2026 executive update · HHS LEAP AI behavioral health projects · Leadership action
HHS Announces LEAP in Health Special Emphasis Notice for AI Behavioral Health Projects
The 2024 Leading Edge Acceleration Projects in Health Information Technology Special Emphasis Notice offered healthcare leaders more than a grant announcement. It identified two connected infrastructure problems: artificial intelligence cannot…
At a Glance
The eventual awards made the strategy tangible. Columbia University began work on scalable methods for evaluating nursing data used by AI, while Oregon Health & Science University pursued an open source, FHIR based behavioral health care planning application for stand alone clinics. For health systems, the…
Executive perspective
The 2024 Leading Edge Acceleration Projects in Health Information Technology Special Emphasis Notice offered healthcare leaders more than a grant announcement. It identified two connected infrastructure problems: artificial intelligence cannot be trusted when its training and validation data do not reflect clinical reality, and behavioral health cannot participate fully in coordinated care when its technology remains isolated. HHS, through the Office of the National Coordinator for Health Information Technology, invited projects focused on improving healthcare-data quality for responsible AI and accelerating health IT adoption in behavioral health settings.
The eventual awards made the strategy tangible. Columbia University began work on scalable methods for evaluating nursing data used by AI, while Oregon Health & Science University pursued an open-source, FHIR-based behavioral health care-planning application for stand-alone clinics. For health systems, the executive opportunity is not to wait for a finished federal product. It is to become a disciplined research partner: contribute real operational problems, protect patients, test solutions in representative workflows, and turn findings into reusable capabilities.
That requires governance beyond a conventional innovation pilot. A strong partnership joins clinical, behavioral health, nursing, informatics, research, legal, privacy, security, community, and patient expertise. It defines what evidence would justify adoption before enthusiasm or grant timelines create pressure to scale.
Leadership priorities
Build an integrated leadership response
Select a Problem Worth Solving Together
Begin with a care problem, not a technology. Suitable projects may include incomplete behavioral health care plans at transitions, missing social context in risk models, inconsistent nursing documentation, or poor exchange between specialty clinics and hospitals. Quantify the current harm through delayed decisions, duplicate assessment, manual reconciliation, avoidable utilization, inequitable performance, or clinician burden.
Choose a question that requires complementary strengths. A health system can offer workflow knowledge, representative data, implementation settings, and clinical accountability. An academic or technology partner may contribute methods, standards expertise, evaluation design, and open-source engineering. Community behavioral health organizations bring realities that hospital-centric programs routinely miss, including resource constraints, consent complexity, and limited interface capacity.
Write a one-page partnership thesis before discussing a tool. State the population, decision, workflow, baseline, intended benefit, principal risks, and what result would change practice. Identify who benefits if the project succeeds and who could be burdened. A narrow, consequential question produces more transferable evidence than an expansive demonstration with no accountable decision.
Assess partnership readiness with the same rigor as scientific merit. Confirm that every site has an executive sponsor, operational lead, data steward, clinical champion, implementation capacity, and protected time. Document competing projects and technology changes that could distort results. Smaller partners may need funding for interfaces, backfill, patient engagement, or project management. Budgeting for those needs strengthens the evidence by preventing resource imbalance from becoming an unacknowledged study variable. Agree on how conflicts will be resolved and how participants can raise concerns outside the normal reporting line. A collaboration is credible when each organization can influence the design, not merely supply data or patients.
Build a Trustworthy Data and Evidence Compact
Create a data compact that covers provenance, definitions, quality thresholds, permitted uses, retention, access, linkage, security, publication, and disposition at project end. Inventory the data actually generated in care, including notes, structured fields, assessments, orders, patient-reported information, and operational context. Do not assume a populated field is clinically meaningful or comparable across sites.
For AI work, examine missingness, label construction, temporal leakage, subgroup representation, changes in documentation practice, and the distance between development data and intended use. Include nurses and other frontline professionals who understand why a record looks the way it does. Their knowledge can distinguish true absence from workflow-driven absence and reveal shortcuts that an algorithm might mistakenly treat as clinical signals.
Define evidence in advance. Use technical performance only as one layer. Add workflow fit, human factors, subgroup performance, override patterns, safety events, patient experience, and resource requirements. Predefine stop rules for harmful performance or operational instability. Independent validation and transparent limitations protect the partnership from converting preliminary findings into premature claims.
Design Behavioral Health Interoperability Around Consent and Use
Behavioral health exchange succeeds when information is usable, appropriately governed, and available at a real decision point. Map the minimum data needed for each use case rather than moving every available record. Care goals, medications, allergies, crisis preferences, assessments, responsible team members, and follow-up commitments may matter differently during emergency care, referral, or longitudinal planning.
Design consent and segmentation with privacy, legal, clinical, and patient input. Federal and state requirements can differ, and sensitive information deserves careful handling. Explain to patients what will be shared, with whom, for what purpose, and how corrections are managed. Test whether staff can execute the process under time pressure without creating a new access barrier.
Use standards as a shared language, not an end state. FHIR interfaces, structured questionnaires, and write-back capabilities must work across the EHRs and supplemental stores used by participating organizations. Test failure modes such as duplicate records, unmatched patients, stale plans, unavailable endpoints, and partial data. The practical outcome is a reliable closed loop: information arrives, is understood, changes appropriate action, and generates a visible response.
Govern the Partnership as a Learning System
Establish a joint steering group with one accountable executive and voting representation from participating clinical settings, research, technology, compliance, and patients or community members. Define decision rights for protocol changes, model releases, safety escalation, publication, intellectual property, and continuation after grant funding. A partnership agreement should also address whether resulting methods, code, or implementation materials will be shared.
Separate research oversight from operational authorization. Institutional review board review, privacy review, security assessment, contracting, and clinical change control answer different questions. Build one integrated calendar so teams can resolve issues early without treating safeguards as last-minute approvals. Maintain a living risk register that includes privacy, bias, automation, workflow, vendor, reputational, and sustainability risks.
Use regular learning reviews instead of status presentations. Compare results with the baseline, inspect subgroup and site variation, review near misses, and ask what assumption failed. Document adaptations so other organizations can judge transferability. When findings are negative, publish the lesson when possible. A clear account of what did not work can be more valuable than a polished demonstration that conceals implementation conditions.
Convert a Grant Project Into Durable Capability
Plan sustainability before the first build. Identify who will own software, interfaces, terminology, monitoring, training, support, patient communication, and cybersecurity after the project period. Estimate recurring costs and workload, not only development expense. If a smaller behavioral health partner cannot support the design without special grant staffing, the solution is not yet scalable.
Create a translation package for each validated component: use case, eligibility criteria, workflow map, data specification, configuration, test scripts, training, safety controls, evaluation results, limitations, and implementation cost. Reusable materials help the system adopt evidence consistently and support dissemination to outside organizations. Open-source components still require governance, maintenance, and local validation.
Set a scale decision with three options: expand, redesign, or stop. Expansion should require predefined clinical, operational, equity, and safety thresholds across representative settings. Redesign is appropriate when the need remains valid but evidence reveals correctable barriers. Stopping is responsible when benefit is weak, risk is unacceptable, or sustainability is unrealistic. The capability gained is disciplined learning, not perpetual piloting.
Leadership cadence
Start, strengthen, and measure the system in 90 days.
Phase 1, days 1 to 30
Form the joint steering group, select one decision-grade use case, document the baseline, and complete a stakeholder map. Draft the partnership thesis, data compact, consent approach, intended-use statement, and preliminary risk register. Include at least one resource-constrained behavioral health setting and patient or community representation.
Phase 2, days 31 to 60
Map the end-to-end workflow and data lineage, profile quality and subgroup representation, and define technical, clinical, operational, equity, and safety measures. Complete privacy, security, research, and contracting pathways. Prototype the smallest viable workflow with synthetic or appropriately governed test data before exposing real care processes.
Phase 3, days 61 to 90
Run a limited simulation or silent-mode evaluation, review failures with frontline teams, and test consent, downtime, and escalation procedures. Approve a bounded pilot only if entry criteria are met. Present leaders with cost, evidence gaps, risk controls, publication commitments, and explicit expand, redesign, or stop thresholds.
Decision-grade measurement
Decision-Grade Metrics
- Data completeness, validity, timeliness, provenance, and missingness by site and population
- Model or rules performance, calibration, subgroup variation, false alerts, and clinician overrides
- Behavioral health records successfully matched, exchanged, integrated, and used at the intended decision
- Consent completion, patient questions, access barriers, privacy incidents, and inappropriate disclosure events
- Workflow time, duplicate work, adoption, support demand, and staff experience
- Clinical process reliability, follow-up completion, safety signals, and patient-reported experience
- Recurring operating cost, partner burden, implementation time, and reuse of project assets
- Milestones delivered, adaptations documented, findings disseminated, and scale decisions completed
SEO
SEO title: HHS LEAP AI Behavioral Health Projects: Executive Guide
Meta description: An executive guide to HHS LEAP AI behavioral health projects, covering research partnerships, data quality, interoperability, governance, and scale.
Focus keyphrase: HHS LEAP AI behavioral health projects
Conclusion
Turn strategy into an accountable operating system.
The LEAP notice connected responsible AI with behavioral health interoperability because both depend on trustworthy data, inclusive workflow design, and cross-organizational learning. Health systems can accelerate that work by acting as rigorous research partners rather than passive technology customers.
The durable return is not one model or application. It is a repeatable way to select meaningful problems, govern data, test in real care, protect patients, share evidence, and sustain what works. That capability can turn a time-limited federal opportunity into lasting improvement across the care continuum.
Executive questions
Frequently Asked Questions
1. What was the central purpose of the 2024 LEAP special emphasis notice?
The notice sought projects in two areas: improving healthcare-data quality to support responsible AI development and accelerating health IT adoption in behavioral health settings. The two priorities were distinct, but both addressed barriers to interoperable, evidence-based care.
2. Does a health system need to be a grant recipient to apply the lessons?
No. An organization can use the same partnership disciplines in internally funded research, vendor pilots, academic collaborations, and quality-improvement programs. The essential practices are a defined use case, representative partners, governed data, prospective evidence standards, and a sustainable operating owner.
3. How should leaders distinguish research from operational improvement?
The determination depends on purpose, design, risk, and applicable policy, not the label used by the project team. Engage the institutional review board, compliance, privacy, and operational leaders early so the appropriate oversight pathway is documented before data access or testing begins.
4. Why include stand-alone behavioral health organizations in design?
They often face technology, staffing, consent, and exchange constraints that differ from those of large integrated systems. Their participation helps prevent a solution from working only in well-resourced environments and improves the relevance of standards, workflows, support, and cost assumptions.
5. What evidence should be required before an AI-enabled workflow scales?
Require performance in the intended population and setting, acceptable subgroup results, reliable human oversight, demonstrated workflow fit, manageable alert and support burden, patient protections, cybersecurity controls, and evidence that the change improves a meaningful clinical or operational outcome without unacceptable balancing harm.
