NIH Bridge2AI Stage 2: Inside the $130 Million Push to Make Health Data Actually Work for AI

The dirty secret of health AI is that most biomedical data is useless for machine learning. Electronic health records are riddled with missing fields and inconsistent coding. Imaging datasets lack the standardized annotations that training algorithms require. Genomic data sits in siloed repositories with incompatible formats. Researchers spend 80 percent of their time cleaning and reformatting data and 20 percent doing actual science. The NIH Common Fund's Bridge to Artificial Intelligence program exists to fix this — and on January 29, the NIH Council of Councils approved its second stage with $130 million in new funding over four years (Granted News).

That dollar figure, while substantial, understates the program's significance. Bridge2AI is not funding another round of AI model development. It is building the infrastructure layer — the datasets, standards, ethical frameworks, and training materials — that will determine whether the next decade of health AI produces reliable clinical tools or expensive hallucination machines.

What Stage 1 Built

Bridge2AI launched in 2022 with approximately $130 million for its first four years. The program funded four Data Generation Projects, each tasked with creating AI-ready datasets in domains where machine learning could transform clinical practice but lacked adequate training data:

Voice as a Biomarker of Health focused on collecting and annotating voice recordings to detect conditions ranging from neurological disorders to respiratory disease. The concept — that vocal patterns encode clinically useful information — required building entirely new annotation standards because no existing medical coding system captures voice features in machine-learning-compatible formats.

AI-READI (Artificial Intelligence Ready and Equitable Atlas for Diabetes Insights) assembled a multimodal dataset combining retinal imaging, wearable sensor data, and clinical records for Type 2 diabetes research. The project explicitly addressed representation gaps, enrolling participants from communities historically underrepresented in biomedical datasets.

CHoRUS (Collective Hospital-Based Observations to Research Underlying Syndromes) built datasets from intensive care unit records to improve prediction of patient outcomes. ICU data is notoriously messy — high-frequency physiological signals, inconsistent clinical notes, variable medication documentation — and CHoRUS developed standardization protocols to make it AI-trainable.

CM4AI (Cell Maps for AI) generated multimodal cell maps integrating imaging, proteomic, and genetic data to model how cells function and malfunction. The goal: create training data for AI systems that can predict cellular responses to drugs and disease.

Beyond the datasets themselves, Stage 1 produced software tools, data standards, best practices documents, and a Bridge2AI portal where researchers can access the results. These are not locked in proprietary databases. The program's design mandate requires open access — any researcher can register and use the datasets, tools, and training materials.

What Stage 2 Changes

Stage 2 shifts from data generation to data deployment. The $130 million will fund two complementary tracks:

Innovation Funnels take the AI-ready datasets from Stage 1 — and potentially from other sources — and use them to build actual tools, devices, and clinical insights. This is where the program crosses from infrastructure into application. If Stage 1 was about creating the raw materials, Stage 2 is about proving those materials can produce reliable clinical products.

The funnels are designed as focused problem-solving pipelines: start with a specific health challenge, apply AI methods to Stage 1 datasets, and produce a working prototype — whether that is a diagnostic tool, a predictive model, or a clinical decision support system. NIH has not yet published the specific funding opportunity announcements for these funnels, but the structure suggests they will favor multidisciplinary teams that combine AI expertise with clinical domain knowledge and regulatory awareness.

Network for AI Health Science is the governance and safety layer. This initiative will assemble a network of scientific experts to develop standards for responsible AI use in biomedical research, create safety measures for health AI tools, and build a framework to guide future programs. Think of it as the institutional infrastructure that surrounds the technical infrastructure — the guidelines, evaluation criteria, and ethical guardrails that determine whether an AI tool gets from prototype to clinical deployment.

Who Should Pay Attention

Bridge2AI operates through the NIH Common Fund, which means it crosses institute boundaries. You do not need to be an existing NIH grantee at a specific institute to participate. The program's interdisciplinary design actively seeks researchers from outside traditional biomedical circles.

Health AI researchers with expertise in foundation models, few-shot learning, or transfer learning should watch for Innovation Funnel solicitations. The program's emphasis on using existing datasets (rather than generating new ones) means computational researchers can lead proposals without running their own clinical studies.

Clinical researchers sitting on unique datasets should consider how Bridge2AI's standards and annotation frameworks could make their data more valuable. Even if you do not apply directly, adopting Bridge2AI's data standards positions your institutional data for future AI collaborations.

Bioethicists and responsible AI researchers have a direct path through the Network for AI Health Science. The program needs people who can develop evaluation frameworks, bias detection methodologies, and governance models. This is a rare NIH funding opportunity where humanities and social science expertise is structurally required, not bolted on as broader impacts.

Medical device and health tech companies should track the Innovation Funnels closely. Stage 2 bridges the gap between research dataset and commercial product. Companies with FDA experience, clinical trial infrastructure, or deployment platforms could be valuable partners on Innovation Funnel proposals — and early involvement positions them for licensing or partnership opportunities downstream.

Data scientists and ML engineers working on data quality, annotation tools, or data harmonization have natural alignment with Bridge2AI's core mission. The program's emphasis on making data AI-ready creates demand for tools that detect bias, fill gaps, standardize formats, and validate quality.

Why This Program Matters More Than Its Budget

At $130 million over four years, Bridge2AI is a rounding error in NIH's $48.7 billion annual budget. The program's influence, however, extends far beyond its direct funding.

First, Bridge2AI is establishing the standards that other NIH programs will adopt. When NINDS, NCI, NHLBI, or any other institute funds AI-related research, they increasingly require data to meet interoperability and annotation standards. Bridge2AI is producing those standards. The datasets it creates become reference implementations — the models that other programs point to when they say "make your data look like this."

Second, the program solves a coordination problem that no single institute can address. Health AI requires combining data types that span institutional boundaries — genomic data from NHGRI, imaging from NIBIB, clinical data from multiple disease-specific institutes. The Common Fund is the only NIH mechanism designed for cross-cutting infrastructure of this kind.

Third, Bridge2AI creates public goods in a landscape dominated by private AI companies training models on proprietary health data. Every dataset Bridge2AI produces is accessible to academic researchers, small companies, and international collaborators. In a world where Google, Microsoft, and Amazon are building health AI products on data they control, publicly funded, openly accessible AI-ready health datasets are a strategic counterweight.

How to Position a Proposal

The specific funding opportunity announcements for Stage 2 have not been published as of this writing. But the program's structure and Stage 1 experience suggest several principles for competitive proposals:

Start with the clinical problem, not the AI method. Bridge2AI Innovation Funnels are structured around health challenges, not algorithmic novelty. A proposal that says "we will apply transformer architectures to ICU data" is weaker than one that says "we will predict sepsis onset 12 hours earlier using AI-ready ICU datasets from CHoRUS, enabling earlier intervention and reducing mortality."

Demonstrate dataset readiness. If you are proposing to use Stage 1 datasets, show that you have already accessed the Bridge2AI portal, understand the data formats, and can articulate exactly which variables you need and how they map to your analytical approach. Reviewers will distinguish between applicants who have done their homework and those waving at the dataset from a distance.

Build a team that spans the pipeline. Stage 2 is explicitly about moving from data to tools. A team of only computational researchers cannot credibly promise clinical deployment. A team of only clinicians cannot credibly build AI systems. Include a regulatory expert if your tool will require FDA clearance.

Address bias and equity from the outset. Stage 1's AI-READI project made equitable representation a core design principle, and Stage 2 will expect the same. Articulate how your approach handles underrepresented populations, how you will detect and mitigate algorithmic bias, and how you will validate performance across demographic groups.

Connect to the Network for AI Health Science. Even if your proposal is primarily technical, referencing the governance and safety frameworks that the Network will develop shows reviewers you understand the full ecosystem Bridge2AI is building.

The Bigger Picture

Bridge2AI exists at the intersection of two massive trends: the explosion of AI capability and the recognition that health data infrastructure has not kept pace. The program will not solve either problem on its own. But it is building the connective tissue — the shared datasets, common standards, and evaluation frameworks — that allow a fragmented research ecosystem to work together on health AI problems too large for any single team.

For researchers watching the space, the immediate action item is to register on the Bridge2AI portal, explore the Stage 1 datasets, and join the program listserv for Stage 2 funding announcements. The Innovation Funnel solicitations, when they arrive, will move quickly. Teams that have already engaged with the data will have a significant head start over those scrambling to understand it after the announcement. Discovery tools like Granted can help you track when those solicitations drop and match them against your research profile before the deadline clock starts ticking.