💻 computer science

An Analysis of Artificial Intelligence Adoption in NIH-Funded Research

This paper analyzes 58,746 NIH-funded biomedical projects from 2025 using a human-in-the-loop LLM methodology to reveal that while AI constitutes 15.9% of the portfolio with a funding premium, it suffers from a significant research-to-deployment gap and critically underrepresents health disparities research.

Original authors: Navapat Nananukul, Mayank Kejriwal

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Navapat Nananukul, Mayank Kejriwal

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the National Institutes of Health (NIH) as a massive, bustling construction company responsible for building the future of human health. Every year, they hand out billions of dollars to thousands of different teams (researchers) to build new tools, discover cures, and understand diseases.

For a long time, this company has been hiring a new, super-smart foreman: Artificial Intelligence (AI).

This paper is like a giant audit of that construction company. The authors used a special kind of AI (called a Large Language Model) to read through the blueprints of 58,746 different projects to figure out exactly how much AI is being used, where it's being used, and where it's being ignored.

Here is the breakdown of their findings, using some everyday analogies:

1. The "AI Premium" (It's Popular and Expensive)

The Finding: About 16% of all NIH projects now involve AI.
The Analogy: Imagine if you went to a grocery store and found that 1 out of every 6 items on the shelf was "Smart Food." Even more interesting? The "Smart Food" items cost 13% more than the regular food.
What it means: AI is a hot commodity. The NIH is willing to pay extra for it because they believe it's the key to solving hard problems. It's not a niche experiment anymore; it's a major part of the budget.

2. The "Fancy Neighborhoods" vs. The "Forgotten Streets"

The Finding: AI money is heavily concentrated in Cancer, Aging, and Mental Health. However, research on Health Disparities (helping underserved communities, minorities, and rural areas) gets only 5.7% of the AI funding.
The Analogy: Think of the NIH budget as a city planner deciding where to build high-tech, self-driving bus lanes.

They built massive, shiny bus lanes in the wealthy, well-connected neighborhoods (Cancer and Aging research).
But they completely forgot to build any lanes in the poorer, rural, or minority neighborhoods (Health Disparities).
The Problem: Even though the NIH says, "We want to help everyone equally," the AI tools are mostly being built for the people who already have the best data and infrastructure. The people who need help the most are getting left behind.

3. The "Toy Factory" vs. The "Real World"

The Finding: 79% of AI projects are still in the "research and development" phase. Only 14.7% are actually being used in hospitals or clinics.
The Analogy: Imagine a toy factory that is churning out amazing, futuristic robots.

79% of these robots are sitting in a warehouse, being tested in a lab, or having their batteries charged. They are cool, but they aren't playing with kids yet.
Only 15% of the robots have actually been shipped out to schools and homes to do real work.
The Problem: There is a huge "translation gap." We are great at inventing the AI, but we are terrible at getting it out of the lab and into the doctor's office where it can actually save lives.

4. The "Clubhouses" and the "Connectors"

The Finding: The researchers found that universities are working together in tight-knit groups. A few big universities (like Johns Hopkins and the University of Minnesota) are the "Clubhouses" where most of the work happens. A few other schools act as "Connectors" (bridges) that link these groups together.
The Analogy: Think of the research world as a social network at a giant party.

Most people are hanging out in small, tight circles with their best friends (the "Clubhouses").
Only a few people (the "Connectors") are walking around shaking hands between the different groups.
The Problem: If you aren't one of the popular kids at the "Clubhouse" or one of the friendly "Connectors," it's very hard to get invited to the party. Smaller universities or those in less wealthy areas might be locked out of these important AI collaborations.

5. The "Magic Detective" (Human + AI Teamwork)

The Finding: The authors didn't just let the AI guess. They used a "Human-in-the-Loop" method.
The Analogy: Imagine trying to sort a massive pile of mixed-up mail.

If you just ask a robot to sort it, it might put a letter about "Chronic Pain" into a generic "Other" pile because it's confused.
In this study, the AI did the heavy lifting (sorting 58,000 letters), but a human detective stepped in to check the "Other" pile.
The Result: The human found hidden gems! They realized that "Chronic Pain," "Alcohol Use Disorder," and "Autoimmune Diseases" were actually huge categories that the standard system had been hiding under the label "Other."
The Lesson: AI is fast, but humans are needed to make sure the AI doesn't miss the important, nuanced stories.

The Big Takeaway

The paper concludes that while the NIH is doing a great job buying AI tools, they need to fix three things:

Spread the wealth: Stop only building AI for cancer and aging; build it for the communities that are currently being ignored.
Finish the job: Stop just building the robots in the lab; get them into the hospitals to actually help patients.
Open the doors: Make sure smaller universities and diverse groups can join the "Clubhouses" so the best ideas don't get stuck in just a few places.

In short: We have the technology, but we need to be smarter about who gets to use it and where.

1. Problem Statement

The National Institutes of Health (NIH) is the primary funder of biomedical research in the U.S., yet there is a lack of large-scale, granular understanding of how Artificial Intelligence (AI) and Machine Learning (ML) are adopted across its entire funding portfolio. Existing studies are limited to specific disease areas, narrow technology domains, or specific funding mechanisms.

This gap creates three critical challenges for policy and strategy:

Equity Misalignment: It is unknown whether AI investment patterns align with the NIH's strategic commitment to health equity and reducing disparities.
Translation Bottleneck: There is a suspected gap between fundamental AI research and real-world clinical deployment, but the scale of this gap is unquantified.
Infrastructure Disparities: High-profile disease areas (e.g., cancer) may have infrastructure advantages that concentrate AI funding, potentially leaving health disparities research under-resourced.

2. Methodology

The authors developed a human-in-the-loop (HITL) Large Language Model (LLM) pipeline to analyze 58,746 NIH-funded research projects from Fiscal Year 2025. The workflow consists of five integrated stages:

Data Ingestion: Data was sourced from the NIH RePORTER system. Records were deduplicated, normalized, and filtered to retain project titles, abstracts, terms, funding amounts, and institution types.
Two-Step LLM Pipeline:
1. Portfolio-Wide Screening (Pass 1): A conservative zero-shot prompt using GPT-4o-mini classified all 58,746 projects as "AI-positive" or "non-AI." The model was instructed to identify substantive AI/ML roles (development or application) rather than generic computational support. This identified 9,363 AI projects (15.9%).
2. Structured Rubric Coding (Pass 2): Only AI-positive projects were processed through a fixed-schema prompt to extract structured labels: what AI is used for, contribution type, role, primary focus areas, application domain, data type, and thematic notes.
Human-in-the-Loop Validation: Human investigators defined research questions and iteratively refined the LLM's decision rules (particularly for ambiguous "Other" categories). The LLM generated executable Python scripts for data aggregation, cross-tabulation, and visualization, producing 249 output files.
Network Analysis: A weighted, undirected collaboration network was constructed where nodes represent universities and edges represent co-participation in project clusters. Community detection was performed using Louvain modularity optimization. Network metrics included weighted degree, betweenness centrality, and power-law distribution analysis.

3. Key Contributions

Methodological Innovation: Demonstrated a scalable, reproducible HITL framework using LLMs to transform unstructured biomedical abstracts into structured, policy-ready intelligence. This approach reduced the "Other" disease category ambiguity by 23.8 percentage points (from 41.5% to 17.7%) through iterative refinement.
Comprehensive Portfolio Characterization: Provided the first large-scale empirical characterization of AI adoption across the entire NIH portfolio (58k+ projects), moving beyond siloed disease-specific studies.
Network Science Application: Mapped the structural inequality of AI research collaborations, identifying a "hub-and-spoke" topology with a small, critical bridge layer connecting dense communities.

4. Key Results

A. Funding and Prevalence

Prevalence: AI constitutes 15.9% of the NIH portfolio (9,363 projects).
Funding Premium: AI projects receive a 13.4% funding premium (mean award: $675,129) compared to non-AI projects (mean: $595,135).
Primary Applications: The portfolio is dominated by Discovery (44.1%), Prediction/Risk Assessment (36.7%), and Data Integration (26.2%). Data integration projects attract the highest funding intensity ($815,884/award).

B. Disease Area Concentration & Equity Gaps

Concentration: Three areas dominate AI funding: Cancer/Oncology (19.4%), Aging/Neurodegenerative (16.8%), and Mental Health (14.9%). Together, these account for 50.1% of all AI funding.
Underrepresentation: Health disparities and minority health research comprises only 5.7% of AI-funded work, despite being central to the NIH mission.
Methodological Mismatch: Imaging-intensive methods (common in cancer) account for 37% of cancer AI work but only 0.9% of health disparities research, indicating a failure to adapt established tools to equity-focused domains.

C. Research-to-Deployment Gap

Stage Distribution: 79.0% of AI projects remain in the Research/Development stage.
Clinical Deployment: Only 14.7% of projects engage in clinical deployment or implementation.
Implication: The translation infrastructure for bringing AI innovations into real-world clinical care is severely underdeveloped.

D. Collaboration Network Structure

Community Clustering: The network (79 universities, 191 edges) revealed 6 distinct communities anchored by specific institutions (e.g., Johns Hopkins, U of Minnesota, Emory).
Hub vs. Bridge:
- Hubs (High Degree): Johns Hopkins, U of Minnesota, and Wake Forest anchor dense local collaborations.
- Bridges (High Betweenness): The University of Washington acts as the primary connector (highest betweenness centrality), linking otherwise separated communities.
Inequality: The network follows a heavy-tailed power-law distribution ( $\alpha = 2.08$ ), indicating that collaboration opportunities and influence are concentrated in a small minority of institutions, creating structural barriers for others.

E. Semantic Recovery

Human-guided LLM analysis recovered hidden domains previously classified as "Other," including Alcohol Use Disorder (25 projects), Chronic Pain (23), Autoimmune Diseases (23), and Autism Spectrum Disorder (21).

5. Significance and Policy Implications

This study establishes an evidence base for strategic interventions in public research funding:

Correcting Structural Mismatches: The data proves that the underrepresentation of health disparities research is not due to a lack of scientific maturity but a structural failure to adapt computational tools to these domains. Policy interventions must target infrastructure and partnership building in equity-focused areas.
Bridging the Translation Gap: With 79% of AI research stuck in the R&D phase, funding mechanisms must be redesigned to incentivize clinical implementation and community health partnerships, particularly for underrepresented diseases.
Network Equity: The concentration of collaboration hubs suggests that institutions outside the "bridge" layer face barriers to multi-institutional AI research. Funding strategies should actively support capacity building for non-hub institutions to foster a more equitable research ecosystem.
Methodological Standard: The HITL LLM pipeline offers a scalable model for future portfolio analyses, enabling dynamic, real-time monitoring of research trends that traditional manual coding cannot achieve.