NIH PRIMED-AI Model-to-Clinic Grant 2027: Up to $1 Million a Year for Clinical AI Tools

NIH PRIMED-AI Model-to-Clinic Grant 2027: Up to $1 Million a Year for Clinical AI Tools

NIH Common Fund’s Model-to-Clinic (M2C) opportunity is for teams that have moved beyond an interesting healthcare AI prototype and are ready to test whether an image-centered clinical decision-support tool can work in real care settings. The call, RFA-RM-27-013, is part of the new Precision Medicine with AI: Integrating Imaging with Multimodal Data (PRIMED-AI) program. It supports tools built as Software as a Medical Device (SaMD) that combine clinical imaging with at least one other relevant health-data type, then move through technical validation, clinical implementation, and evidence generation for adoption.

This is a substantial, phased NIH cooperative agreement rather than a prize for an early concept. The UG3 phase can receive up to $450,000 in direct costs per year, and teams that meet NIH-approved milestones may transition to a UH3 phase with up to $1,000,000 in direct costs per year. The combined project period can last no more than five years. NIH expects approximately six to eight awards, subject to appropriations and the quality of applications. The first submission date is September 19, 2026; applications are due October 19, 2026 at 5:00 p.m. local time of the applicant organization. NIH lists July 2027 as the earliest start date.

The opportunity is unusually specific about what “clinical AI” must mean in practice. A strong proposal needs more than an accurate retrospective model. It needs credible access to clinical-grade data, a clinically meaningful problem, a plan for users and workflows, rigorous validation, monitoring after deployment, and a multidisciplinary team capable of managing regulatory and implementation work. This guide summarizes the published call as checked on July 29, 2026. Read the full official RFA-RM-27-013 announcement before committing resources or submitting.

Key details

What NIH is trying to fund

PRIMED-AI is aimed at clinical decision-support, or CDS, tools that help care teams make a decision related to patient care within a clinical workflow. The M2C call is not asking applicants to build a general-purpose medical model or a standalone research analysis. Its central output is an AI-enabled, image-centered, multimodal CDS tool developed as SaMD and directed at a defined, important unmet clinical need.

“Image-centered” is a hard program feature. Clinical imaging must be the anchor data type, while one or more complementary modalities add clinical value. The RFA gives examples including laboratory tests, omics data, digital-sensor data, and other clinical information. It recognizes radiologic, ophthalmologic, endoscopic, dermatologic, and video imaging used in patient care. A project may work with several imaging types, but microscopy-based ex vivo digital pathology cannot be the sole imaging data type. A team should therefore be able to explain both why the selected imaging is central and why the additional data make a measurable difference to the proposed decision.

The award’s practical focus is translation. NIH describes the desired path as taking a technically validated prototype toward clinical applications with potential to improve patient outcomes or healthcare processes. That means the application must identify the intended users, the care setting, the current standard of care, the decision the tool will inform, and what a usable deployment would look like. A proposal that calls an algorithm “clinical” without a workflow, implementation partner, or adoption route is poorly aligned.

Who is a good fit

The best fit is an established multi-institutional team with a promising, image-anchored multimodal CDS tool and a credible route to real-world deployment. Teams may include academic medical centers, hospitals, nonprofit organizations, companies, and other eligible organizations. NIH permits domestic and foreign organizations to apply and allows foreign components, but it says applications from foreign organizations require strong justification and encourages U.S. collaborators. Foreign subawards or subcontracts are not allowed, although unfunded foreign collaborations, foreign consultants, and procurement of unique foreign equipment or supplies are not automatically excluded.

This call fits teams that already have a defined clinical use case and enough prior work to justify the scientific and clinical rationale. The program does not require a clinical trial, but a team should not treat “clinical trial optional” as “clinical evidence optional.” The phased program expects a route from rigorous technical validation to clinical implementation and validation. If the tool is still mainly a laboratory demonstration, if access to clinical data is only hoped for, or if the intended clinical user has not been identified, a team may be better served by a less translational funding mechanism.

NIH requires an integrated multidisciplinary structure. The RFA names medical-device/SaMD development, AI and multimodal-data science, relevant clinical-domain knowledge, regulatory science, and healthcare-system workflow implementation as required areas of expertise. It also encourages expertise in user acceptance and human-AI interaction where adoption is central, along with patient representatives or advocates for patient-centered design. Multiple PDs/PIs are permitted. Regardless of mechanism, the team must show how it will operate as one accountable project rather than a collection of disconnected specialists.

Eligibility and the first checks to make

There is no individual citizenship restriction stated for the lead applicant in the RFA; applications are submitted by eligible organizations. NIH says any individual with the skills, knowledge, and resources to lead the work may work with their organization to develop an application. An organization can submit more than one application only when each is scientifically distinct. NIH will not accept duplicate or highly overlapping applications that are under review at the same time.

Start with organization readiness. Required registrations must be complete before submission, and NIH warns that registration can take six weeks or longer. Depending on the applicant organization and its situation, this includes a Unique Entity Identifier, active SAM registration, eRA Commons, Grants.gov, and an NCAGE code where applicable. Every PD/PI needs a valid eRA Commons ID entered in the Credential field; without it, an electronic NIH submission cannot succeed. Treat these as operational requirements, not cleanup items for the final week.

Next, test the project itself against the program’s core definition. The proposed tool should integrate clinical imaging with at least one other modality, address a significant unmet clinical need, and be intended for clinical decision support. The team must be able to obtain and manage clinical-grade imaging and multimodal datasets through development, technical validation, and clinical validation. NIH asks for verifiable evidence of access, not a vague assertion that a partner could provide data later.

Finally, map international participation carefully. An eligible foreign organization may apply, and a U.S. applicant may include a foreign component, but a monetary foreign subaward or subcontract would make an application noncompliant. This is a detail for the institutional grants office and project finance lead to resolve early. Do not assume that a scientifically valuable collaborator can be budgeted in any structure.

How the UG3 and UH3 phases work

The UG3 phase is the exploratory and developmental stage and may last up to two years. It establishes the technical and operational foundation for clinical translation: refining models, preparing for regulatory work, building implementation infrastructure and deployment plans, and showing preliminary feasibility in the proposed setting. NIH requires milestones, and the team cannot simply roll into the advanced phase on a calendar date. Transition to UH3 depends on achieving specified milestones and NIH approval.

The UH3 phase is the advanced-development stage. It emphasizes putting the CDS tool into clinical workflows, conducting prospective clinical studies, evaluating clinical performance, showing feasibility with existing health-information systems, and producing evidence relevant to broader adoption and sustained use. The total UG3/UH3 period cannot exceed five years. Funding ceilings are not a promise that every successful proposal receives the maximum: the requested budget must still be appropriate and justified for the scope.

Build the phase boundary around an observable decision point. For example, the UG3 plan should identify what technical, data-quality, implementation, and safety or reliability evidence would make the tool ready to progress. The UH3 plan should then explain what real-world clinical performance, workflow outcome, or patient-care evidence will be collected. A milestone that says only “complete validation” leaves the key standard undefined; a milestone linked to the intended use, dataset, performance measurement, and implementation site is more credible.

What the application must demonstrate

The RFA organizes the research plan around five pillars. The first four belong within the 12-page application limit, while the fifth is a required separate attachment. This structure should drive proposal planning from the beginning.

First, the application needs a data-quality and governance strategy. NIH requires verifiable access to comprehensive, large-scale clinical-grade imaging and multimodal datasets and a Data Risk Assessment and Mitigation Plan. The team must address access, quality, sharing, provenance, curation, harmonization, and fitness for the intended use. Do not present a generic data-sharing policy in place of an account of the actual datasets, permissions, gaps, and fallback plans.

Second, describe the clinically grounded AI technology. The central tool must be image-centered and multimodal. Explain the clinical problem, current standard of care, intended users, clinical environment, and how the CDS output will be integrated into the care pathway. The RFA expects a suitable advanced analytical framework and a reason that the proposed fusion of modalities is necessary for the clinical task.

Third, provide the route to implementation and adoption. NIH requires a plan for clinical adoption and dissemination, including proactive regulatory engagement, and a post-deployment system to monitor model performance and reliability. A development plan that ends at model accuracy does not answer this pillar. The applicant must confront deployment conditions, changing data, user interaction, and whether the tool can remain trustworthy in practice.

Fourth, present the integrated team and leadership plan. Explain decision rights, collaboration structure, publication and dissemination policies, governance, intellectual-property arrangements where relevant, and a conflict-resolution process. NIH asks for an organizational chart and expects the leadership plan to show that the joint work can be administered as a single project. Fifth, attach the required Error Mitigation and Technical Management Plan, using NIH’s specified filename, and keep it within the three-page limit stated in the RFA.

A practical preparation plan

Begin with a short internal fit memo rather than drafting prose. State the clinical decision, patient population, imaging anchor, second modality, proposed users, clinical site or sites, evidence already available, and the decision that would trigger UG3-to-UH3 transition. If any answer depends on an unconfirmed partner, dataset, or workflow owner, resolve that before writing around it.

Then bring the clinical lead, data steward, AI lead, implementation lead, regulatory adviser, and grants office together. The purpose is not merely to collect biographies. It is to make sure the model design, data agreements, implementation architecture, human-subjects plan, budget, and governance all describe the same project. Inconsistency is especially damaging here because reviewers explicitly assess data access, clinical feasibility, workflow integration, and team resources.

Build an evidence register. For each major claim, identify the supporting artifact: data-access documentation, a partner letter, preliminary validation result, workflow map, letters from potential adopters, or a documented governance arrangement. NIH says support letters from healthcare systems, industry partners, or technology-transfer offices are not mandatory, but they can strengthen the case for clinical applicability, licensing, adoption, collaboration, or dissemination. Ask for specific letters that confirm a real contribution rather than generic endorsements.

Plan submission early. NIH accepts electronic submissions through NIH ASSIST, Grants.gov Workspace, or an institutional system-to-system solution, with tracking through eRA Commons. Applications received after 5:00 p.m. local time on October 19, 2026 are late, and the RFA says late applications will not be accepted. Build in time for organizational review and correction of system errors. Paper applications are not accepted.

How reviewers will judge the proposal

Reviewers will score the application’s likely overall impact through the NIH peer-review system. The RFA highlights three scored areas: importance of the research, rigor and feasibility, and expertise and resources. It then gives program-specific prompts that are particularly useful as a drafting checklist.

On importance, reviewers will assess whether the multimodal integration genuinely exceeds the limits of a single-modality solution and could produce a new level of diagnostic precision, prognostic accuracy, or therapeutic personalization. They will also look for a clear value proposition and an important unmet need in real settings. The lesson is to compare the proposed clinical decision against the actual limitation of current practice and explain why each modality changes that decision.

On rigor and feasibility, reviewers will examine experimental design, controls, sample size, analysis, reporting, generalizability, and the appropriateness of the study population. For this program they also assess whether the team can harmonize datasets with different formats, spatial or temporal resolutions, and quality; whether technical validation measures the contribution of each modality; whether data access is durable; and whether clinical workflow integration is realistic across the intended sites. A performance metric without a comparator or a deployment plan without an operational owner will leave a gap.

On expertise and environment, reviewers consider the investigators’ experience and institutional resources. For multiple-PI teams, the leadership plan is part of the evidence. Make responsibilities visible: who owns model development, data governance, clinical implementation, regulatory engagement, monitoring, evaluation, and final decisions. Good qualifications are more persuasive when they connect directly to a named workstream and an enforceable coordination process.

Common mistakes to avoid

Do not treat this as a generic health-AI grant. A model trained on images alone is not responsive to the central multimodal requirement. Likewise, a proposal that uses multiple data types but fails to anchor the work in clinical imaging is a poor fit. The application should make the imaging anchor and the contribution of the additional data visible from the specific aims onward.

Avoid describing datasets as if access and quality were automatic. NIH wants verifiable clinical-grade data access and a risk-and-mitigation plan. Be candid about data heterogeneity, missingness, provenance, permissions, throughput, and the transition from retrospective work to clinical validation. Unsupported claims of “large-scale data” are unlikely to satisfy this requirement.

Do not postpone clinical adoption until the end. The RFA requires an end-user design and clinical integration plan, co-design across both phases, performance monitoring, and a sustainable path. Engage clinicians, health-system implementation staff, and prospective users while defining the technical approach. If human-subjects research or a clinical trial is proposed, follow the specific NIH forms and protections rather than assuming the technical plan covers them.

Finally, do not ignore small compliance instructions. The first four pillars must fit within the 12-page research-plan limit; the error-mitigation attachment has its own three-page limit and required filename; all registrations must be active; and all PDs/PIs need eRA Commons credentials. NIH states that incomplete, noncompliant, or nonresponsive applications will not be reviewed.

Frequently asked questions

Is this only for U.S. applicants?

No. The RFA says foreign organizations are eligible, non-U.S. components of U.S. organizations are eligible, and foreign components are allowed. However, foreign organizations require strong justification and NIH encourages U.S. collaborators. Budgeting foreign subawards or subcontracts is not allowed, so international collaboration must be structured with care.

Is a clinical trial required?

No. The notice is clinical-trial optional. Still, the phased design expects rigorous movement toward clinical implementation and validation. If a team proposes a trial, it must meet the applicable NIH human-subjects and clinical-trial instructions.

Can a team submit both M2C and the companion Data-to-Model call?

NIH’s PRIMED-AI FAQ says a team may apply to both only if the applications are scientifically distinct. NIH generally does not accept duplicate or highly overlapping applications under review at the same time. Document the difference in aims, datasets, technical work, and intended endpoint before considering parallel submissions.

Where can applicants verify changes or ask questions?

Use the official RFA-RM-27-013 announcement as the controlling source for instructions, contacts, and policy notices. NIH Common Fund also maintains a PRIMED-AI funding opportunities page and a program FAQ. Confirm the application package, organizational registrations, and any later NIH notices before submission.

The immediate next step is a fit review with your sponsored-research office and clinical implementation partners. If the team can show a real multimodal clinical decision-support tool, durable data access, a patient- and user-informed deployment plan, and measurable UG3 transition milestones, this is a timely 2026 call with a potential July 2027 start.