI has partnered with Avanade, a global tech services company specializing in Microsoft solutions, to develop the model card registry software.

This platform, which will be free and open to use, has been funded through donations from Avanade and CHAI.

CHAI is also building an automated system to review model card submissions for completion and adherence to responsible AI standards, including compliance with HTI-1 regulations.

The review process will be mostly automated, allowing vendors to quickly upload their model cards, while incomplete submissions will be sent back for revision.

As part of troubleshooting the registry, CHAI is exploring the role of human oversight in reviewing flagged submissions. Ghane confirmed that an engineer is working full-time on the model registry to refine the system and determine when human intervention is necessary.

The launch of CHAI’s model card registry has garnered support from leading healthcare institutions and AI vendors, including Cleveland Clinic, Kaiser Permanente, Memorial Sloan Kettering, Mercy, Mount Sinai Health System, Providence, Rush University System for Health, Sharp HealthCare, Stanford Medicine, UMass Memorial, and the University of Texas Health System.

These institutions recognize the importance of AI transparency and believe the registry will streamline AI procurement, ensuring that health systems receive standardized, critical information about AI models upfront.

Creating an Industry Benchmark for AI Model Cards

Beyond individual model submissions, CHAI plans to develop gold-standard sample model cards for different AI use cases to serve as industry benchmarks. Initial efforts have focused on discharge summaries and sepsis detection.

Additionally, the model card includes a checkbox to indicate if a model has been externally validated through independent third-party testing, including potential CHAI-certified AI assurance labs.

CHAI’s broader vision includes establishing AI assurance labs to certify AI models through rigorous, standardized evaluation.

However, the organization has not provided a recent public update on this initiative since October 2024, when it released a draft framework for certification criteria.

At that time, CHAI had identified 32 potential assurance lab candidates and planned to have two operational sites by the end of 2024.

Stay tuned for more such updates on Digital Health News.