Equity in AI Health Research Institute

Words in [brackets] are blanks you fill in with your own details when you adopt this sample in Supreme Plan AI.

Executive Summary

Equity in AI Health Research Institute is a nonprofit research institute in Maryland that develops and openly releases AI models for clinical decision support, with a specific focus on conditions that affect populations underrepresented in most clinical AI research: rural communities, low-income communities, communities of color and communities with limited access to specialist care. Every model we release is tested across demographic subgroups, trained on datasets that include patients from those communities, and released under an open license with a full model card that describes performance, limitations and known failure modes. Health systems, public health agencies and researchers can use our models without a license fee.

Most clinical AI research uses data from large academic medical centers with patient populations that skew toward urban, insured and white. Models trained on those datasets can perform meaningfully worse for the patients who most need better diagnostic support, and the problem is often invisible until a community health center or a rural hospital deploys the model and measures outcomes. We build and test for the populations that are left out.

In Year 1 we will complete validation of our first model, a chest X-ray triage tool for detecting signs of pneumonia and pleural effusion in resource-limited settings, on a multi-site, demographically diverse dataset, and release the validated model and its model card publicly. Our Year 1 budget is about $949,000, with about $150,000 from federal research contracts, $80,000 from health system and foundation partners and $20,000 from licensing fees on open models. Grants fall from about 74% of revenue in Year 1 to about 41% in Year 5.

We do not claim that our models will improve health outcomes in every setting, and we do not recommend that any health system deploy a model without adapting it to their patient population and clinical workflow and obtaining appropriate regulatory or ethics review. Our role is to build a better starting point and to publish what we find, including failures.

I am Dr. Amara Osei, Executive Director and Principal Research Scientist, an internal medicine physician and machine learning researcher. The numbers in this proposal are estimates based on the assumptions listed at the end, not promises.

Company Description

Equity in AI Health Research Institute is a nonprofit research institute organized as a 501(c)(3) public charity in Maryland. We are governed by an independent board of directors and will provide our IRS determination letter, articles of organization, board list, conflict-of-interest and research ethics policies and most recent financial statements with this proposal. We will confirm eligibility with the Equity in Health Innovation Fund before we accept any award.

Our mission. To build and openly release AI models for clinical decision support that work as well for the patients who are hardest to reach as for those who are easiest to study.

What we believe. Clinical AI trained without diverse data fails the patients who most need it. Open access to validated models and model cards gives community health centers, public health departments and researchers in lower-resourced settings the foundation they need to apply AI responsibly. Research published in the open moves the field faster and creates accountability. We do not claim our models are safe to deploy without clinical review; we claim they are better starting points with better documentation.

How we started. The institute grew from a research collaboration between two physicians, an epidemiologist and a data scientist who were studying why a commercially licensed clinical AI model performed significantly worse in a network of federally qualified health centers than in the academic hospital where it was tested. They investigated the dataset, the model architecture and the implementation context, published the findings, and were contacted by other researchers and health centers with similar experiences. A foundation grant supported a symposium, and the consensus at that symposium was that the field needed an independent institute to build and release models with equity in mind. This proposal asks for support for that institute's first year of full operation.

What we offer:
- Open clinical AI models. We train, validate and release clinical AI models under open licenses, with full model cards that describe training data, validation datasets, performance by demographic subgroup, known failure modes and recommended use conditions.
- Multi-site diverse validation. Every model is validated on data from multiple sites, including community health centers, federally qualified health centers, rural hospitals and academic medical centers, so performance differences across settings are measured and documented.
- Community research panels. People who use the health systems where our models might be deployed serve on a panel that reviews research questions, model outputs and plain-language summaries. Panelists are paid by check or payroll, not cash.
- Health equity analyses. We publish subgroup analyses and investigate causes when performance gaps are found, and we release both the analysis and the code.
- Training and technical assistance. We train health system staff and researchers in how to use, adapt and evaluate our released models, and we provide technical assistance for implementation.
- Policy and regulatory input. We comment on proposed FDA guidance, NIH grant announcements and state policies affecting clinical AI, representing the perspective of equity-focused research.

Who we serve. Health systems, community health centers, public health departments, researchers and policymakers who want open, equity-tested clinical AI models and the evidence behind them. Our models and publications are free to access. Health systems pay for training, technical assistance and custom validation services.

What we are not. We are not a commercial software vendor, and we do not offer models cleared by the FDA for clinical use as a medical device without a health system's own regulatory and ethics review. We do not provide individual clinical care. We do not accept funding that requires us to suppress or modify a finding.

Problem & Opportunity

Clinical AI research has a well-documented dataset diversity problem. The largest and most widely used clinical datasets in the United States come from major academic medical centers that see patient populations that are wealthier, more urban and less racially and ethnically diverse than the country as a whole. Models trained primarily on those datasets may perform well in the settings where they were trained and worse in community health centers, rural hospitals and clinics that serve patients whose imaging, lab values and clinical presentations differ from the training distribution.

The consequences are not hypothetical. Published studies have found performance gaps in commercial clinical AI tools across race, sex, age, body habitus and imaging equipment type. A model that performs 10 percentage points worse in sensitivity for Black patients compared to white patients, deployed at scale, could generate a systematic difference in diagnosis rates that is hard to detect from aggregate statistics and nearly impossible to detect without data stratified by race at the model level.

Independent open research on this problem is scarce. Most clinical AI research is funded by or affiliated with companies that have a financial interest in the model performing well. Academic researchers who want to study performance gaps often cannot access the code or training data of commercial models. Health centers and hospitals that want to evaluate a model before deploying it often lack the technical staff and the validated datasets to do it.

We do not claim to have solved the dataset diversity problem. Any model trained on available data reflects the gaps in that data. Our contribution is to document those gaps honestly, to include more diverse sites in validation, to release the model and the model card together so users know what they are getting and to investigate causes when gaps are found rather than smoothing them into aggregate metrics.

Products & Services

Equity in AI Health Research Institute will develop, validate and openly release AI models for clinical decision support, with equity in design and a full model card for every release, in three phases: model development with a diverse training dataset, multi-site validation with demographic subgroup analysis and open release with documentation.

How the work proceeds:
1. Research question and dataset. We identify a clinical question where an AI model could support decision-making in resource-limited settings. Our first model targets chest X-ray triage for pneumonia and pleural effusion. We license de-identified datasets from multiple sites, including community health centers and rural hospitals, under data-use agreements and IRB approval.
2. Model development and internal validation. We train and validate the model with a held-out test set from each site. We test by age, sex, race and ethnicity, imaging equipment type and site type, and we document performance by subgroup in a preliminary model card.
3. Community panel review. A panel of patients and community members from the health systems we study reviews the research question, a plain-language description of what the model does and the preliminary results, and tells us whether the question matters to them and whether our description of limitations is honest and useful. Their feedback shapes the model card language.
4. Multi-site prospective or external validation. We validate the model at external sites not used in training, with the same subgroup analysis. An independent biostatistician at a partner university analyzes results.
5. Open release. We release the model weights, training code, validation datasets (where consent and data-use agreements allow) and the full model card under an open license, with a clear statement of what the model is not cleared or validated for.
6. Dissemination and technical assistance. We publish the results in a peer-reviewed journal, present at relevant conferences and provide technical assistance to health systems that want to adapt the model to their patient population.

What the grant will pay for. Part of one Clinical Research Scientist's salary, IRB fees and community panel stipends paid by check or payroll, cloud computing and model training, clinical data licensing and curation and open-source publication and dissemination costs. The rest of the budget comes from federal research contracts, health system and foundation partners and licensing fees.

Independence. We accept no funding from a company whose model we are evaluating or whose commercial interest conflicts with our findings. We disclose all funders for every study and project.

Market Analysis

Who benefits: health systems, community health centers, public health departments and researchers who want equity-tested, openly licensed clinical AI models, and, through them, the patients who are too often left out of AI development.

Our primary users are clinical AI researchers, health system data science teams, public health agencies and federally qualified health centers that want a model to start from and adapt, rather than building from scratch. Our secondary users are health system leaders and policymakers who want published equity analyses to inform purchasing and regulatory decisions.

We do not target individual patients directly. Our models may eventually reach patients through the health systems that adapt and deploy them.

Market for open clinical AI. The market for open models in health is small today and growing as FDA, NIH and CMS increase their focus on equity in AI and as health systems face scrutiny over disparate outcomes from deployed AI. We will use NIH grant databases, health system technology budgets and published surveys of health AI adoption in our final application and will not overstate the market.

Revenue model. Health systems and researchers pay for training and technical assistance, and we earn licensing fees when organizations use our released models commercially with a service layer (the model itself remains free). Foundation and health system partners contribute to our research agenda. Federal research contracts cover specific studies.

Competitive Landscape

Other sources of clinical AI models and equity research, and how we are positioned.

- Commercial clinical AI vendors. They offer cleared tools with vendor support and proprietary models. They rarely publish subgroup analyses or release model weights. Our open models and published equity analyses complement and, over time, can inform commercial products.
- Academic research groups. They publish clinical AI studies and sometimes release models. They rarely have resources for multi-site validation, community panel engagement or technical assistance. We partner with them.
- NIH and government datasets and models. NCI, NLM and other agencies release datasets and, increasingly, models. These are valuable and rarely include the implementation support and equity focus of our work.
- Open-source medical AI repositories. Resources like PhysioNet and HuggingFace's medical AI models are growing. They vary widely in documentation and equity analysis. We will release our models there as part of dissemination.

What makes us different.
- Every model trained and validated on demographically diverse, multi-site data.
- A community panel that reviews findings before release.
- A full model card that includes subgroup performance and known failure modes.
- Published equity analyses with the code to reproduce them.
- No vendor funding for evaluations of that vendor's product.

Marketing & Sales Strategy

How we will reach health systems, researchers, communities and funders.

Health systems and community health centers. We will reach health system data science and informatics leaders, CIOs and CMOs through the American Medical Informatics Association, the Healthcare Information and Management Systems Society and direct outreach to federally qualified health center networks. Every partner site that contributes data to a study receives technical assistance and co-authorship credit.

Researchers. We will publish in open-access venues, present at NeurIPS Health, AMIA and MLHC, release code and model cards on standard repositories and actively recruit co-investigators from under-resourced institutions.

Communities. We will hold community panel sessions at partner health centers, publish plain-language summaries in the languages of our patient populations where we have capacity and report back to communities about what we found and how it affects them.

Funders. We will report to funders twice a year and publish our methods and results publicly. We disclose all funders for every project and manage conflicts of interest under written policy.

Policymakers and regulators. We will comment on FDA, NIH and CMS guidance on AI in health, representing the equity perspective and drawing on our published evidence.

Operations Plan

How the work will be carried out.

Timeline for Year 1:
- Months 1 to 3: Hire the Clinical Research Scientist, set up the IRB submissions and data-use agreements for the multi-site chest X-ray study, set up the secure HIPAA-compliant data environment and convene the community research panel.
- Months 4 to 6: License and curate training data from multiple sites, train the model and conduct internal validation with subgroup analysis.
- Months 7 to 9: Run external validation at two additional sites. Community panel review of preliminary findings and model card language.
- Months 10 to 12: Release model version 1.0 with model card, submit for peer review and provide technical assistance to the first two health system partners.

Infrastructure. We run a HIPAA-compliant data environment with encrypted storage, role-based access and audit logs, backed by a business associate agreement with our cloud provider. Every dataset is used under a signed data-use agreement and IRB approval. Cloud computing and model training are about $50,000 in Year 1. Security, HIPAA and compliance tooling is about $18,000.

IRB and research ethics. Every use of patient data is covered by an IRB-approved protocol and a data-use agreement. Community panel participation follows informed consent procedures. An independent research ethics review confirms our protocols meet the standards for research with human subjects data that apply to our work.

Community panel. About ten community members from the health systems and communities we serve review our research questions, preliminary findings and model card language. Panelists are paid stipends by check or payroll, not cash, and may withdraw at any time. Their feedback is documented and influences the model card.

Open release policy. Model weights, training code and, where data-use agreements allow, validation datasets are released under an open license. Model cards include performance by demographic subgroup, training data description, known failure modes and recommended conditions of use. We do not release a model without a model card, and we do not claim a model is safe for clinical deployment without the health system's own regulatory review.

Independence and conflicts. We accept no funding from a company whose model we are evaluating, and we disclose all funders for every project. Conflicts of interest for staff and board are managed under written policy.

Legal and insurance. Our attorney reviews data-use agreements, IRB submissions and open-source licensing terms. We carry general liability, professional liability, cyber liability and directors and officers coverage. Insurance is about $16,000 a year in Year 1. An outside accountant reviews our books every year.

Management & Organization

The people who will carry out the project.

Dr. Amara Osei, Executive Director and Principal Research Scientist. An internal medicine physician and ML researcher with published work in algorithmic fairness in clinical AI and experience at an FQHC and an academic medical center. She leads the institute, its research agenda and its partnerships. (Replace this paragraph with your own director's background.)

Principal Research Scientist (AI and Health). A machine learning researcher with experience in medical imaging or clinical NLP, who designs the model architecture, runs training and validation and leads the technical work.

Clinical Research Scientists (one in Years 1 and 2, two from Year 3). Physicians or clinicians who design the clinical evaluation, manage IRB submissions and data-use agreements, oversee data curation and interpret clinical findings.

Community Engagement and Health Equity Lead (one in Years 1 to 3, two from Year 4). Convenes and coordinates the community panel, manages community communications and ensures that equity is considered at every stage.

Biostatistician. Designs the statistical analysis plan, conducts independent analysis and co-authors publications.

Research Coordinator. Manages IRB submissions, data agreements, panel logistics and grant reporting.

Board of directors. Seven independent directors who meet quarterly, including a community health center medical director, a public health researcher, a health policy attorney, a bioethicist, a patient advocate and two health system CIOs. No board member may be employed by a company whose product we evaluate.

Community research panel. About ten community members who review our research questions and findings before public release, paid by check or payroll.

Staff pay in this plan includes payroll taxes and benefits. Salaries are Executive Director $130,000, Principal Research Scientist $170,000, Clinical Research Scientist $120,000, Community Engagement Lead $90,000, Biostatistician $130,000 and Research Coordinator $60,000.

Milestones Narrative

Our goals for each year and how we will measure them.

- Year 1: Release validated model version 1.0 with a full model card on a multi-site, demographically diverse dataset, complete community panel review and submit for peer review. A budget of about $949,000 in revenue, with a surplus of about $23,600. 6 paid staff.
- Year 2: Release version 2.0 incorporating community panel feedback and external validation, and begin the second model development cycle targeting a new clinical question. A budget of about $971,500 in revenue, with a surplus of about $23,200. 6 paid staff.
- Year 3: Add a second Clinical Research Scientist, publish three peer-reviewed papers and provide technical assistance to at least five health system partners. A budget of about $1,119,500 in revenue, with a surplus of about $24,900. 7 paid staff.
- Year 4: Add a second Community Engagement Lead, grow the federal research contract base and expand to a second open model release with community panel review in three languages. A budget of about $1,240,000 in revenue, with a surplus of about $26,200. 8 paid staff.
- Year 5: Sustain the program with federal contracts, partner fees and licensing revenue, with grants covering a minority share of the budget. A budget of about $1,273,000 in revenue, with a surplus of about $25,600. 8 paid staff.

What we measure:
- Model releases with a full model card and open license.
- Subgroup performance gaps documented and, where found, investigated and reported.
- Community panel participation and the share of panel feedback that changes a model card or research question.
- Peer-reviewed publications and open dataset or code releases.
- Health system and community health center technical assistance engagements.
- Federal and foundation research contracts signed.
- Equity: the share of validation data from community health centers, FQHCs and rural hospitals.

How we review results. The Principal Research Scientist and the Executive Director review research progress every two weeks. The board reviews scientific progress, independence and finances quarterly. The community panel reviews findings before each release. We publish our methods and results regardless of direction.

Funding Request

What we are asking for, and how the money will be used. Our request: $150,000 from the Equity in Health Innovation Fund for one Clinical Research Scientist, IRB fees and community research panel stipends, cloud computing and model training, clinical data licensing and curation and open-source publication and dissemination, for July 2027 through June 2028.

The request is spread across these budget lines:
- Personnel: Clinical Research Scientist salary, $60,000 (other funds add $60,000).
- Other: IRB fees and community research panel stipends (paid by check or payroll, never cash), $28,000.
- Other: Cloud computing and model training, $30,000 (other funds add $20,000).
- Contractual: Clinical data licensing and curation, $20,000 (other funds add $4,000).
- Other: Open-source publication and dissemination costs, $12,000.

Other money in the program. Federal research contracts, health system and foundation partners and licensing fees pay the other $84,000 of those lines, plus the Executive Director, Principal Research Scientist, Biostatistician, Community Engagement Lead and Research Coordinator salaries, office rent and security and compliance costs.

Year 1 budget: about $949,000 in revenue. Year 1 money coming in: grants $699,000 (this request plus other grants we are applying for), federal and state research contracts $150,000, health system and foundation partners $80,000 and licensing fees from open models $20,000. Year 1 costs: Executive Director $130,000; Principal Research Scientist $170,000; Clinical Research Scientist $120,000; Community Engagement and Health Equity Lead $90,000; Biostatistician $130,000; Research Coordinator $60,000; cloud computing and model training about $50,000; irb fees and community research panel stipends about $28,000; clinical data licensing and curation about $24,000; office rent and utilities about $28,000; open-source publication and dissemination costs about $10,000; security, hipaa and compliance tooling about $18,000; insurance about $16,000; legal and ethics review about $20,000; travel and community convenings about $14,000; administration and accounting about $14,000; and about $8,200 a year for the wear on equipment.

About federal research contracts. NIH (NCI, NIDCD, NIMHD, NLM), AHRQ and PCORI fund health equity, clinical AI and comparative effectiveness research. NSF, NIST and DARPA fund AI fairness and trustworthy AI. These are competitive, peer-reviewed programs that reward rigorous methods and equity focus. We plan to apply and do not count specific awards until they are received. The amounts in this plan are our planning estimates.

One-time startup costs. Startup costs of $59,500 are paid once from startup money, not from the yearly budget: $17,000 for equipment, $6,000 for leasehold improvements, $6,000 for deposits, $8,000 for opening supplies and inventory, $1,000 for licenses and permits, $12,000 for legal and professional help, $4,000 for signage, brochures and a website, $5,500 for background checks, training and first insurance.

How we got our startup number. The Funding Need Calculator in this plan adds the startup costs ($59,500), the cash we need to cover the wait for the first payments (about $39,400) and a cash cushion of about $230,500, which is about three months of Year 1 costs. That comes to $329,500. We start with $40,000 of our own savings and $300,000 in startup support from a health equity foundation, a health system innovation fund and our board, $340,000 in all, which covers that amount.

The next five years. Revenue grows from about $949,000 in Year 1 to about $1,273,000 in Year 5. Grants fall from about 74% of revenue in Year 1 to about 41% in Year 5 as federal research contracts and licensing fees grow.

Risks and how we will handle them:
- A study finds large performance gaps. We document, investigate and report them fully, and we do not release a model without documenting known gaps.
- A data-use agreement or IRB amendment is delayed. We stage the research so early phases can proceed while agreements are finalized.
- A funder pressures us to change a finding. We decline, and our contracts with funders state that we control all publications.
- A community panel member raises an ethical concern about the research. We pause, address the concern and decide whether to proceed.
- Federal contracts are delayed or not awarded. We plan with a cash cushion and apply to multiple sources.

Savings. We plan to keep several months of costs in cash. The lowest year-end cash balance in the plan is about $283,600, at the end of Year 1. Customers and funders often pay after we invoice or submit reports, so we plan for about $66,430 owed to us at the end of Year 1.

If you are just starting. A new institute can begin as a project inside a fiscal sponsor, publish an initial equity analysis of a publicly available dataset and release the code, and apply for an NIH R01 or R21. Strong published equity analyses and a community partner make larger awards more competitive.

Appendix Notes

Documents to attach with this proposal:

- Your IRS 501(c)(3) determination letter, or your fiscal sponsor agreement.
- Your most recent financial statements, Form 990 and current year budget.
- A list of board members, their roles and any conflicts of interest.
- Resumes for the Executive Director and Principal Research Scientist, with publications.
- Your research ethics committee or IRB approval letter for the chest X-ray study.
- Your data-use agreement template and HIPAA security policies.
- Your open release policy, including model card standards and licensing terms.
- Your conflict-of-interest policy and funder disclosure policy.
- Your community panel charter and consent procedures.
- Results from prior equity analyses or publications, if any.
- Proof of general liability, professional liability, cyber liability and D&O insurance, or quotes.
If you don't have any of this yet, no worries - let's continue.

The assumptions behind our numbers (change them in the Financials section to match your own):
- Grant Revenue: $699,000 in Year 1, then $616,500, $639,500, $628,000, $521,000.
- Federal Research Contracts: $150,000 in Year 1, growing to $460,000.
- Health System and Foundation Partners: $80,000 in Year 1, growing to $182,000.
- Licensing Fees from Open Models: $20,000 in Year 1, growing to $110,000.
- Pay: Executive Director $130,000, Principal Research Scientist $170,000, Clinical Research Scientist $120,000, Community Engagement Lead $90,000, Biostatistician $130,000, Research Coordinator $60,000, with headcount growing as shown.
- Other yearly costs in Year 1: Cloud Computing and Model Training $50,000, IRB Fees and Community Research Panel Stipends $28,000, Clinical Data Licensing and Curation $24,000, Office Rent and Utilities $28,000, Open-Source Publication and Dissemination Costs $10,000, Security, HIPAA and Compliance Tooling $18,000, Insurance (General, Professional, Cyber and D&O) $16,000, Legal and Ethics Review $20,000, Travel and Community Convenings $14,000, Administration and Accounting $14,000. Growing between 4% and 20% a year.
- Equipment: $17,000 at the start, then $5,000 to $7,000 a year.
- Startup costs: $59,500, paid once from $340,000 in startup money.
- About 7% of each year's revenue is still owed to us at year end.
- Tax: none, as a 501(c)(3) public charity.
- Every dollar amount is an estimate.

Startup Costs