1. What is the PRIMED-AI consortium?

The “PRIMED-AI consortium” constitutes members of PRIMED-AI, excluding NIH program staff. The “PRIMED-AI Program” is an umbrella term encompassing the consortium, NIH staff, and overall programmatic objectives.

2. What is an Error Mitigation and Technical Management Plan?

Applicants for the Validation Center, Data-to-Model Academic-Industrial Partnerships, and Model-to-Clinic funding opportunities must submit a required Error Mitigation and Technical Management Plan as a separate attachment. Guidance on what should be included in this document is provided here. Applications that lack this required attachment are incomplete, will not be reviewed, and will be withdrawn. The Error Mitigation and Technical Management Plan may be revised during the award period in accordance with Consortium policies and updated NIH guidance.

3. Can I submit a late application due to my recent peer review service?

No.

4. Can one institution apply to multiple NOFOs?

Yes. Applicant organizations may submit more than one application, provided that each application is scientifically distinct.

5. If awarded, would receipt of any of the PRIMED-AI awards end my early-stage investigator status through the NIH if I am the sole primary PI? What about if I am one of multiple PIs?

Yes, being the PD/PI or being a part of a group of multiple PD/ PIs on any of the PRIMED-AI awards will terminate an investigator’s ESI status. The U24, U54, U01 and UG3/UH3 award mechanisms are considered substantial research grants that would terminate an investigator's ESI status.

6. Are there specific considerations for applications involving the NIH Intramural Research Program?

Title	RFA number	Considerations for Intramural Involvement
Validation Center	RFA-RM-27-014	Extramural only: Intramural investigators are not eligible to participate in these NOFOs.
Logistics Center	RFA-RM-27-015	Extramural only: Intramural investigators are not eligible to participate in these NOFOs.
Development and Testing of a Multi-use Frameworks Playbook	RFA-RM-27-011	Extramural and intramural competition: The requests by NIH intramural scientists will be limited to the incremental costs required for participation. As such, these requests will not include any salary and related fringe benefits for career, career conditional or other Federal employees (civilian or uniformed service) with permanent appointments under existing position ceilings or any costs related to administrative or facilities support (equivalent to Facilities and Administrative or F&A costs). These costs may include salary for staff to be specifically hired under a temporary appointment for the project, consultant costs, equipment, supplies, travel, and other items typically listed under Other Expenses. Applicants should indicate the number of person-months devoted to the project, even if no funds are requested for salary and fringe benefits. Applications from the NIH Intramural Program, either as primary applicants or as collaborators, submitted in response to this NOFO must include a current (i.e. within 2 months of application due date) letter from the Scientific Director of their Division indicating that the intramural scientist will be able to collaborate on the project.
Data-to-Model Academic-Industrial Partnerships	RFA-RM-27-012	Extramural only: U.S. Federal Government Agencies (e.g., NIH Intramural Research Program, DOE National Laboratories) may participate as partners but are not eligible to apply as the primary applicant institution. While their expertise and resources are highly valued, NIH intramural scientists cannot receive salary support or any other direct financial compensation from funds awarded through this extramural NOFO.  Their involvement should be clearly outlined in the application, including a description of their scientific contribution, the number of person months devoted to the project, and a formal letter of collaboration from their Institute/Center Scientific Director or equivalent, confirming their commitment to the project and that no grant funds will be used for their support or the operational costs of NIH intramural facilities.
Model-to-Clinic	RFA-RM-27-013	Extramural only: U.S. Federal Government Agencies (e.g., NIH Intramural Research Program, DOE National Laboratories) may participate as partners but are not eligible to apply as the primary applicant institution. While their expertise and resources are highly valued, NIH intramural scientists cannot receive salary support or any other direct financial compensation from funds awarded through this extramural NOFO.  Their involvement should be clearly outlined in the application, including a description of their scientific contribution, the number of person months devoted to the project, and a formal letter of collaboration from their Institute/Center Scientific Director or equivalent, confirming their commitment to the project and that no grant funds will be used for their support or the operational costs of NIH intramural facilities.

7. Does smartphone-acquired imaging fall within scope as the anchor data type, given the definition of clinical imaging as an FDA-approved imaging modality used in patient care?

High-quality, standardized clinical photographs could serve as the anchor imaging data type, as long as photography is the established/accepted clinical standard for evaluation in this context. In terms of the use of smartphone image acquisition, the PRIMED-AI funding announcements do not specify that the image-acquisition device itself must have FDA clearance. Instead, the emphasis is on the quality and standardization of the data collection. If photography is not the accepted clinical standard, you should consider anchoring on another technology.

8. Can microscopy-based imaging of biospecimens ex vivo or post-mortem like digital pathology represent the primary imaging data type/technique in PRIMED-AI award applications?

No, microscopy-based imaging of biospecimens ex vivo (e.g., digital pathology) cannot represent the primary imaging data type. However, ex vivo and post-mortem imaging can be included as a part of the broader set of multimodal data. On the other hand, integrating digital pathology, molecular, and clinical data around an FDA-approved clinical-imaging modality is responsive. The key thing here is that the model must be anchored on the clinical imaging data, as opposed to the digital pathology data.

9. Can EEG and/or EKG/ECG imaging represent the primary imaging data type/technique in PRIMED-AI award applications?

No, EEG and/or EKG/ECG imaging, even if high density, cannot represent the primary imaging data type.

10. Can data derived from non-human sources be incorporated into PRIMED-AI award applications?

Although non-human imaging and/or MMD data may have assisted in development of an AI-model, overt representation and reliance on data derived from non-human sources for CDS tool development, testing, and validation will be given low programmatic priority.

11. Can non-DICOM standard imaging be incorporated in PRIMED-AI award applications?

Digital Imaging and Communications in Medicine (DICOM) standard, the most widely used by the community to address interoperability challenges, is strongly encouraged but not required. Inclusion of non-DICOM standard clinical imaging must include a plan to develop standards in conjunction with the PRIMED-AI community if none currently exist.

12. The NOFOs emphasize the integration of imaging with multimodal data; however, how broadly is "multimodal" intended to be interpreted? For example: Would the integration of two distinct imaging modalities be considered sufficiently multimodal? Or, is the expectation that applications integrate imaging data with a fundamentally different data type, such as pathology, genomics, laboratory results, EHR data, or other clinical information? More generally, is there a minimum expectation regarding the number or type of modalities that should be included to align with the goals of the program?

The expectation is that applications must integrate imaging data with a fundamentally different data type, such as pathology, genomics, laboratory results, EHR data, or other clinical information. Multiple imaging modalities alone will not suffice as “multimodal,” although they can be a part of the mix of data elements used. For example, MRI images + CT images +X-ray images is not responsive. In contrast, MRI images + CT images + EHR data + genetic analysis is responsive.

13. What project outputs does NIH expect awardees to share with the PRIMED-AI consortium and the broader research community? How should we distinguish shared outputs from proprietary academic or industry components?

The NIH cooperative agreement framework is designed to support translation in a pre-competitive space, allowing commercial partners to protect their proprietary software while fully participating in the network. Some additional important issues around IP are as follows:

• Under the federal Bayh-Dole Act of 1980, any organization receiving federal research funding (including universities and for-profit startups) is legally entitled to retain ownership (e.g., hold the title, patent rights, and copyright to any intellectual property (IP), software code, or trained model weights generated under the grant) as well as license or directly commercialize the technology. See more about NIH IP Policy and Compliance.
• It is strongly encouraged that the MPI Leadership plan addresses communication plans, processes for making decisions on scientific direction, data rights, inventions and intellectual property, and procedures for resolving conflicts.
   

Other IP issues, including pre-existing IP, are addressed in the NOFOs

14. Are external collaborators allowed to work with multiple and/or all of the centers simultaneously?

Yes, external collaborators may be permitted to work with multiple PRIMED-AI awardees. However, this will depend on the specific circumstances, including the collaborator’s role and the nature of the work being performed. Collaborators would need to disclose and demonstrate the absence of conflicts of interest and adhere to appropriate data-sharing and confidentiality safeguards. This may include specific mitigation measures to ensure that benchmark, proprietary, or otherwise restricted information from one awardee is not shared with or used by another awardee (e.g., information from the Validation Center being disclosed to a Data-to-Model or Model-to-Clinic project team).

15. Can you clarify “new” versus “existing” AI models?

An existing single-modality (e.g., images only) model would be considered  a new multi-modal model once it is modified to integrate non-imaging data. Once you include non-imaging data, you may also add other secondary imaging modality data.

16. Will NIH facilitate matching between PRIMED-AI center awardees and external collaborators? Is there a forum to find collaborators?

No, NIH does not set up collaborations or host a matchmaking forum; doing so would be a conflict of interest. Applicants are responsible for independently building their team, including deciding how to use restricted funds for them, whether by leveraging existing working relationships or establishing new collaborations on their own.

1. Can foreign organizations apply?

For the Playbook (RFA-RM-27-011), Validation Center (RFA-RM-27-014), and Logistics Center (RFA-RM-27-015) funding opportunities, foreign organizations cannot apply. Non-domestic (non-U.S.) entities (foreign organizations) are not eligible to apply. Non-domestic (non-U.S.) components of U.S. Organizations are also not eligible to apply. Foreign components are not allowed.

For the Data-to-Model Academic-Industrial Partnerships (RFA-RM-27-012) and Model-to-Clinic (RFA-RM-27-013) funding opportunities, foreign organizations can apply. Non-domestic (non-U.S.) entities (foreign organizations) are eligible to apply. Non-domestic (non-U.S.) components of U.S. Organizations are also eligible to apply. Foreign components are allowed. Foreign subawards are not allowed. 

Please refer to Information for Foreign Grants, NIH Grants Policy Statement, and NOT-OD-25-104 for the most up-to-date information.

2. Are there restrictions on using datasets from other countries?

No, as long as NIH funds are not used for foreign image/data acquisition, there are no restrictions on using contributing datasets from other countries.

3. If a foreign collaborator contributes a proprietary AI model or algorithm as an unfunded/in-kind collaborator, are there any restrictions on that collaborator retaining ownership of the underlying intellectual property while participating in a PRIMED-AI-funded project?

According to NOT-OD-06-005, NIH allows “foreign entities to retain rights in intellectual property developed under NIH funding agreements (grants, cooperative agreements, contracts, subgrants, and subcontracts), consistent with current NIH policies and any other terms and conditions of the specific funding agreement awarded.” 

You may also refer to NOT-OD-09-010 which states: “Regarding intellectual property, NIH allows foreign grantees to have the same rights and obligations regarding invention ownership as U.S. grantees. NIH allows foreign entities to retain rights in intellectual property developed under NIH funding agreements (grants, cooperative agreements, contracts and subcontracts), consistent with current NIH policies and any other terms and conditions of the specific funding agreement awarded. NIH believes that the current policy protects the U.S. public's interest while not unduly compromising productive research and intellectual property relationships. (See NIH Guide Notice NOT-OD-06-005). All foreign grantees, contractors, consortium participants, and/or subcontractors are reminded that they must comply with Bayh-Dole invention reporting requirements. Information on reporting requirements and policy, as well as electronic systems to fulfill reporting requirements, may be found at Interagency Edison.

Please also refer to NIH’s IP Compliance/Policy and be aware of reporting requirements for inventions and patents.

4. If a foreign parent company has a separately-incorporated U.S. subsidiary or affiliate, can that U.S. entity participate as a funded collaborator or subrecipient under the NOFO?

As outlined in the NIH Grants Policy Statement definition of terms, a foreign organization is an organization located in a country other than the United States or its territories. A domestic organization is defined as an organization that is located in the U.S. or its territories.  The NIH does not specifically address multi-national companies with foreign-owned US affiliates, but prospective applicants should consider these definitions when identifying their eligibility. 

NIH will not issue awards to domestic or foreign entities that include a subaward to a foreign entity (NOT-OD-25-104 and NOT-OD-25-130), except under the collaborative international research award structure, as identified in NIH Grants Policy Statement Section 16.8. For multi-national companies, as long as a proposed subaward is executed by the domestic (U.S.-based) entity, and all grant funds, the designated MPI, and contractual work remain within the U.S. branch, this does not trigger NIH foreign subaward restrictions. In this scenario, as you draft your application, we recommend you explicitly specify the US entity, i.e., always refer to the partner by its exact domestic corporate name and list its domestic address. Additionally, if any non-U.S.-based scientists will contribute to the project's scientific development in a way that leads to co-authorship, or if any sequestered data will be processed on servers physically located outside the U.S., you must declare a "Foreign Component" in your application. Under NIH rules, a Foreign Component does not require a foreign subaward, but it does require programmatic disclosure. Note that foreign components are only allowed for the Data-to-Model and Model-to-Clinic NOFOs.

Regardless of whether the prime applicant is a domestic or foreign entity, no NIH funding may go to a foreign subaward except under the PF5/UF5 model, which allows for domestic prime applicants with funded foreign subawards. There is currently no opportunity for foreign entities to incorporate funded foreign subawards into their proposal. NIH is exploring technical solutions to allow for PRIMED-AI NOFOs to use the PF5 to submit international collaboration applications. We encourage the extramural community to subscribe to the NIH Guide for Grants and Contracts, as that will be the venue that NIH will publish Guide Notices with the updates on the issue.

1. How does PRIMED-AI define “benchmark datasets”?

To the PRIMED-AI program, benchmark datasets consist of carefully curated and labeled data that are essential for validating and comparing models developed in the PRIMED-AI Consortium.

2. How does PRIMED-AI define “clinical decision support (CDS) tool”?

PRIMED-AI defines clinical decision support (CDS) tools as a type of software, computational model, or digital system that is incorporated into clinical workflows to assist in determining a course of action related to patient care.

3. How does PRIMED-AI define “clinical imaging”?

PRIMED-AI defines clinical imaging as any FDA-approved imaging modality used in patient care, including radiologic (e.g., radiographic, computed tomographic, magnetic resonance, molecular, radionuclide imaging), ophthalmologic (e.g., Optical Coherence Tomography), endoscopic, and dermatologic imaging, and video. Clinical imaging of human participants is intended to be the anchor data type that multimodal data are integrated within the PRIMED-AI Program, which will form the basis for AI algorithm development and testing of clinical decision support (CDS) tools.

4. How does PRIMED-AI define “harmonization”?

PRIMED-AI defines harmonization as the process of bringing together data from different sources and ensuring that it is consistent, comparable, and compatible. This involves standardizing data formats, structures, and definitions so that data from various sources can be integrated and analyzed together effectively.

5. How does PRIMED-AI define “interoperability”?

PRIMED-AI defines interoperability as the ability for AI models and associated data and metadata to be understood and work across different AI platforms and have the potential to be used consistently across different health systems.

6. How does PRIMED-AI define “multimodal data (MMD)”?

PRIMED-AI defines multimodal data as different types of data and information from multiple sources that may include multiple clinical imaging modalities and non-imaging health data (e.g., electronic health records, EEG, EKG, laboratory test results (-omics), wearable sensor data, medical reports). Multiscale data are encouraged.

7. How does PRIMED-AI define “Playbook”?

PRIMED-AI defines a Playbook as a collection of actionable guidelines, standardized protocols, and/or standardized operating procedures for the reliable and effective development and deployment of multimodal clinical decision support tools. The Playbook is a collection of frameworks.

8. How does PRIMED-AI define “Precision Medicine”?

Sometimes called personalized medicine or individualized medicine, precision medicine refers to a healthcare approach that uses information based on a patient’s individual characteristics such as health measures, genotype, phenotype, environment, and lifestyle information to guide, tailor, and optimize decisions related to their medical care and management.

9. How does PRIMED-AI define “real-world data”?

PRIMED-AI defines real world data as data relating to patient health status and/or the delivery of health care that is routinely collected from a variety of sources during everyday clinical practice, rather than from clinical trials or controlled experiments. Within the PRIMED-AI framework, this typically encompasses multimodal data—such as electronic health records (EHRs), clinical imaging (radiology, pathology, fMRI etc.), laboratory test results, claims/billing data, and wearable sensor data. RWD is essential for training, testing, and externally validating Clinical Decision Support (CDS) tools to ensure the AI models are accurate, generalizable, and perform reliably across diverse, real-world patient populations.

10. How does PRIMED-AI define “uncertainty quantification”?

To PRIMED-AI, uncertainty quantification is measuring or quantifying the impact of uncertainties in complex systems, including quantifying the confidence in outcomes predicted by multimodal AI models.

11. How does PRIMED-AI define “validation”?

Validation exists on a continuum in the PRIMED-AI Program. Analytical or technical validation is based on the evaluation of algorithmic performance and the ability of a multimodal AI model to make accurate predictions.  Initially, a model or algorithm can meet expected performance on retrospective and/or entirely new clinical datasets within the confines of a specific hospital or healthcare system. It is useful locally (internally) but is not yet applicable (generalizable) to the wider real-world population. Subsequently, for clinical validation, a model or algorithm can be tested (externally) on new wider real-world population datasets to predict a meaningful outcome and meet regulatory criteria for the claimed use case. The PRIMED-AI Program anticipates validation of projects along this continuum as outlined in the NOFOs.

12. How does PRIMED-AI define “verification”?

PRIMED-AI defines verification as the process by which data integrity and construction of models is assessed for appropriateness within the context of use or intended purpose.

1. What is the announcement number of this NOFO?

RFA-RM-27-014

2. What is the purpose of this announcement?

The purpose of this NOFO is to solicit applications for the creation of a Validation Center for the PRIMED-AI Program to serve as a dedicated hub for the comprehensive evaluation and characterization of AI-enabled, image-based, multimodal CDS tools developed by the PRIMED-AI Consortium, to ensure these tools are reliable, reproducible, and generalizable.  The Validation Center will focus on verification, validation, interoperability, and uncertainty quantification to comprehensively characterize performance of the CDS tools developed by the program. The core functions of the Validation Center will involve the systematic validation of PRIMED-AI Consortium deliverables, including those emerging from award recipients under the Playbook, Data-to-Model Academic-Industrial Partnerships, and Model-to-Clinic NOFOs.

3. How many awards will be funded?

The program anticipates funding one award.

4. What is the award budget?

$2,000,000 direct costs for the first year and up to $3,200,000 direct costs for the remaining four years of the project.

5. When are applications due?

The receipt date is October 2, 2026. All applications are due by 5:00 PM local time of applicant organization. 

6. To whom can I reach out with questions about this announcement?

All questions are welcome. Please reach out to [email protected] and indicate which funding opportunity your question relates to.

What are the key dates in the timeline for planning our application?

1. Are Letters of Intent (LOI) required?

No, NIH no longer requests or accepts Letters of Intent.

2. What are the requirements for Multi-PD/PI (MPI) and other Validation Center staff?

A multi-PD/PI application is required - one PD/PI is expected to commit at least 2 person months annually and the other(s) should devote at least 1 person month each. The PRIMED-AI Program will have regular conference calls and meetings. Applicants should request funds for 2-6 group members to attend annual meetings, Consortium-led tutorials, and open innovation meetings, as appropriate. The Validation Center should budget at least 6 person months for a dedicated project manager/director (PM/PD) for the project with the appropriate scientific expertise and project management responsibilities, who would support the PI(s) with project management. The PM/PD will be the primary liaison with the PRIMED-AI Program and Steering Committee.

3. How do I submit my application?

Organizations must submit applications to Grants.gov (the online portal to find and apply for grants across all Federal agencies). Applicants must then complete the submission process by tracking the status of the application in the eRA Commons, NIH’s electronic system for grants administration. NIH and Grants.gov systems check the application against many of the application instructions upon submission. Errors must be corrected and a changed/corrected application must be submitted to Grants.gov on or before the application due date and time.  If a Changed/Corrected application is submitted after the deadline, the application will be considered late. Applications that miss the due date and time are subjected to the NIH Grants Policy Statement Section 2.3.9.2 Electronically Submitted Applications. 

Applicants are responsible for viewing their application before the due date in the eRA Commons to ensure accurate and successful submission.

4. The page limit for the Research Strategy section for each of the Core Components (i.e. the Validation Core, the Translation Core, and the Fast-track Core) is 12 pages. Does the information requested in "Other Plans" have to fit in the 12 page limit?

Yes, the requested information listed in the "Other Plans" section for each core must be included within the 12-page limit. Except for the Error and Technical Management Plan, the "Other Plans" are part of the 12-page Research Strategy and are not separate attachments.

5. Specific Aims are required for the other components (Validation, Translational, and Fast-Track Cores). For the Overall Component, is are Specific Aims required? 

No, a Specific Aims page is not required for the Overall Component section of your response, but its inclusion is acceptable.

1. Section II. Award Information indicates that the application budget is up to $3,200,000 direct costs for the remaining four years of the project. Does this mean up to $3,200,000 per year for years 2-5, or $3,200,000 in total for years 2-5?

The application budget for the Validation Center NOFO is limited to $2,000,000 direct costs for the first year and up to $3,200,000 direct costs per year for the remaining four years of the project. 

2. Section I. Notice of Funding Opportunity Announcement indicates that the Validation Center will continuously sequester and annotate multimodal data for validation and benchmarking purposes and that the benchmark data sets will be made public through the PRIMED-AI web portal. How should this best be accomplished to maintain sequestered data as private for testing purposes?

The Validation Center is responsible for maintaining datasets that are independent and distinct from the data used by other PRIMED-AI grantees for model training. These independent datasets should remain sequestered within the Validation Center and should not be shared through the PRIMED-AI web portal. Data used for model training may be publicly shared, to the extent possible.

3. Can a single institution hold both the Logistics Center and the Validation Center? Is there any programmatic preference or expectation that these two centers reside at different institutions?

The NOFOs do not prohibit one institution from holding both centers. However, the programmatic expectation is that the Logistics and Validation Centers will be at different institutions to preserve independent validation, data sequestration, and unbiased evaluation. Applicants are strongly encouraged to attend the NIH public webinar announced in NOT-OD-26-099 for further guidance.

4. If a single institution held both centers, what separation, e.g.  distinct MPI leadership, firewalled budgets, a formal conflict-of-interest plan. Would NIH expect to preserve the Validation Center's independence and the integrity of program-level evaluation?

The NOFOs do not prescribe specific institutional firewalls. However, a co-located structure should include:

• Separate scientific leadership and decision-making;
• Separate budgets, accounts, and cost allocation;
• Independent validation protocols and data access;
• Formal conflict-of-interest, recusal, and dispute-resolution procedures; and
• Independent program evaluation.

These controls would support NIH requirements concerning scientific integrity, allocable costs, data sequestration, and unbiased validation.

1. What is the announcement number of this NOFO?

RFA-RM-27-015

2. What is the purpose of this announcement?

The purpose of this NOFO is to establish a Logistics Center for the PRIMED-AI Program to facilitate and coordinate PRIMED-AI activities to maximize impact for the PRIMED-AI community.  Within the Logistics Center, three integrated Cores will serve the functions of Administration, Evaluation, and Outreach for the PRIMED-AI Consortium. It is expected that the Logistics Center will work closely with other PRIMED-AI consortium members to collect, curate and disseminate information regarding tools developed by the PRIMED-AI award recipients, and will facilitate widespread access and awareness across the relevant scientific, patient, and clinical communities.

3. How many awards will be funded?

The program anticipates funding one award.

4. What is the award budget?

$575,000 direct costs for the first year and up to $2,500,000 direct costs per year for the remaining four years of the project.

5. When are applications due?

The receipt date is October 2, 2026. All applications are due by 5:00 PM local time of applicant organization. 

6. To whom can I reach out with questions about this announcement?

All questions are welcome. Please reach out to [email protected] and indicate which funding opportunity your question relates to.

What are the key dates in the timeline for planning our application?

1. Are Letters of Intent (LOI) required?

No, NIH no longer requests or accepts Letters of Intent.

2. What are the requirements for Multi-PD/PI (MPI) and other Logistics Center staff?

A multi-PD/PI application is required - one PD/PI is expected to commit at least 2 person months annually and the other(s) should devote at least 1 person month each.

The Logistics Center is responsible for managing in person meetings and regular conference calls. Budgets should reflect costs associated with managing and hosting these meetings, both in person and virtual.  Applicants should also request funds for 2-6 group members to attend annual in person meetings, Program-led tutorials, and open innovation meetings, as appropriate.

The Logistics Center should budget at least 6 person months for a dedicated project manager/director (PM/PD) for the project with the appropriate scientific expertise and project management responsibilities, who would support the PI(s) with project management. The PM/PD will be the primary liaison with the PRIMED-AI Program and Steering Committee.

3. How do I submit my application?

Organizations must submit applications to Grants.gov (the online portal to find and apply for grants across all Federal agencies). Applicants must then complete the submission process by tracking the status of the application in the eRA Commons, NIH’s electronic system for grants administration. NIH and Grants.gov systems check the application against many of the application instructions upon submission. Errors must be corrected and a changed/corrected application must be submitted to Grants.gov on or before the application due date and time.  If a Changed/Corrected application is submitted after the deadline, the application will be considered late. Applications that miss the due date and time are subjected to the NIH Grants Policy Statement Section 2.3.9.2 Electronically Submitted Applications. 

Applicants are responsible for viewing their application before the due date in the eRA Commons to ensure accurate and successful submission.

1. Are the required $2.5M in restricted funds from the R&R budget included within the $2.5M annual direct-cost cap for Years 2–5, or is this value in addition to the funds available for Logistics Center operations and administrative functions? 

Yes, the $2.5M total cost restricted funds must be included in the $2.5M direct cost budget cap for years 2-5 – this is not in addition to the $2.5M direct cost budget cap.  This means that direct costs associated with restricted funds as well as the logistics center itself must fall under the $2.5M direct cost cap.  Direct costs associated with the restricted funds must be entered under the “Other Expenses” category of the budget section.  The amount of direct costs requested for the logistics center as well as the restricted funds must reflect the needs of the proposed project and may not exceed $2.5M direct costs each year for years 2-5.

2. Can a single institution hold both the Logistics Center and the Validation Center? Is there any programmatic preference or expectation that these two centers reside at different institutions?

The NOFOs do not prohibit one institution from holding both centers. However, the programmatic expectation is that the Logistics and Validation Centers will be at different institutions to preserve independent validation, data sequestration, and unbiased evaluation. Applicants are strongly encouraged to attend the NIH public webinar announced in NOT-OD-26-099 for further guidance.

3. If a single institution held both centers, what separation, e.g.  distinct MPI leadership, firewalled budgets, a formal conflict-of-interest plan. Would NIH expect to preserve the Validation Center's independence and the integrity of program-level evaluation?

The NOFOs do not prescribe specific institutional firewalls. However, a co-located structure should include:

• Separate scientific leadership and decision-making;
• Separate budgets, accounts, and cost allocation;
• Independent validation protocols and data access;
• Formal conflict-of-interest, recusal, and dispute-resolution procedures; and
• Independent program evaluation.

These controls would support NIH requirements concerning scientific integrity, allocable costs, data sequestration, and unbiased validation.

4. The Logistics Center requires a multi-PI team (one PI ≥2 person-months, others ≥1) plus a dedicated ≥6-month project manager. If one institution held both centers, may the same individuals serve as MPIs or key personnel across both, or must leadership and key personnel be entirely distinct?

The NOFOs do not require leadership and key personnel to be entirely distinct; the same individuals may serve on both applications. However, separate MPI leadership and dedicated project managers are strongly preferred to preserve the Validation Center’s scientific independence and sequestered data. Any shared personnel must have clearly distinct roles, meet each award’s effort requirements, disclose all support, and avoid scientific, budgetary, or commitment overlap, including total effort exceeding 12 person-months.

5. The Logistics Center must pass most of its year 2–5 restricted funds to non–Logistics-Center recipients (≥75% of the teaming fund, ≥50% of the curricula fund). If the Validation Center were at the same institution, may it receive those restricted funds as a collaborator, or would an intra-institutional transfer be disallowed or fail the pass-through requirement?

The NOFO does not clearly state whether an internal allocation to a Validation Center at the same institution would satisfy the requirement that specified funds go to collaborators or non-Logistics Center recipients. Because an internal transfer is not a subaward to a separate legal entity, applicants should not assume it counts toward the required percentages without written NIH confirmation. Any supported work must also be scientifically distinct from the Validation Center’s existing award scope and free of duplicate funding.

6. The Logistics Center's Evaluation Core external advisory group must include stakeholders "not associated with the PRIMED-AI Program." If both centers were at one institution, does that broaden what counts as "associated" and narrow who can serve as an independent advisor?

The NOFO requires advisors who are not associated with the PRIMED-AI Program. Co-location does not necessarily disqualify everyone at the same institution, but it could create a perceived lack of independence. Advisors from outside the institution who have no role, funding, or decision-making responsibility in either center would provide the strongest compliance with the NOFO’s scientific and programmatic independence expectations.

7. The Logistics Center hosts the web portal and ingests validated tools into the Common Fund Data Ecosystem. If validation occurred at the same institution, is there any concern about the portal-hosting entity and the validating entity being the same, and are there specific neutrality or information-security expectations?

The NOFOs do not prohibit this arrangement. However, validation datasets, benchmark information, interim findings, and validation decisions should remain controlled by the Validation Center and separate from Logistics Center portal personnel. Appropriate safeguards should include role-based access, data sequestration, audit trails, cybersecurity controls, documented publication procedures, and Validation Center authority over the release and interpretation of validation findings.

1. What is the announcement number of this NOFO?

RFA-RM-27-011

2. What is the purpose of this announcement?

The purpose of this NOFO is to solicit applications for the design, development, and preliminary validation of robust frameworks for the application of multimodal-artificial intelligence (AI) models for clinical use. Frameworks developed through these awards would populate a "playbook", defined as a collection of actionable guidelines, standardized protocols, and/or standard operating procedures (SOPs) for reliable and effective development and deployment of multimodal AI tools.  It is expected that the frameworks delineated in this playbook will directly address PRIMED-AI objectives and needs, while also remaining flexible enough to enable sufficient extensibility and interoperability for use across a broad spectrum of multimodal biomedical AI applications, both internal and external to the PRIMED-AI Program.

3. How many awards will be funded?

The program anticipates funding five awards.

4. What is the award budget?

$300,000 direct costs per year.

5. When are applications due?

The receipt date is October 9, 2026. All applications are due by 5:00 PM local time of applicant organization. 

6. To whom can I reach out with questions about this announcement?

All questions are welcome. Please reach out to [email protected] and indicate which funding opportunity your question relates to.

What are the key dates in the timeline for planning our application?

1. Are Letters of Intent (LOI) required?

No, NIH no longer requests or accepts Letters of Intent.

2. How do I submit my application?

Organizations must submit applications to Grants.gov (the online portal to find and apply for grants across all Federal agencies). Applicants must then complete the submission process by tracking the status of the application in the eRA Commons, NIH’s electronic system for grants administration. NIH and Grants.gov systems check the application against many of the application instructions upon submission. Errors must be corrected and a changed/corrected application must be submitted to Grants.gov on or before the application due date and time.  If a Changed/Corrected application is submitted after the deadline, the application will be considered late. Applications that miss the due date and time are subjected to the NIH Grants Policy Statement Section 2.3.9.2 Electronically Submitted Applications. 

Applicants are responsible for viewing their application before the due date in the eRA Commons to ensure accurate and successful submission.

3. The NOFO lists several items under Research Strategy (relevant expertise, management plan, consortium collaboration plan) — should these be incorporated into the Research Strategy attachment, or submitted as separate attachments?

These items belong in the Research Strategy attachment itself, since they are listed specifically under that section in the NOFO. Only items listed elsewhere in the NOFO (under "Other Attachments") should go in separate attachments.

4. Are preliminary results expected to be included in the submission draft?

Preliminary results are not explicitly required but are encouraged if relevant. Applicants should follow the NOFO instructions and review criteria closely, and include preliminary data if it helps demonstrate the framework will work or if it supports the applicant's expertise and track record.

1. What constitutes a "framework," and how does it differ from a model or an image processing pipeline?

A framework is a flexible guiding tool — such as a set of guidelines, standards, or protocols — that helps standardize the development and real-world use of multimodal AI-based clinical decision support (CDS) tools. It is not a model (the actual tool that performs prediction, diagnosis, etc. — that falls under other NOFOs) and not an image processing pipeline (not considered multimodal AI). A framework refers to protocols and guidelines that help move multimodal AI-based CDS tools out of the development/testing stage and into validated, real-world clinical use. An illustrative example (outside this specific field) is the FAIR data principles: a widely used set of guidelines defining what data must meet to be usable across the data community. Full definitions are available in the Glossary above.

2. Could you clarify what is meant by "distinct” frameworks? Would an end-to-end framework be more responsive than a narrower one?

"Distinct" frameworks would address separate aspects or stages of CDS tool development — for example, a data preprocessing framework and a data analytics framework count as separate, even within the same pipeline. It does not mean applying a similar framework redundantly across different data types (e.g., near-identical metadata standards for two different data types would not count as distinct). Building an actual AI model/tool for diagnosis or prognosis does not itself count as a "framework" under this NOFO, either. There is no specific synergy NIH is targeting between the two frameworks — that depends on each applicant's project and needed collaborations. Full end-to-end pipeline frameworks are not more encouraged or prioritized than narrower, single-stage frameworks; what matters is a clear, well-justified proposal with two or more distinct frameworks that reviews well. Applicants proposing a full pipeline should realistically weigh whether it's achievable in two years given their lab's capacity and collaborators.

3. How many frameworks can I propose to develop?

All applicants are required to propose at least two distinct frameworks, but more are allowable if feasible within the proposed team expertise, budget and two year award time period.  Please keep in mind that applications proposing frameworks or methods that are duplicative to existing frameworks or are incremental advancement to existing frameworks will be considered of low programmatic relevance, which will decrease likelihood of funding.

4. Why is more than one distinct framework required? 

Two or more distinct frameworks are required. This requirement exists to give applicants flexibility. Since awardees will need to coordinate with other awardees working on related efforts (and with the Validation Center), requiring two or more frameworks allows applicants room to shift or adjust their approach as the overall Playbook effort evolves through this collaboration. 

5. Do both frameworks need to be equally weighted? For example, one aspect may be expanding OMOP to accommodate echocardiography parameters (a smaller, less exciting but necessary framework).

No, frameworks do not need to be equally weighted — applicants can propose a mix of simpler and more complex frameworks, or two complex ones, as long as there's a clear rationale for the chosen focus and how the frameworks fit together. Regarding the OMOP example specifically, a well-designed case study (e.g., extending OMOP for one condition) that could be generalized to other conditions would be valuable, and demonstrating this kind of extensibility/flexibility strengthens a proposal's competitiveness.

6. What makes a proposed framework "duplicative" or "low effort," and does this include frameworks based on prior publications or existing tools?

A framework is considered duplicative/low effort if it is already published and complete, combines several existing components with minimal original contribution, or if the proposal only makes incremental changes to an existing framework or tool. Using an existing open-source tool or method as a starting component within a genuinely novel framework is not duplicative -  extending something narrow into new, previously unaddressed applications, is more likely to be viewed favorably as novel. The key question is how much genuinely new contribution the proposal makes. Applicants unsure whether their approach counts as duplicative are encouraged to reach out with a specific aims page for feedback.

7. Would a standalone multimodal software application and a cloud-based web version of the same AI platform be considered two separate frameworks, or a single framework with different deployment options?

This describes developing an actual clinical decision support tool/platform, which is not in scope for this NOFO. Applicants building the tool itself should instead look at the related Data-to-Model and/or Model-to-Clinic NOFOs.

8. Is anything considered non-responsive to this NOFO?

Projects that do not propose two or more distinct frameworks for development will be considered non-responsive and will not be reviewed. Each applicant is required to propose two or more distinct frameworks for development – the frameworks should be on two separate framework topics, rather than two methods of addressing the same framework topic.

9. How will the Playbook program interact with the Validation Center, including for validation of specific frameworks?

All awardees across the funding components work together as part of a consortium. Playbook awardees will form their own working group to develop frameworks, but must coordinate closely with the Validation Center, particularly when a framework relates specifically to validation. Whether an applicant needs to work directly with the Validation Center on validation depends on the framework itself: if the framework is specifically about validating whether a tool works, applicants may be able to partner with the Validation Center to conduct those studies. If the framework is more front-end focused (e.g., a training framework for clinical use), the Validation Center is less likely to directly support that work. In all cases, applicants should demonstrate — through their expertise and prior work — why their proposed approach is sound, and should review the NOFO's review criteria closely, since this is explicitly addressed there.

10. How does my planned interaction with other components of the PRIMED-AI Program need to be addressed in my application?

All PD(s)/PI(s) funded under the PRIMED-AI Program will be expected to participate in PRIMED-AI Consortium activities, including attending and participating in Steering Committee meetings and accepting and implementing the consensus guidelines and procedures, as appropriate.  In addition, all Playbook award recipients are required to establish a Playbook working group that will enable the recipients to coordinate development of the final playbook product and ensure minimal overlap between each of the recipients.  This Playbook working group will be required to meet at least monthly to streamline playbook development, and PIs of Playbook projects are expected to participate in these meetings.  When applying, all applicants are required to include a Consortium Collaboration Plan to briefly describe their plans to work with other Playbook award recipients, the Logistics Center for dissemination and web portal inclusion, and the Validation Center to harmonize processes for appropriate development and validation of multimodal AI tools.

11. Are Playbook awardees allowed to create innovative software tools as part of their framework?

Yes, software tools are allowable as submissions to this NOFO as long as they address the underlying need of the framework selected.

12. Can the Playbook be comprised of code or is it intended to be written guidance only?

Code is allowable as part of the framework, however, as stated in the NOFO, all frameworks developed under this NOFO and the resulting playbook are expected to be shared and adopted by the broader research community and must be designed in such a way as to ensure that this is achievable.  Any framework developed under this NOFO that incorporates code will need to ensure that enough written guidance is provided to allow the broader research community to utilize the framework.

13. Does a CDS tool need to already exist for a framework to be developed around it, and is developing a tool as part of a framework in scope?

This is case-by-case. Frameworks focused on areas like metadata standards don't necessarily require an existing tool. Frameworks addressing validation or downstream clinical compatibility may require that a tool already exists. Separately, developing a tool is not inherently out of scope — a tool counts as a framework if it assists with validation or helps move a CDS tool from development into clinical use. A tool designed to sit on top of and analyze a CDS tool would be relevant here. However, if the tool being built is itself the CDS tool (e.g., the tool that predicts, diagnoses, or generates clinical output), that belongs under a different NOFO.

14. How specific should we be with each framework when defining data modalities, use cases, and audience? For example, should guidelines be defined narrowly (e.g., for a specific AI tool, modality, and disease application) or more broadly (e.g., for a general application area), or somewhere in between?

Broader, more generalizable frameworks are preferred, though narrower frameworks tied to specific data types aren't excluded. The key consideration is real-world utility — frameworks should ideally be usable beyond the specific program, helping others developing similar tools in practice. As with other responsiveness questions, applicants unsure whether their scope fits the NOFO are encouraged to contact NIH staff directly (e.g., via a specific aims page) for feedback.

15. Is it acceptable for a proposed framework to be primarily EHR-focused, with imaging included as an additional modality, or is more balanced multimodal integration expected?

Imaging needs to be the central, forward-facing component of the framework, with EHR and other data types supporting it rather than the reverse. That said, applicants are encouraged to reach out to NIH staff directly to confirm responsiveness. Overall, the goal is a reasonable balance across modalities while keeping imaging central.

16. The NOFO says a final framework must be posted as a 12-month milestone. What is the expectation for how the 2nd year would be used if the framework is already done?

The 12-month deliverable is not expected to be a finished framework, but rather a defined outline or "framework for a framework" that goes beyond a basic outline. Because awardees may need to revise their approach to stay compatible with other groups in the consortium (e.g., their Playbook working group), the second year is intended for making these adjustments and refinements based on input from other awardees, ultimately producing the finalized framework.

17. A U01 is a two-year project with only $300K in funding. Given that multimodal data are typically distributed across multiple partners, building collaborations, collecting data, and curating data alone can take up to two years. Under these constraints, is a project focused solely on an "AI playbook" justifiable, and should the priority instead be on what data will be used, how it will be curated, and how supporting infrastructure will be developed?

Applicants are not expected to curate large datasets — the $300K/year (direct costs) is meant to fund development of frameworks/protocols, not data collection. For example, rather than acquiring imaging and biomarker data, the focus should be on researching and defining requirements (e.g., metadata standards) needed to ensure a downstream tool would work reliably across multiple or generalizable clinical settings. Note the budget is up to $300K in direct costs per year (not total for the project), so applicants can request more than $300K overall if justified. More broadly, NIH sees a real gap in standardized guidance for this type of work (e.g., benchmarking, clinical workflow integration, privacy preservation), which is why this opportunity emphasizes developing broadly useful protocols — including revisiting and adapting them in year two as the technology (e.g., generative AI) evolves — rather than data curation itself.

18. Should the proposal focus on actionable aims rather than hypotheses?

For this NOFO, the focus should likely be on actionable aims that will result in a framework “product," rather than on exclusively hypothesis-driven projects.

19. Do frameworks need to be validated with real-world use cases and data? How can a framework be validated without building a tool?

Yes, all frameworks must be validated — an unvalidated framework has limited value to the community. The scope and method of validation depend entirely on what the framework covers. Narrower frameworks (e.g., metadata standards, FAIR-type guidelines, or training-focused frameworks) generally don't require building a new tool and may be validated using existing tools. Broader frameworks spanning an entire CDS pipeline are inherently more complex to validate. Applicants should carefully scope their framework to be both realistic and “validatable,” and much of the specific path forward is expected to be worked out through collaboration with other consortium members and awardees — which is part of why a strong collaboration plan matters in the application.

20. Are there specific sources of data you are looking for?

There is no required list of data sources. The only firm requirement is that the data must include an imaging component and must support development of a tool usable in a real clinical environment. Data sources can come from Bridge2AI, other NIH programs, or external sources (e.g., an institution's own patient data system). The key criterion is that the data fits your framework's purpose.

21. For small businesses, procuring imaging data can be difficult. Does the proposal require us to specify how we will obtain this data? Can we use publicly available data for proof of concept?

Yes, publicly available data can be used, as long as it fits the context and goals of the proposed project. Using private data instead is also acceptable.

22. Are there specific disease areas you are interested in?

No, there's no disease-area preference. The only requirement is that the project be multimodal and include an imaging component. As a Common Fund initiative, the program is intentionally designed to address issues relevant across many NIH institutes and centers.

23. Do you allow use of open source tools on top of which we can build the software?

Yes, use of open source tools is allowed, as long as you're not working outside that tool's intended protocol/use or violating any intellectual property. Applicants should cite any open source tools used so it's clear what was incorporated.

24. Are projects encouraged to work with datasets generated by Bridge2AI?

Use of Bridge2AI datasets is allowed but not required. Regardless, projects need to have imaging data.

25. Can a subscription to Claude Max be a budget line item for this NOFO?

Budget items must be justified, and a well-justified item is not inherently disallowed at the application stage. However, even if an application scores well and is selected for funding, NIH grants management has final say on allowability.

26. Are NIH Intramural Research Labs allowed to apply?

Yes, applications may be submitted by or include collaborations with NIH intramural research programs.  Special budgetary requirements are associated with inclusion of intramural labs and are included under Section IV. Application and Submission Information, 2. Content and Form of Application Submission, R&R Budget. Please also keep in mind that applications from the NIH Intramural Program submitted in response to this NOFO must include a current (i.e. within 2 months of application due date) letter from the Scientific Director of their Division indicating that the intramural scientist will be able to collaborate on the project.

1. What is the announcement number of this NOFO?

RFA-RM-27-012

2. What is the purpose of this announcement?

The purpose of the Data-to-Model Academic-Industrial Partnership (D2M-AIP) NOFO is to support multi-sector and multi-disciplinary research teams, including investigators from both academia and industry, to create mutually beneficial opportunities for partners in the pre-competitive development stage.  D2M-AIP projects are primarily focused on the integration and harmonization of novel multiscale, multimodal data with clinical imaging data and the development and testing of truly novel AI-enabled, image-centered, multimodal CDS tools, developed in pursuance as Software as a Medical Device (SaMD). D2M-AIP projects will leverage existing resources across the partnership, such as high-performance computing capabilities and access to clinical data, to generate robust validation data and engage with regulators, positioning the technology for rapid post-award translation into a viable and impactful clinical product.

3. What is the focus of Phase 1 (the UG3 phase)?

The UG3 phase is expected to focus on systematic PRIMED-AI CDS tool development and comprehensive technical validation, which will generally rely on retrospective or de-identified, curated datasets for technical validation. This phase will last up to 2 years.

4. What is the focus of Phase 2 (the UH3 phase)?

Following successful transition, the UH3 phase is expected to focus on further development of the PRIMED-AI CDS tool and validation of the tool's potential for delivering future clinical value.

5. How many awards will be funded?

The program anticipates funding six to eight awards.

6. What is the award budget?

$450,000 in direct costs per year for the UG3 phase and $800,000 in direct costs per year for the UH3 phase.

7. When are applications due?

The receipt date is October 19, 2026. All applications are due by 5:00 PM local time of applicant organization. 

8. To whom can I reach out with questions about this announcement?

All questions are welcome. Please reach out to [email protected] and indicate which funding opportunity your question relates to.

What are the key dates in the timeline for planning our application?

1. Are Letters of Intent (LOI) required?

No, NIH no longer requests or accepts Letters of Intent.

2. Are clinical trials allowed?

While they are not required, applicants may propose to conduct clinical trials if appropriate and feasible within the project period. If a trial is proposed, the applicant must describe the trial thoroughly and provide strong justification for their chosen design. If a trial is proposed as part of a D2M-AIP project, it is anticipated that such trials will focus primarily on the efficacy of the developed CDS tools and be evaluated within de novo dedicated prospective trials or evaluated leveraging ongoing or planned trials to assess CDS performance as secondary objectives, integrating with existing infrastructure and data streams.

3. How do I submit my application?

Organizations must submit applications to Grants.gov (the online portal to find and apply for grants across all Federal agencies). Applicants must then complete the submission process by tracking the status of the application in the eRA Commons, NIH’s electronic system for grants administration. NIH and Grants.gov systems check the application against many of the application instructions upon submission. Errors must be corrected and a changed/corrected application must be submitted to Grants.gov on or before the application due date and time.  If a Changed/Corrected application is submitted after the deadline, the application will be considered late. Applications that miss the due date and time are subjected to the NIH Grants Policy Statement Section 2.3.9.2 Electronically Submitted Applications. 

Applicants are responsible for viewing their application before the due date in the eRA Commons to ensure accurate and successful submission.

4. Is there a special structure required for the 12-page research strategy? Should it be organized by the five pillars, or do the standard Significance/Innovation/Approach sections still apply? 

The traditional Significance, Innovation, and Approach sections remain the primary structure, but should be explicitly mapped to the five pillars: Significance → clinical value proposition of the SaMD; Innovation → novel data integration and model structures; Approach → pillars 1–4 explicitly (data quality/governance, clinically grounded technology, pathway to implementation/adoption, integrated multidisciplinary team). Pillar 5 (error mitigation and technical management) is a separate required attachment and does not count toward the 12-page limit.

5. Is this one application that includes both the UG3 and UH3 phases together, or do we submit the UG3 section now and the UH3 later once milestones are reached? 

You must submit a single, combined application including both the UG3 and UH3 phases together within one 12-page Research Strategy, along with a clearly defined list of quantitative milestones you will achieve by the end of the UG3 phase to demonstrate technical feasibility. You do not submit a new grant application for the UH3 phase. As you near the end of Year 2, NIH will conduct an administrative/programmatic review of your progress. If you have met your pre-negotiated UG3 milestones and remain scientifically aligned with consortium goals, your award will be prioritized and administratively transitioned to the UH3 phase.

1. What level of preliminary data would you like to see in the proposal? I understand this is a very fast paced timeline. Would collection of new data points in the UG3 phase be allowed?

The NOFO requires that the pilot test of the proposed CDS tool in the UG3 phase generate initial preliminary data to support subsequent UH3 development. The UG3 phase generally relies on retrospective or de-identified, curated datasets for technical validation, so applicants should plan for rigorous retrospective analysis during this stage.

In practice, applicants need to convince reviewers that the proposed aims are achievable, so documenting access to the data needed for curation, harmonization, and model development will likely be score-driving. Collecting new data during the UG3 phase is allowed, but applicants should consider how reviewers may perceive the associated risk. As with many aspects of these NOFOs, this is highly project-specific, and applicants are encouraged to reach out to program staff with questions about their particular approach.

2. May a D2M-AIP project build upon an existing NIH-supported imaging pipeline, if the proposed project is scientifically distinct (novel multimodal integration, new intended use, independent validation, SaMD translation)?

Yes — a new intended use and a scientifically distinct project (different population, different data) qualifies it as a new model. Applicants are encouraged to reach out to program officers with tentative specific aims for a clearer determination.

3. What level and type of substantive industry involvement does NIH expect under D2M-AIP? What capabilities, activities, or deliverables should the industry partner contribute for the application to be a true academic-industrial partnership rather than an academic project with a passive collaborator? 

NIH expects the industry partner to be a co-leader of the project, serving as MPI with decision-making authority. This is a hard eligibility requirement. The industry partner is expected to contribute across three areas: commercialization pathway expertise, regulatory engagement, and shared governance, covering not just scientific direction, but also data rights and IP/inventions, as outlined in the required MPI Leadership Plan.

4. I'm not affiliated with an academic institution but the institution I am affiliated with has a clinical institute and research infrastructure. Am I still eligible to apply for the D2M-AIP?

At least one PI must be from the applicant academic institution with expertise in preclinical and/or translational AI research, but may include additional institutions to ensure the group has a critical mass of expertise to accomplish the goals of the project.

5. Should the industry partner be an AI company to scale or a company that uses the AI?

The primary industrial partner must be a for-profit organization of any size that brings a proven track record or strong potential for the development, regulatory approval, and/or commercialization of medical software. While they do not have to be exclusively an "AI company," they must possess the capabilities to scale, validate, and commercialize the resulting software product.

6. Does the academic-industry partnership need to be pre-existing; e.g., a year of demonstrated collaboration, joint publications, or patents — or is strong complementary expertise with no prior collaboration sufficient?

No, the partnership does not need to be pre-existing. An academic-industrial partnership is required for D2M-AIP. The program is specifically designed to support mutually beneficial opportunities at the pre-competitive development stage. Reviewers evaluate the partnership based on the strength, synergy, and completeness of the team assembled for the 5-year project, rather than its historical connection (prior joint papers or patents are not required).

7. Can a small for-profit company or startup satisfy the required industrial partner role based on strong potential for SaMD development and commercialization, even without a prior FDA-authorized device? May the academic institution serve as prime recipient while the company's principal serves as the required industrial MPI through a substantive sub-award?

Yes, a startup or small for-profit company can satisfy the requirement. The NOFO explicitly defines the eligible industry partner as "a for-profit organization of any size" and does not require a preexisting track record of FDA clearance. Evaluation is based on the company's potential capability to support development, evidence generation, and eventual commercialization. The academic-institution-as-prime, company-as-MPI-via-sub-award structure is acceptable, and may be optimal, since academic institutions often have more experience administering grants.

8. Does a nonprofit healthcare system or medical center/hospital qualify as the required industry partner — for example, if it provides clinical-grade imaging data access, supports clinical validation, and has technology translation capabilities?

No, not on its own. A for-profit company is still required. This is a hard eligibility requirement, and non-profit partners explicitly do not satisfy it, regardless of their data access, validation support, or translation capabilities, unless the entity is legally incorporated as a for-profit organization. The industrial partner must also bring a proven track record or strong potential for development, regulatory approval, and/or commercialization of medical software.

9. Are there restrictions or guidelines on the budget allocation for industry partners.

Rather than enforcing a rigid split, we designed the NOFO allow for the level of participation and budget may vary among partners based on the specific translational goals of the project. However, the academic-industrial partnership must establish a mutually beneficial collaborative structure, which should be explicitly detailed within a Multiple PD/PI Leadership Plan covering budget and resource allocation.

10. If the industry partner is a startup founded by the academic PI (who also has an ownership interest), is that generally acceptable? Are there specific expectations regarding the PI's role or conflict-of-interest management? 

This arrangement is generally acceptable, but subject to rigorous conflict-of-interest (COI) management and explicit role separation. Under the NOFO's Academic-Industrial MPI structure, the academic PI cannot serve as the Principal Investigator for both the academic institution and the startup simultaneously on the grant application. Because the PI has an ownership interest and potentially a leadership role in the startup, applicants should review pertinent Financial Conflict of Interest (FCOI) policies under NIH regulations (42 CFR Part 50, Subpart F).

11. Must the required industry MPI be at a domestic organization, or can a foreign for-profit hold that role? 

The required industrial partner and its designated MPI can be at a foreign (non-U.S.) for-profit organization. However, proposing a foreign for-profit company to hold this co-leadership role triggers strict NIH administrative rules and review guidelines that you must explicitly address in your application. Please see section above, “Questions Regarding Non-U.S. Entities.”

12. What constitutes a "multi-institutional team"? Could it be cardiology/ophthalmology/AI within the same academic environment, or AI/engineering with clinical teams at the same academic institution? 

The D2M-AIP NOFO outlines a partnership structure designed to bridge gaps in knowledge and expertise by combining academic, industry, and other investigators. The multi-institutional research team described in the NOFO must include at least one academic institution and one for-profit industrial institution.

13. If there is a partnership with the Validation Center, will D2M applicants need to also secure their own external validation sites? 

No, you do not need to list, recruit, or budget for external validation sites as part of your D2M-AIP application. The PRIMED-AI program uses a centralized, cooperative infrastructure specifically designed to eliminate the need for individual developers to find and fund their own external validation sites.

14. Are D2M-AIP awardees required to submit models or data to the Validation Center, and what access or rights does that entail? Are proprietary code and models owned by the industrial partner required to be released publicly, and how will proprietary models be validated by the Validation Center?

Close collaboration with the Validation Center is required throughout Phase 2 (UH3), including: comprehensive sharing of relevant data (consistent with approved data-sharing plans and IRB protocols) and the sequestration strategy; strict adherence to the Validation Center's standardized evaluation protocols; systematic use of shared benchmark datasets for comparative analysis; and regular submission of CDS tools for independent validation to ensure objective assessment of robustness, fairness, and performance. The industrial partner is not required to release proprietary source code or model weights to the public, but must release them to the Validation Center for this independent evaluation. Three Reporting and Sharing Milestones apply to all awardees: Year 1 — begin working with the PRIMED-AI Logistics Center to establish IP-management and data/model-sharing agreements with the Validation Center; Year 2 — sequester data and share models for independent validation; Year 3 — initial submission of the CDS tool to the Validation Center for iterative, independent testing.

15. Will Intramural Research Program costs will be counted against the $450,000 direct cost cap for the UG3 phase and the $800,000 direct cost cap for the UH3 phase?

No grant funds will be used for NIH intramural scientist salary support or the operational costs of NIH intramural facilities. However, cost of IRP reagents, consumables and data analysis (e.g. software licenses) should be submitted as a separate budget page in the budget section. These costs may include salary for staff to be specifically hired under a temporary appointment for the project. The IRP budget may also include consultant costs, equipment, supplies, travel, and other items typically listed under other expenses. These costs would constitute consortium costs and count against the direct cost cap.

16. What level of engagement with the FDA would you expect to happen within the 5 year timelines? A Q-sub or 510(k)/De Novo?

A Q-sub is anticipated by the end of the project period.

17. Can the final SaMD product be a specific application (for a particular vendor etc.) or does it need to be generic that can work with multiple vendors?

It is anticipated that the  final SaMD tool should emphasize broad interoperability and compatibility across different platforms. You proposal should describe how the tool has the potential for use in different  systems and address source-specific variables such as diverse electronic health record (EHR) vendors and systems.  That said, the NIH does not weigh in on IP considerations or have a say in the eventual commercialization strategy.

1. What is the announcement number of this NOFO?

RFA-RM-27-013

2. What is the purpose of this announcement?

The purpose of this NOFO is to catalyze the translation of Artificial Intelligence (AI)-enabled, image-centered, multimodal CDS tools, developed as Software as a Medical Device, from training and testing of prototypes towards clinical applications that address unmet health challenges in precision medicine. These projects are expected to have high potential for demonstrable, positive impact on patient outcomes and/or healthcare processes.

3. What is the focus of Phase 1 (the UG3 phase)?

The UG3 phase is expected to focus on the robust development and technical validation of CDS tools, preparing them for subsequent clinical application. This phase will last up to 2 years.

4. What is the focus of Phase 2 (the UH3 phase)?

Following successful transition, the UH3 phase is expected to focus on gathering compelling evidence of the CDS tool's real-world clinical utility, validating its feasibility for integration into the intended clinical workflows, and demonstrating its positive impact on patient outcomes and/or healthcare processes, thereby catalyzing its clinical adoption.

5. How many awards will be funded?

The program anticipates funding six to eight awards.

6. What is the award budget?

$450,000 in direct costs per year for the UG3 phase and $1,000,000 in direct costs per year for the UH3 phase.

7. When are applications due?

The receipt date is October 19, 2026. All applications are due by 5:00 PM local time of applicant organization. 

8. To whom can I reach out with questions about this announcement?

All questions are welcome. Please reach out to [email protected] and indicate which funding opportunity your question relates to.

What are the key dates in the timeline for planning our application?

1. Are Letters of Intent (LOI) required?

No, NIH no longer requests or accepts Letters of Intent.

2. Are clinical trials allowed?

While not required, applicants may propose to conduct clinical trials if appropriate and feasible within the project period. If a trial is proposed, the applicant must thoroughly describe it and provide strong justification for their chosen design. The work supported through an M2C award that includes clinical trial(s) shall be focused on the performance and effectiveness testing of the developed CDS tools in the real-world clinical workflow, and not on trial objectives outside of M2C scope (e.g., costs related to testing a therapeutic). M2C trials could be either de novo dedicated prospective trials or leverage ongoing or planned trials to assess CDS performance as secondary objectives, integrating with existing infrastructure and data streams.

3. How do I submit my application?

Organizations must submit applications to Grants.gov (the online portal to find and apply for grants across all Federal agencies). Applicants must then complete the submission process by tracking the status of the application in the eRA Commons, NIH’s electronic system for grants administration. NIH and Grants.gov systems check the application against many of the application instructions upon submission. Errors must be corrected and a changed/corrected application must be submitted to Grants.gov on or before the application due date and time.  If a Changed/Corrected application is submitted after the deadline, the application will be considered late. Applications that miss the due date and time are subjected to the NIH Grants Policy Statement Section 2.3.9.2 Electronically Submitted Applications. 

Applicants are responsible for viewing their application before the due date in the eRA Commons to ensure accurate and successful submission.

4. Is there a special structure required for the 12-page research strategy? Should it be organized by the five pillars, or do the standard Significance/Innovation/Approach sections still apply? 

The traditional Significance, Innovation, and Approach sections remain the primary structure, but should be explicitly mapped to the five pillars: Significance → clinical value proposition of the SaMD; Innovation → novel data integration and model structures; Approach → pillars 1–4 explicitly (data quality/governance, clinically grounded technology, pathway to implementation/adoption, integrated multidisciplinary team). Pillar 5 (error mitigation and technical management) is a separate required attachment and does not count toward the 12-page limit.

5. Is this one application that includes both the UG3 and UH3 phases together, or do we submit the UG3 section now and the UH3 later once milestones are reached? 

You must submit a single, combined application including both the UG3 and UH3 phases together within one 12-page Research Strategy, along with a clearly defined list of quantitative milestones you will achieve by the end of the UG3 phase to demonstrate technical feasibility. You do not submit a new grant application for the UH3 phase. As you near the end of Year 2, NIH will conduct an administrative/programmatic review of your progress. If you have met your pre-negotiated UG3 milestones and remain scientifically aligned with consortium goals, your award will be prioritized and administratively transitioned to the UH3 phase.

1. In the UH3 phase of M2C, is prospective collection of new data appropriate, or is the emphasis mainly on deployment/evaluation using existing data and workflows? 

Yes, it is appropriate and expected. M2C trials could be either de novo dedicated prospective trials or leverage ongoing/planned trials to assess CDS performance as secondary objectives. The emphasis is on testing the model in real-world clinical workflows, which inherently involves prospective data as the tool is deployed.

2. How far does the validated model in M2C need to be — could a model validated in a different, limited use case be transferred to a different use case for this proposal?

This presents a risk. The NOFO states that "Projects without an existing prototype of an AI-powered, image-centered, multimodal CDS tool... will not be reviewed". If you are pivoting to a completely new clinical use case where you have zero retrospective technical validation for that specific problem, your application may be considered non-responsive. You should ideally have at least retrospective validation for the specific clinical problem you're proposing. Otherwise, you're probably a better fit for D2M-AIP than M2C.

3. What are the requirements for clinical sites and workflow development in M2C? Does it have to be multi-site (given cost concerns above $1 million), and would a single-center prospective study be considered responsive for the UH3 phase, or must it be multi-site?

Yes, M2C requires multiple sites, and a single-center prospective study is unlikely to be deemed fully responsive for the UH3 phase. The NOFO specifically mandates multi-site clinical validation to demonstrate generalizability of performance across sites in real-world settings. The tool must be interoperable across sites, with a plan for a clinical workflow that works as a common denominator across them. An industrial-partnership component can also be added to M2C and may supplement NIH's budget with additional funds (not expected, but possible). In general, less M2C funding goes toward model development, leaving more available for multi-site validation, done in collaboration with the validation center.

4. Do we need to have a clinical trial unit (CTU) or clinical research organization (CRO) involved as a partner for M2C? That could be expensive given financial constraints.

No, a formal CTU or CRO is not required. The NOFO explicitly states: "While not required, applicants may propose to conduct clinical trials if appropriate and feasible within the project period."  You must conduct clinical performance testing in the real world during the UH3 phase, but this does not mandate an expensive formal clinical trial unit, provided your multi-disciplinary team has the logistical and scientific rigor to evaluate the tool's effectiveness in clinical workflows.

5. Can the disease entity be a rare disease or rare cancer? If so, and prospective/real-world validation must use a smaller sample size, what is the criteria for success given limited statistical power for the M2C NOFO? 

Yes, rare diseases are acceptable. The NOFO's peer review criteria specifically instruct reviewers to evaluate "whether the results will be generalizable or, in the case of a rare disease/special group, relevant to the particular subgroup." Success criteria rely on a well-justified approach: applicants must rigorously justify why the sample size is sufficient for that specific rare population and ensure the analytical framework is robust enough to handle limited data without overfitting.

6. Post-deployment model performance monitoring is mentioned under both the M2C NOFO and the Validation Center NOFO. Who is responsible for developing the monitoring tools? 

You, the M2C applicant, are responsible for developing the monitoring pipeline for your specific tool. The NOFO requires that "All projects must include a plan for a system to track model performance and reliability. This plan must describe a post-deployment monitoring pipeline ready for clinical settings". While you will sequester data and share models for independent evaluation by the PRIMED-AI Validation Center, the core continuous monitoring framework (e.g., AI run charts) must be integrated into your project's UH3 deployment strategy.

7. The NOFO states that independent Validation Center evaluation should be conducted on a sequestered retrospective dataset. Would a prospectively accrued but subsequently sequestered dataset be acceptable for the required Validation Center evaluation?

A prospectively accrued dataset, once completed and sequestered, can be considered for independent Validation Center evaluation, provided the outcome data remains locked and is not used for model tuning prior to the evaluation. Applicants should ensure the dataset strictly adheres to the NOFO’s requirements for independence and sequestration.

8. Can the clinical workflow expert (for M2C) be a clinician? Can such a clinician be hired externally if the academic/hospital partner lacks that expertise? 

Yes, a clinician can serve as your clinical workflow expert. In fact, the NOFO requires a "Core Multidisciplinary Expertise" which includes "Implementing clinical workflows for healthcare systems" alongside AI, medical device development, and relevant clinical domain knowledge. You are permitted to hire this expertise externally. The NOFO explicitly calls for a "multi-institutional research team" to assemble a critical mass of expertise, meaning you can subaward or collaborate with external academic or hospital partners to fill this gap.

9. For M2C, is the commercialization/licensing partner expected to be identified in the application, or can partner identification be a project activity in the UG3 or UH3 phase? 

The expectation is that the commercial and business component will be identified in the application itself. All expertise — clinical workflow, IP, AI, etc. — must be described in the proposal. Note D2M-AIP and M2C requirements are not identical: M2C requires a clinical workflow expert, while D2M-AIP requires an industrial partner (not present in the M2C NOFO).

10. Can a company apply for M2C without FDA approval or an IDE already in place, and can M2C funding be used for the FDA application/clearance process?

Yes, a company can apply without FDA clearance or an Investigational Device Exemption (IDE) already obtained. FDA interaction is typically expected as a milestone in the UG3 phase, which is specifically designed to support regulatory preparations and clinical implementation infrastructure. A pre-submission consultation (pre-sub) to identify the regulatory pathway (De Novo, 510(k), or PMA) is encouraged while the design is still changeable. If an IDE is needed, applicants should outline a clear strategy and timeline for securing it with their collaborating company as part of their UG3 milestones. NIH grant funds cannot be used for FDA regulatory submission or application user fees.

11. The grant guidelines state that a validated multimodal AI prototype is required. Would a proposal be considered responsive if the model has full external validation (supported by publications) for the imaging component, but only internal validation for the multimodal component? If so, what specific documentation is required to demonstrate internal validation of the multimodal components?

Yes, this would be considered responsive. The M2C guidelines define a technically validated AI prototype as a model that has undergone technical validation at a minimum of a single site or unimodally. Therefore, having full external validation for the imaging component alongside internal validation for the multimodal component meets the baseline eligibility. Ensure your application clearly documents the retrospective datasets and technical performance of this internal validation.

12. Is it acceptable under this NOFO for licensed/proprietary inputs to remain required inputs to the CDS tool, provided licensing, implementation feasibility, and cost-effectiveness are addressed and that the associated costs are demonstrated to be reasonable for clinical deployment?

Yes, it is acceptable for the CDS tool to include licensed or proprietary inputs as required inputs, as long as you have addressed the licensing requirements, feasibility, and cost-effectiveness in your application. Demonstrating that the associated costs are reasonable and manageable for clinical deployment is important for review (show that this is scalable beyond your sites). 

1. What is the distinction between the Data-to-Models: An Academic-Industrial Partnership (RFA-RM-27-012) and the Model-to-Clinic (RFA-RM-27-013) NOFOs?

The Data-to-Model Academic-Industrial Partnership (D2M-AIP) NOFO is intended for applications from academic and industrial partners proposing commercially-driven projects. The primary innovation for D2M-AIP projects lies in the novel integration of complex datasets and the development of new AI models, followed by rigorous analytical validation and performance testing. These projects are not required to conduct clinical validation studies within the award period, though they may propose to do so. The goal is to develop and de-risk novel AI technologies through academic-industrial collaboration to a stage where they are primed for subsequent commercialization and translation. 

The Model-to-Clinic (M2C) NOFO is intended for applications that explicitly focus on broad adoption of AI models into practical clinical use, with a primary emphasis on the clinical endpoint. M2C projects will start with a promising AI model and must assess and validate the PRIMED-AI CDS tool's clinical performance, utility, and impact in real-world healthcare settings. The projects should also address other challenges arising from adoption of the PRIMED-AI CDS tools in clinical workflow and associated adoption cost and feasibility.  M2C projects do not require an industrial partner or a commercial driver, though collaboration with end-users and potential dissemination partners is strongly encouraged. 

2. M2C seems like an extension of D2M-AIP without the mandatory partnership between industry and academia. Can I apply to both D2M-AIP and M2C?

You can apply to both NOFOs, but those applications must be scientifically distinct. An NIH applicant generally cannot send the same application (or highly overlapping applications) to multiple NOFOs at the same time. The NIH prohibits duplicate applications under review simultaneously to ensure fair, independent evaluations.

• Choose the D2M-AIP NOFO if your project: Is commercially-driven and led by an academic-industrial partnership; has its primary innovation in novel data integration and/or new AI model development; and will focus on analytical validation and performance testing, without a requirement for clinical validation studies.     
• Choose the M2C NOFO if your project: Has a primary focus on assessing clinical adoption and clinical impact; will translate a promising AI model into a clinical workflow; and will conduct required clinical validation studies to evaluate the tool’s real-world utility and feasibility for adoption in the real world clinical care environment.                                                                                                                                                       

3. Is it problematic for a D2M or M2C project and the Validation Center to be at the same institution, in terms of independent validation? 

Applicants from the same institution may apply to all three of the aforementioned NOFOs. Nonetheless, the Validation Center (VC) is a consortium-wide resource and funded as an independent, centralized third-party hub. Its sole programmatic purpose is to provide objective, unbiased "third-party verification, validation, and uncertainty quantification." To this end in the event of D2M-AIP or M2C projects taking place at the same institution as the VC, the PRIMED-AI Steering Committee and NIH Program Staff will need to negotiate protocols to ensure this independent validation and eliminate any potential local institutional bias or data leakage between the VC and D2M or M2C project.

4. The M2C NOFO states that projects without an existing prototype of an AI-powered, image-centered, multimodal CDS tool are non-responsive. What level and type of prior evidence qualifies a model as a "validated prototype" for M2C — is retrospective validation enough, does it need to be published, tested in large populations, or demonstrate real-world clinical efficacy? How much additional model refinement is appropriate in the UG3 phase of M2C?

This determines which NOFO the project fits. If the model doesn't yet exist and the underlying data is available, that's D2M-AIP territory (more "blue sky," starting from data integration). A "technically validated AI prototype," defined as a model that has undergone technical validation at a minimum of a single site, unimodally, or in retrospective data.

There are two different thresholds based on the phase in M2C: Entry Requirement (UG3): You must have a "technically validated AI prototype, defined as a model that has undergone technical validation at a minimum of a single site, or unimodally, or in retrospective data". You do not need large-population clinical efficacy evidence prior to applying as generating that multi-site clinical evidence is the precise goal of the UH3 phase. Entry Requirement (UH3): This requires compelling evidence of real-world utility, meaning prospective clinical studies/evaluation in actual clinical workflows to demonstrate positive impact on patient outcomes or healthcare processes.

At bare minimum for either phase, the model must be image-centered and able to make a CDS prediction on the population of interest; adding modalities beyond that counts as model refinement, not developing a new model. Borderline cases are decided case-by-case. Applicants are strongly encouraged to reach out to program officers directly, though their internal read doesn't bind reviewers.

5. What qualifies as the required clinical imaging "anchor"? Does this include clinical photographs (e.g., of skin conditions), 3D wound imaging, or physiologic waveform data like EKG traces and arterial blood pressure signals?

Clinical imaging is defined as any FDA-approved imaging modality used in patient care — radiologic (radiographic, CT, MRI, molecular, radionuclide), ophthalmologic (e.g., OCT), endoscopic, dermatologic imaging, and video. Photographs can qualify as the anchor (particularly for skin conditions) as long as photography is the accepted clinical standard for the disease in question; if it is not yet the clinical standard, there's more uncertainty and FDA involvement is often required, though the modality could still be used as a non-anchor input. The same logic applies to 3D wound imaging — it can serve as the anchor if photography/imaging is the established standard for that use, or another standard imaging modality (e.g., vascular ultrasound) can serve as the anchor while 3D wound imaging is a complementary modality. EKG traces and invasive waveforms (e.g., arterial BP) are not considered images and cannot serve as the anchor, regardless of how they're visualized or transformed. Not everything that can be turned into a picture counts as imaging for this purpose. Contact program staff for specific cases.

6. What role are industry partners allowed to play in the development of the proposal, protocol, and clinical trial?

Industry partners are permitted and strongly encouraged to play an active role across both D2M-AIP and M2C opportunities. For D2M-AIP, an academic-industrial partnership is explicitly required, meaning teams must include at least one academic and one industry organization. For M2C, your industry partners can actively participate in refining the models, driving regulatory submissions, and establishing the deployment strategy. While standard NIH compliance rules apply, the PRIMED-AI program is fundamentally designed to support these cross-sector partnerships to accelerate the adoption of clinical decision support tools into real-world healthcare settings.

7. How should applicants balance foundational methodological AI innovation with the program's emphasis on validation, usability, and translation? Is substantial new AI-method development appropriate, or are these mechanisms primarily for advancing mature technologies?

These NOFOs are primarily translational in nature. Novel algorithmic architectures are welcome, but must not be developed in a clinical vacuum — the ultimate goal is advancing technologies toward clinical implementation, independent validation, and regulatory authorization. 

8. Is the requirement to show "verifiable access to clinical-grade imaging data" required at proposal submission, or just verified access by project start date?

This must be demonstrated at submission. Reviewers need to be able to evaluate whether the proposal is ready to go at the time of review.

9. What role will the PRIMED-AI Validation Center play? Does it mean radiology/imaging data will feed into a validation center where the AI model is hosted, or do sites supply data manually? How do D2M and M2C awardees interact with it? Is validation retrospective, prospective real-time, or in silent mode?

The Validation Center is a standalone U54 NOFO (RFA-RM-27-014). It is responsible for continuously sequestering and annotating independent multimodal data for validation and benchmarking, which must be distinct from the data D2M and M2C sites use for model training. No validation center has been selected yet; the validation center and the D2M/M2C projects will need to work out validation approaches together in year one. Options include the funded project setting aside data for the validation center to test the model, or the validation center providing outside datasets for the project to validate against. This will need to be discussed in the first year of the projects.

10. How will the proposals be reviewed, by a standing or ad hoc study section?

Applications will be reviewed by a special emphasis panel convened explicitly for these NOFOs, assembled by the Center for Scientific Review (CSR) at NIH in response to the expertise needed to review the applications received. Reviewers will have knowledge of the funding opportunity itself and will evaluate applications on the standard review criteria (Factor 1: Importance of the Research, Factor 2: Rigor and Feasibility, Factor 3: Expertise and Resources), plus NOFO-specific additional review criteria. Both NOFOs share review criteria, but they are not identical — check Section V of each NOFO for specific details.

11. How are the applications being selected? How will all the awardees integrate and work together? 

Applications will be selected based on review scores, programmatic priority, and addressing key gaps. There will be many opportunities to collaborate — each D2M-AIP and M2C awardee will have the opportunity to meet with the validation center and develop a plan for what validation means for their specific project. Applicants should review the cooperative agreement terms in each NOFO, which describe the requirements of these collaborative interactions.

12. Who would be the optimal translational partners? For example, drug companies, imaging companies? 

The optimal partner is determined by the program's ultimate goal: advancing clinical decision-support AI toward regulatory clearance (FDA 510(k) or De Novo) as a Software as a Medical Device and integrating it into real-world clinical workflows. NIH does not prescribe who is the most optimal partner, as this is inherently dependent on the specific application.

13. The FDA has a specific definition of CDS tools. Is the term "Clinical Decision Support" for PRIMED-AI intended to mean the same as the FDA device class?

Yes, the program was designed to align with the FDA definition of CDS tools. However, NIH does not evaluate whether applications align with FDA guidance. This is why both NOFOs recommend discussion with the FDA to determine alignment.

14. If an M2C or D2M-AIP applicant's model relies on imaging or data preprocessing standards that differ from those later developed by the Playbook or Validation Center, could this jeopardize the project's transition from UG3 to UH3?

Your UG3 to UH3 transition will not be jeopardized by concurrent developments from the Playbook or Validation Center, as NIH transition determinations are strictly governed by your project's ability to achieve its pre-negotiated, quantitative Go/No-Go milestones, regulatory readiness, and clinical workflow integration. Within the PRIMED-AI Consortium architecture, the Playbook and Validation Center develop generalized best-practice frameworks and conduct independent robustness testing/uncertainty quantification rather than imposing top-down, retroactive mandates that would invalidate locked, validated Clinical Decision Support (CDS) pipelines. Methodological differences in data or imaging preprocessing are resolved collaboratively through Consortium working groups and treated as comparative robustness benchmarks, ensuring that mature M2C projects advance to UH3 funding based on their demonstrated analytical and clinical performance.

15. What happens if the Validation Center does not currently have the capability to obtain the specific modalities or population of interest needed to validate a given model? Are there contingency plans if an applicant's own external sites can provide the necessary labeled outcome data and datasets, but the Validation Center cannot?

Third-party, independent validation is a fundamental principle of the PRIMED-AI program, and this scenario is addressed through built-in administrative and structural contingency mechanisms rather than an outright barrier to transition. The Validation Center works collaboratively with D2M-AIP and M2C awardees prior to independent validation, and if a unique dataset is needed, the details of its acquisition and its relationship to the model developer will be considered on a case-by-case basis — with the Validation Center responsible for these determinations and both real and perceived conflicts of interest factored into the decision.

Applicants should identify potential impediments like this in their UG3 phase plans and propose clear alternative approaches and risk-mitigation milestones. During Year 1 of the UG3 phase, teams are expected to actively collaborate with the PRIMED-AI Logistics Center and NIH program staff to establish data-sharing policies, intellectual property frameworks, and standardized evaluation protocols. This collaborative process provides a structured pathway to integrate external validation datasets — such as those from an applicant's own planned multi-site collaborators — into the Validation Center's testing pipeline, onboarding the necessary modalities and labeled outcomes to the program's shared benchmark resources ahead of the UH3 transition.

1. Would creating a workflow that combines various existing software programs to communicate better with each other (interoperability, data pipelining, EHR/PACS integration) be considered M2C or D2M? 

M2C. Creating a workflow that integrates and combines various existing software programs to improve communication (interoperability, data pipelining, and EHR/PACS integration) is likely more responsive to Model-to-Clinic (M2C).

2. Would collecting new data be considered responsive, or must the project rely on existing datasets (public or private)? 

The project does not need to rely exclusively on existing datasets.

3. Would a CDS tool anchored in cardiac MRI or echocardiography, with ex vivo label-free monocyte morphomics as a complementary input, be responsive if its intended use is to predict immune-fibrotic remodeling and/or pathway-specific therapeutic response? 

Yes. The non-negotiable responsiveness criterion under both NOFOs is that the CDS tool must be anchored in clinical imaging as the primary data type — cardiac MRI/echo satisfies that. The ex vivo morphomics input aligns with the mandate to integrate multiscale multimodal data, and the intended use fits the program's precision-medicine goal. Post-mortem/ex vivo imaging can be part of the multimodal dataset but cannot itself serve as the anchor.

4. What capabilities would NIH expect a proposed digital twin to demonstrate beyond conventional longitudinal prediction (e.g., patient state updating, simulation of alternative interventions, uncertainty quantification)? 

Digital twins as described would be responsive. Start from the general CDS definition (assisting clinicians in diagnostic, prognostic, or treatment-assignment decisions). A digital twin should ingest clinical data (one type of which must be clinical imaging), simulate "what-if" scenarios, and calculate uncertainty quantification (confidence intervals, probability distributions). Continuous monitoring/error tracking ties into the five required pillars, particularly the error-mitigation pillar reviewers will scrutinize closely.

5. Could a multi-organ digital twin (cardiovascular, hepatic, adipose, muscle, renal, inflammatory) be responsive if all components support one clearly defined clinical decision, using population-level enriched cohorts (CKD, HIV, long COVID-19)? 

Yes, provided there is a clear clinical connection — the system must support one clearly defined, clinically justified intended use. A broad, generalized model without a clear clinical focus risks being classified as non-responsive. Diverse enriched cohorts are encouraged for testing transportability and generalizability across diverse clinical samples.

6. Would analyzing CT images already acquired for another indication be responsive, and may a hybrid architecture combine image-level learning with quantitative features (e.g., Agatston score, liver attenuation)? 

Yes to both. NIH does not restrict models to raw images only — a hybrid architecture pairing image-level learning with established quantitative biomarkers is a clinically interpretable way to satisfy the multimodal integration requirement.

7. Does integrating electronic health record (EHR) data with imaging data tend to make an application stronger, or is it usually better to keep the model focused on imaging alone? 

Integrating EHR data with imaging data will likely make an application stronger. A model focused on imaging alone is likely to be viewed as non-responsive to the core mandates of the program.

8. For multimodal prediction using imaging and longitudinal EHR data, would you recommend integrating extracted features or learned representations within a single model, or developing separate imaging and EHR machine-learning models and combining their outputs through late-fusion ensemble learning? 

Neither architecture is explicitly mandated or recommended. The NOFOs are intentionally algorithm-agnostic to allow for maximum scientific innovation. Applications are evaluated on functional capabilities, safety, and translational readiness rather than a prescribed machine-learning architecture.

9. Will the CDS/imaging data be limited to the diagnostic stage, or will it cover all stages of clinical decision support? 

Not limited to the diagnostic stage. Proposed CDS tools are explicitly intended to cover a broad spectrum of clinical decision stages, including screening, diagnosis, risk stratification, prognosis, and treatment selection/monitoring.

10. If you have a published unimodal imaging model plus preliminary data showing potential for multimodal integration, is that responsive? Must the multimodal model already be fully developed, validated, and published before submission? 

Yes, this is responsive and represents a potentially strong entry point. The multimodal model does not need to be fully developed, validated, or published before submission — the relative maturity of the model and plans for multimodal integration can be the basis for either D2M-AIP or M2C. Discussion with NIH program staff about specific proposals is highly encouraged.

11. Is a UG3 combining a large retrospective cohort with a smaller prospective multimodal cohort responsive, or must all modalities be present retrospectively?

At bare minimum, a retrospective analysis is expected in the UG3 phase, but a prospective component is welcome and not required to cover all modalities. There is no limit to the kind of proposals that come in — they can combine retrospective and prospective data.