Written and maintained by CASRAI Editorial Board
Last updated
The Health AI Partnership (HAIP) is a multi-institutional healthcare collaborative established to build consensus, practice-based standards for how healthcare delivery organizations (HDOs) procure, locally validate, integrate, and monitor clinical artificial intelligence software. Spearheaded by the Duke Institute for Health Innovation (DIHI), Mayo Clinic, UC Berkeley, and DLA Piper alongside health systems such as Kaiser Permanente, UCSF Health, and safety-net hospital networks, HAIP focuses on the operational gap that remains after an algorithm receives regulatory clearance or is published in a peer-reviewed journal: how an individual hospital or clinic network determines whether a model is safe, effective, and equitable for its own local patient population and clinical workflow.
While national policy consortia such as the Coalition for Health AI (CHAI) define high-level assurance blueprints and model cards, HAIP translates clinical AI oversight into an operational playbook structured around four lifecycle phases and eight key decision points, reinforced by the HEAAL (Health Equity Across the AI Lifecycle) evaluation methodology.
What Is the Health AI Partnership (HAIP) and Who Runs It?
Healthcare delivery organizations face a recurring governance challenge when adopting predictive analytics and clinical decision support software: a model trained at one academic medical center or cleared via an FDA 510(k) submission can suffer substantial calibration drift when deployed at a community hospital with different Electronic Health Record (EHR) documentation habits, laboratory assay platforms, and patient demographics.
HAIP was formed to synthesize real-world implementation lessons across diverse health systems—ranging from large integrated delivery networks and academic medical centers to federally qualified health centers (FQHCs) and community safety-net hospitals. Rather than acting as a certifying body or software vendor, HAIP publishes open-access decision guides, interdisciplinary role matrices, and equity audit protocols designed for hospital AI governance committees, Chief Medical Information Officers (CMIOs), nursing leadership, and health services researchers.
The 8 Key Decision Points of the HAIP Clinical AI Lifecycle
The core architecture of the Health AI Partnership framework divides the lifecycle of a clinical AI product into four sequential phases encompassing eight key decision points. At each decision point, an institutional governance committee must review documented evidence before authorizing the project to advance to the next stage:
| Lifecycle Phase | HAIP Key Decision Point | Core Governance Question & Informatics Deliverable |
|---|---|---|
| Phase 1: Procurement & Problem Identification | 1. Identify and prioritize a problem | What specific clinical or operational gap exists, who is affected, and is an AI model actually better suited than a deterministic EHR workflow fix? |
| Phase 1: Procurement & Problem Identification | 2. Identify requirements for an AI product | What technical (FHIR/HL7 inputs, latency), regulatory (FDA CDS vs. SaMD), and clinical specifications must a build-or-buy candidate satisfy? |
| Phase 2: Development & Local Adaptation | 3. Develop measures of outcomes and success | Which primary clinical outcomes, process metrics, and financial/operational key performance indicators (KPIs) define success vs. failure? |
| Phase 2: Development & Local Adaptation | 4. Design a new optimal workflow | Who receives the model output, when in the clinical encounter does it fire, what action is expected, and how is alert fatigue prevented? |
| Phase 2: Development & Local Adaptation | 5. Evaluate safety, effectiveness, and equity concerns | Does silent local shadow-mode validation confirm acceptable discrimination, calibration, and subgroup equity on local hospital data? |
| Phase 3: Clinical Integration | 6. Execute AI product rollout | How is the tool phased into live clinical units (e.g., pilot unit rollout, DECIDE-AI early-stage evaluation, clinician training, and escalation paths)? |
| Phase 4: Lifecycle Management | 7. Monitor the AI solution | What automated telemetry tracks input feature drift, label shift (such as ICD-10 or lab assay changes), override rates, and safety incidents? |
| Phase 4: Lifecycle Management | 8. Update or decommission the AI solution | What pre-specified performance degradation thresholds trigger model recalibration, vendor version re-validation, or complete clinical retirement? |
HEAAL: Embedding Health Equity Across the AI Lifecycle
A signature contribution of the Health AI Partnership is the HEAAL (Health Equity Across the AI Lifecycle) framework, developed through multi-stakeholder consensus to prevent clinical algorithms from exacerbating racial, ethnic, socioeconomic, or geographic healthcare disparities. Rather than treating algorithmic fairness as a narrow mathematical check performed only once during model training, HEAAL maps concrete equity audits across all eight HAIP decision points:
- Problem Selection and Proxy Label Auditing (Decision Points 1–3): Auditing whether the target prediction label uses biased historical proxies—such as predicting future healthcare monetary cost as a proxy for medical illness severity, which systematically under-identifies underserved patients who face barriers to care access.
- Resource Allocation and Workflow Capacity (Decision Point 4): Verifying that downstream clinical interventions triggered by an AI flag (such as care-management nurse outreach or specialist referrals) are adequately staffed at safety-net clinics so high-risk alerts actually translate into patient care.
- Subgroup Stratified Local Validation (Decision Point 5): Disaggregating sensitivity, specificity, positive predictive value (PPV), and calibration curves across race, ethnicity, preferred language, age, sex, and insurance payer status using local retrospective and silent prospective EHR cohorts prior to go-live.
How HAIP Complements FDA SaMD, ONC HTI-1, and CHAI Standards
How does the Health AI Partnership fit alongside federal regulatory rules and other AI assurance initiatives? In health system compliance architectures, these frameworks operate at complementary layers (see Foundation Models vs. FDA’s SaMD Framework):
- Federal Regulatory Baseline (FDA & ONC HTI-1): The FDA regulates medical device manufacturers marketing SaMD or Device CDS products, while the ONC Health Data, Technology, and Interoperability (HTI-1) final rule (§ 170.315(b)(11)) requires certified EHR developers to provide source-attribute transparency and intervention risk management (FAVES: Fair, Appropriate, Valid, Effective, and Safe) for predictive Decision Support Interventions (DSIs).
- Institutional Operational Execution (HAIP): Because neither FDA 510(k) clearance nor ONC EHR certification can guarantee how a local hospital configures order sets, trains nursing staff, or monitors local lab code changes over time, hospital AI governance committees use HAIP’s 8 Key Decision Points as their internal standard operating procedure (SOP) for local clinical deployment and lifecycle maintenance.
Frequently Asked Questions
What is the Health AI Partnership (HAIP) and how is it used in clinical research and healthcare?
The Health AI Partnership (HAIP) is a multi-institutional collaborative initiative founded by healthcare delivery organizations—including Duke Health (DIHI), Mayo Clinic, Kaiser Permanente, and UCSF—alongside UC Berkeley and DLA Piper. Healthcare delivery organizations use HAIP’s practice-based governance framework, structured around 4 lifecycle phases and 8 key decision points, to procure, locally validate, integrate, monitor, and decommission clinical AI software safely and equitably.
What standards or regulatory requirements govern the Health AI Partnership?
HAIP is a voluntary, practice-based governance consortium rather than a federal regulator. Its 8 key decision points and Health Equity Across the AI Lifecycle (HEAAL) framework complement FDA Software as a Medical Device (SaMD) and Clinical Decision Support (CDS) guidance, ONC HTI-1 predictive decision support intervention (DSI) transparency rules, and Coalition for Health AI (CHAI) assurance blueprints.
How does the Health AI Partnership framework integrate with EHR and health informatics workflows?
Health informatics teams embed HAIP’s 8 decision points into institutional AI oversight committees: running silent shadow-mode evaluations inside the EHR before live activation (Decision Point 5), co-designing clinician notification and order-set handoffs (Decision Point 4), and configuring automated telemetry pipelines to detect post-deployment dataset shift or subgroup performance degradation (Decision Point 7).
