---
type: "post"
id: "usc-analysis-of-58-746-nih-grants-finds-29e231"
slug: "usc-analysis-of-58-746-nih-grants-finds-29e231"
title: "USC analysis of 58,746 NIH grants finds AI makes up 15.9% of the FY2025 portfolio."
description: "A USC-led analysis of 58,746 NIH-funded biomedical projects from 2025 estimates that AI accounts for 15.9% of the portfolio, with a 13.4% funding premium."
retrieval_nugget: "A USC-led analysis of 58,746 NIH-funded biomedical projects from 2025 estimates that AI accounts for 15.9% of the portfolio, with a 13.4% funding premium. The paper also finds a large gap between research and clinical implementation, offering teams a reminder that investment, research activity, deployment, and demonstrated outcomes are different measures of progress."
published_at: "2026-11-03"
updated_at: "2026-11-03"
record_date: "2026-11-03"
date_kind: "published_at"
topics: ["ai"]
entities: ["National Institutes of Health","University of Southern California","Navapat Nananukul","Mayank Kejriwal"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "The USC analysis reports new findings on AI’s share, funding premium, deployment gap and health-equity representation in the FY2025 NIH portfolio."
prior_refs: []
source_url: "https://arxiv.org/abs/2604.07424"
source_urls: ["https://arxiv.org/abs/2604.07424"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:usc-analysis-of-58-746-nih-grants-finds-29e231"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/usc-analysis-of-58-746-nih-grants-finds-29e231/","markdown":"https://newruntime.com/posts/usc-analysis-of-58-746-nih-grants-finds-29e231.md","json":"https://newruntime.com/posts/usc-analysis-of-58-746-nih-grants-finds-29e231.json"}
markdown_url: "https://newruntime.com/posts/usc-analysis-of-58-746-nih-grants-finds-29e231.md"
---

# USC analysis of 58,746 NIH grants finds AI makes up 15.9% of the FY2025 portfolio.

## Retrieval answer

A USC-led analysis of 58,746 NIH-funded biomedical projects from 2025 estimates that AI accounts for 15.9% of the portfolio, with a 13.4% funding premium. The paper also finds a large gap between research and clinical implementation, offering teams a reminder that investment, research activity, deployment, and demonstrated outcomes are different measures of progress.

## Summary
A USC-led analysis of 58,746 NIH-funded biomedical projects from 2025 estimates that AI accounts for 15.9% of the portfolio, with a 13.4% funding premium. The paper also finds a large gap between research and clinical implementation, offering teams a reminder that investment, research activity, deployment, and demonstrated outcomes are different measures of progress.

## What happened
Navapat Nananukul and Mayank Kejriwal analyze artificial intelligence and machine learning adoption across the National Institutes of Health's 2025 portfolio in a paper titled “An Analysis of Artificial Intelligence Adoption in NIH-Funded Research.” The paper was submitted to arXiv on April 8, 2026 and revised on October 29, 2026. The record also lists a 2026 IEEE International Conference on Digital Health journal reference. Its central dataset covers 58,746 NIH-funded biomedical research projects from 2025.

The authors frame the work as a way to understand the distribution of AI and ML work in a major biomedical funding portfolio. Their stated motivation is practical as well as policy-oriented: research descriptions are numerous and unstructured, while a portfolio-level view can inform research funding strategy, institutional planning, and health policy. Rather than manually reading every project description in isolation, the paper describes a human-in-the-loop methodology that uses large language models to classify and summarize research descriptions at scale.

The supplied source establishes that LLMs were used for semantic extraction with human involvement. It does not specify the model, prompts, reviewer workflow, annotation protocol, or the rules used to resolve disagreements. That distinction matters because the reported shares depend on how a project description is classified, rather than on direct observation of laboratory activity, clinical systems, or patient outcomes.

The paper's headline estimate is that AI constitutes 15.9% of the NIH portfolio. It further reports that AI projects carry a 13.4% funding premium. According to the abstract, the classified work is concentrated in discovery, prediction, and data integration across disease domains. These findings describe the composition of the portfolio examined by the authors; they do not identify individual grants, rank technical methods, or say which AI systems have proven effective in practice.

The authors' second headline is a “research-to-deployment gap.” They report that 79% of AI projects remain in research and development stages, while 14.7% engage in clinical deployment or implementation. The abstract does not explain whether the categories are mutually exclusive, what share falls outside them, or how “clinical deployment” differs from “implementation.” Still, the direction of the finding is clear: in the authors' classification, most AI-tagged projects have not crossed into the clinical-use category.

Finally, the paper estimates that health disparities research accounts for 5.7% of AI-funded work. The authors characterize that share as severely underrepresented given NIH's equity mission and argue that the results provide a framework for evidence-based policy interventions aligned with strategic priorities and health equity goals. This is an analysis and recommendation from the paper, not an announcement that NIH has adopted a new allocation rule, program, or intervention.

## What the archive adds
We have not covered this before; it opens a new thread.

## Trend signal
No supplied SITE_PATTERNS entry directly addresses public research funding allocation or the gap between biomedical research and clinical deployment. The findings are relevant to AI adoption, but linking them to a broader pattern would imply support the source does not establish.

## Why now
The paper was revised on October 29, 2026, after its April 8 arXiv submission, and its record identifies a 2026 IEEE International Conference on Digital Health reference. The supplied material does not say what changed in the revision, whether the underlying dataset was updated, or whether a funding or policy event prompted publication. Its immediate relevance is therefore the availability of a recent, quantified portfolio analysis rather than a newly announced NIH action.

## Why it matters
For teams building AI for health research, the headline adoption share is only one part of the picture. A portfolio can devote substantial funding to AI while most projects remain upstream of clinical use. That distinction matters when teams plan pilots, set success metrics, or describe what evidence supports a product: research activity is not evidence of clinical implementation, and implementation is not itself evidence of improved outcomes.

The reported concentration in discovery, prediction, and data integration points to where NIH-funded work is being classified. It may help teams distinguish crowded research categories from the harder operational work that comes after a model or analytic method is developed: workflow integration, clinical validation, governance, and sustained use. But the paper does not measure any of those downstream steps separately.

The 5.7% estimate for health disparities research may also prompt funders and builders to examine whether their own portfolios address underserved populations. It is a signal about allocation in this analysis, not a causal account of why disparities persist and not a prescribed product roadmap. Builders should resist turning the figures into claims that a particular product category, model, or vendor has been clinically validated.

## Limits and uncertainty
These figures come from authors' classification and analysis of project descriptions, using a human-in-the-loop LLM methodology. The supplied abstract does not provide enough detail to independently assess the definitions of “AI,” “funding premium,” research stage, clinical deployment, or health disparities work. It also provides no validation metrics, error analysis, confidence intervals, or uncertainty intervals in the material supplied here.

Portfolio shares do not establish the quality, effectiveness, safety, or clinical impact of projects. The 79% and 14.7% figures are not explained as exhaustive or mutually exclusive categories. Nor does the source disclose whether a project classified as deployed was evaluated in routine care, in a pilot, or by another threshold. The paper does not establish that NIH has changed its funding practices in response to its findings.

## Skill to build now
Separate your evidence ledger into at least four fields: funding or budget, research-stage progress, deployment status, and measured outcome. For a current health-AI pilot, write one criterion and one source of evidence for each field. This week's useful result is not a larger dashboard; it is preventing a grant, prototype, or clinical integration from being treated as proof of patient benefit.

## Reader takeaway
The useful finding is not simply that AI represents 15.9% of NIH's 2025 portfolio. Most AI projects were still classified as research and development, while a much smaller share involved clinical implementation. Builders should keep funding, research progress, deployment, and demonstrated patient benefit as distinct measures.

## Open question
How do the authors define and validate the categories behind the reported AI portfolio and clinical implementation shares?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Ai](https://newruntime.com/topics/ai.md): Explore the ai topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)