Coalition for Health AI (CHAI): Blueprint for Trustworthy AI, Applied Model Cards, and HAIP Governance

The Coalition for Health AI (CHAI) (chai.org) is a multi-stakeholder non-profit standards and assurance organization founded by academic medical centers, integrated health systems, technology developers, and patient advocacy organizations—with federal participation and observation from agencies including the FDA, ASTP/ONC, NIH, and CMS—to establish consensus guardrails for responsible artificial intelligence in healthcare. Alongside practice-based implementation networks such as the Health AI Partnership (HAIP), CHAI provides the technical frameworks, Applied Model Cards, and independent Assurance Labs that clinical informaticists use to evaluate predictive and generative AI before bedside deployment.

What Is the Coalition for Health AI (CHAI) and How Does Its Blueprint for Trustworthy AI Work?

Clinical artificial intelligence models—ranging from inpatient sepsis early-warning scores and radiology artificial intelligence triage algorithms to ambient clinical documentation models—can degrade silently when moved across hospital populations with different EHR coding practices, demographics, or imaging scanners. To bridge the gap between high-level management standards (ISO/IEC 42001 and the NIST AI RMF vs. ISO/IEC 42001) and health-system clinical governance, CHAI released its foundational Blueprint for Trustworthy AI Implementation Guidance and Assurance for Healthcare.

The CHAI Blueprint evaluates clinical and operational healthcare AI systems across seven core assurance pillars across the entire model lifecycle (problem formulation, design, local validation, deployment, and continuous monitoring):

1. Usefulness, Usability, and Efficacy: Verifies that the AI model addresses a well-defined clinical or administrative need, integrates into human clinician workflows without inducing alert fatigue, and demonstrates measurable improvement in clinical outcomes or operational efficiency.
2. Fairness, Equity, and Bias Mitigation: Requires stratified subgroup evaluation across age, sex, race, ethnicity, socioeconomic status, and payer category to ensure the algorithm does not amplify health disparities or perform disproportionately poorly on underrepresented patient groups.
3. Safety and Reliability: Assesses failure modes, out-of-distribution behavior, calibration stability, and clinical harm severity in alignment with AI system impact assessments and medical device risk management principles.
4. Transparency and Explainability: Mandates standardized documentation of training data provenance, model architecture, intended clinical user, and known limitations through the CHAI Applied Model Card.
5. Accountability and Governance: Establishes clear institutional roles, human-in-the-loop oversight authorities, incident reporting channels, and model decommissioning triggers.
6. Security and Resilience: Protects model weights, inference APIs, and EHR pipelines against adversarial attacks, prompt injection, and data poisoning.
7. Privacy: Ensures strict protection of protected health information (PHI), minimum-necessary data ingestion, and compliance with HIPAA and federal research regulations.

CHAI Assurance Labs and the National Health AI Registry

Beyond written guidelines, CHAI operates an operational assurance infrastructure consisting of two institutional mechanisms:

• CHAI Assurance Labs: Independent, accredited testing facilities hosted at academic medical centers and technical evaluation organizations that benchmark health AI models against standardized, multi-institutional reference datasets to verify quantitative performance, calibration, and subgroup fairness claims.
• CHAI National Health AI Registry: A centralized catalog where health systems and developers register validated clinical AI models alongside their standardized Applied Model Cards and post-deployment monitoring summaries.

What Information Is Required on a CHAI Applied Model Card for Clinical AI Transparency?

Modeled on a clinical “nutrition label” and aligned with the ASTP/ONC HTI-1 Final Rule’s transparency requirements for predictive Decision Support Interventions (DSIs, 45 CFR § 170.315(b)(11)), a CHAI Applied Model Card requires AI developers and health system deployers to disclose structured, machine- and clinician-readable metadata across six sections:

CHAI Applied Model Card Section	Required Technical & Clinical Disclosures	Governance & Audit Purpose
1. Model Identity & Intended Use	Model name, version, developer, regulatory clearance status (e.g., FDA 510(k)/De Novo SaMD vs. non-device CDS), target clinical setting, target patient population, and intended clinical user	Prevents off-label clinical use in unvalidated care settings (e.g., applying an adult inpatient model to pediatric ED patients)
2. Out-of-Scope Uses & Warnings	Explicit contraindications, excluded patient cohorts, known failure modes, and whether human clinician verification is mandatory prior to clinical action	Protects patient safety and prevents automation bias
3. Training & Evaluation Data Provenance	Source health systems, geographic/temporal coverage, sample size, demographic breakdown, inclusion/exclusion criteria, and ground-truth outcome label definitions	Exposes spectrum bias and label-leakage risks (such as using healthcare cost as a proxy for clinical illness severity)
4. Technical Architecture & Input Features	Model family (e.g., gradient-boosted trees, transformer LLM, CNN), exact EHR/FHIR input variables (LOINC/SNOMED codes), output format, and inference frequency	Allows hospital IT teams to verify whether local EHR data mappings match the model’s expected feature distributions
5. Quantitative Performance & Subgroup Fairness	Discrimination (AUROC, AUPRC), calibration slope/intercept, sensitivity, specificity, PPV/NPV at the operating threshold, and disaggregated metrics across demographic subgroups	Substantiates both overall clinical accuracy and algorithmic equity
6. Maintenance & Drift Monitoring Plan	Recalibration cadence, input feature drift thresholds, outcome decay triggers, and responsible clinical/technical owners	Ensures continuous post-market surveillance as clinical practice patterns evolve

How the Health AI Partnership (HAIP) Guides Local Clinical AI Validation and Governance

While CHAI defines national consensus standards and model card schemas, the Health AI Partnership (HAIP)—a collaborative learning network established by healthcare delivery organizations, academic institutions (including the Duke Institute for Health Innovation), and clinical safety leaders—provides a step-by-step institutional playbook for local health system adoption and validation. HAIP structures local clinical AI governance around four operational phases:

1. Phase 1 — Problem Definition and Organizational Readiness: Multidisciplinary governance committees (comprising clinicians, biostatisticians, informaticists, ethicists, and patient representatives) verify that the clinical problem warrants an AI intervention and that the institution possesses the workflow capacity to act on model predictions.
2. Phase 2 — Local Retrospective and Silent-Mode Prospective Validation: Before any AI score is displayed to bedside clinicians, HAIP requires the health system to test the model on local patient data—first retrospectively, and then in “silent” (shadow) prospective mode inside the live EHR—to confirm that local calibration, positive predictive value, and demographic subgroup fairness meet clinical safety thresholds.
3. Phase 3 — Workflow Integration and Clinician Training: The governance team defines exact clinical escalation protocols, trains frontline staff on how to interpret the model’s Applied Model Card and confidence bounds, and conducts staged pilot rollouts.
4. Phase 4 — Lifecycle Surveillance and Decommissioning: Automated telemetry tracks input data drift (for example, changes in lab assay codes or ICD-10 documentation) and clinical outcome impact, triggering model recalibration or immediate deactivation if performance degrades.