Who is leading medical AI? A systematic review and scientometric analysis of chest x-ray research

This systematic review and scientometric analysis of 928 chest X-ray AI studies reveals that research leadership and training data are overwhelmingly dominated by high-income countries, particularly the US and China, creating significant disparities that risk developing AI systems with inconsistent performance across diverse global populations and exacerbating healthcare inequities.

The Gist

Imagine the world of medical AI as a massive, high-tech kitchen where chefs are trying to invent the perfect automated recipe for diagnosing chest diseases using X-rays. The goal is amazing: to help doctors everywhere, especially in places where there aren't enough specialists, to spot illnesses like pneumonia or tuberculosis quickly and accurately.

This paper is like a food critic who went into that kitchen, looked at nearly 1,000 recipes (research studies) published between 2017 and 2025, and asked a very important question: "Who is cooking, and what ingredients are they using?"

Here is what the critic found, broken down into simple stories:

1. The "Who's Cooking?" Problem (The Chef Disparity)

The Analogy: Imagine a cooking competition where the judges and head chefs are almost entirely from wealthy, famous culinary capitals like New York and Beijing.
The Reality: The study found that High-Income Countries (like the US, China, and South Korea) are doing almost all the leading work.

• China and the US are the top two "chefs," writing about 33% of all the recipes.
• The Missing Voices: Not a single "head chef" (first author) came from a Low-Income country. Even when looking at the "senior chefs" (senior authors), almost none were from the poorest nations.
• The Risk: If you only ask chefs from rich cities to design a menu for the whole world, they might forget that people in rural villages have different tastes, different allergies, and different cooking equipment. The AI is being built by people who might not understand the specific health problems of the poorest regions.

2. The "Ingredients" Problem (The Data Disparity)

The Analogy: Imagine the chefs are trying to learn how to bake a cake. But instead of tasting flour from all over the world, they are only baking with flour from one specific, expensive bakery in the US.
The Reality: The "ingredients" (the medical data/X-rays) used to train these AI computers are mostly from wealthy countries.

• 73.6% of the data comes from High-Income countries.
• The United States alone provided over 40% of the data.
• The Risk: If you train a computer to recognize a disease using X-rays from healthy, wealthy patients, it might get confused when it sees an X-ray from a patient in a poor country who has different lung conditions, different body types, or different types of equipment. It's like teaching a dog to recognize only Golden Retrievers; when it sees a Chihuahua, it doesn't know what it is.

3. The "Teamwork" Problem (The Collaboration Gap)

The Analogy: Imagine the rich chefs and the poor chefs rarely talk to each other. When they do work together, it's usually the rich chef calling the shots, and the poor chef just handing over some vegetables.
The Reality: The study found that cross-country teamwork is extremely rare.

• Only 3.9% of the studies involved a partnership between a rich country and a poorer country.
• Most research is done by teams from rich countries working with other rich countries.
• The Risk: This means the "local experts" in developing nations aren't helping to design the tools. They aren't saying, "Hey, this doesn't work for our patients," because they aren't in the room where the decisions are made.

The Big Picture: Why Should We Care?

The authors warn that if we keep cooking this way, we are building AI tools that are biased.

Think of it like a pair of glasses designed for a person with perfect vision in a sunny city. If you hand those same glasses to someone in a foggy, rainy village, they won't just be blurry; they might make things look worse than they were before.

If medical AI is trained only on data from rich countries, it might fail when used in the places that need it the most (where respiratory diseases are actually killing the most people). This could accidentally make health inequality worse instead of fixing it.

The Solution: A New Recipe

The paper suggests we need to change the kitchen rules:

1. Invite New Chefs: We need to actively support researchers from poorer countries to lead the projects, not just help out.
2. Gather Global Ingredients: We need to collect X-ray data from all over the world, not just the US and Europe, and make it available to everyone.
3. True Partnerships: Rich countries should work with poorer countries as equals, sharing money and technology, rather than just taking data and leaving.

In short: The technology is powerful, but right now, it's being built by a small group of people using a very narrow set of examples. To save lives globally, we need to open the kitchen doors, invite everyone in, and make sure the recipes work for everyone, not just the wealthy.

Technical Summary

1. Problem Statement

The paper addresses a critical gap in the development of medical Artificial Intelligence (AI): the lack of equity and generalizability in computer vision (CV) models for chest X-ray (CXR) interpretation.

• The Issue: While CV models show promise for automating CXR diagnosis, their clinical utility is compromised if they are trained and validated on data that does not represent the global population.
• The Risk: Current research is heavily skewed toward High-Income Countries (HICs). This creates a "generalizability gap" where AI systems perform well in the settings they were developed in but fail when deployed in Low- and Middle-Income Countries (LMICs), where the burden of respiratory diseases (e.g., tuberculosis, pneumonia) is often highest.
• The Question: Who is leading this research, where does the data come from, and are the resulting technologies inclusive of diverse global populations?

2. Methodology

The authors conducted a systematic review and scientometric analysis following the PRISMA 2020 guidelines.

• Data Sources: A systematic search was performed across PubMed, Embase, and SciELO databases.
• Timeframe: Publications from 2017 to 2025 (noting the preprint date of April 2026).
• Inclusion Criteria: Peer-reviewed journal and conference papers focusing on CV or multimodal approaches applied to CXRs.
• Exclusion Criteria: Preprints (initially), reviews/surveys, incomplete papers, and studies using non-CV methods (e.g., NLP or traditional ML).
• Sample Size:
  • Initial search yielded 1,555 articles.
  • After screening and full-text review, 928 eligible studies were retained for quantitative analysis.
• Data Extraction & Analysis:
  • Used the Dimensions API and manual extraction to gather metadata (author affiliations, journal, dataset provenance).
  • Classification: Countries were categorized by World Bank income levels (Low, Lower-Middle, Upper-Middle, High).
  • Normalization: Author contributions were normalized by population size, GDP, and R&D investment to assess relative research output.
  • Focus Areas: First and senior author affiliations, institutional contributions, dataset origins, and cross-income collaboration patterns.

3. Key Contributions

• Comprehensive Scientometric Mapping: This is one of the first large-scale analyses to map the global landscape of CXR-AI research, specifically quantifying disparities in authorship and data provenance.
• Identification of Structural Imbalances: The study provides empirical evidence of a "self-reinforcing ecosystem" where HICs dominate both the production of knowledge (authorship) and the resources (datasets) required to train models.
• Quantification of the "Data Poverty" Gap: It highlights that the populations most in need of AI diagnostic support (LMICs) are almost entirely absent from the research process as both authors and data contributors.

4. Key Results

The analysis of 928 studies revealed significant disparities:

A. Authorship Leadership

• Dominance of HICs: High-income countries accounted for 55.6% of first authors and 59.7% of senior authors.
• Top Countries: China (16.93% first authors) and the United States (16.72%) led in volume. South Korea and India followed.
• Exclusion of Low-Income Countries (LICs): Zero first authors were affiliated with low-income countries. Only 0.7% of senior authors had LIC affiliations.
• Normalization: When adjusted for population, GDP, and R&D, countries like Singapore, Qatar, and Denmark showed high per-capita output, while countries like Jordan and Pakistan showed high output relative to GDP.

B. Dataset Provenance

• HIC Data Dominance: 73.6% of all datasets originated from HICs.
• US Contribution: The United States was the single largest contributor (40.45% of datasets).
• Private vs. Public: Private datasets were the most frequently used (20.52%), followed by public datasets like ChestX-ray14 (8.24%) and CheXpert (6.31%).
• LMIC Representation: Datasets from LMICs were severely underrepresented.

C. Collaboration Patterns

• Intra-HIC Dominance: 53.43% of collaborations occurred between HIC institutions.
• Cross-Income Rarity: Collaborations between HICs and LMICs were rare.
  • LMIC First Author / HIC Senior Author: 3.91%
  • HIC First Author / LMIC Senior Author: 1.68%
• Data-Author Mismatch: Studies using HIC datasets were overwhelmingly led by HIC authors (51.38% first authors), whereas studies using LMIC datasets rarely had LMIC authors leading the work.

5. Significance and Implications

The paper argues that the current trajectory of medical AI development poses a risk of exacerbating global health inequities.

• Performance Disparities: Models trained on HIC data (specifically US data) are likely to underperform in LMIC settings due to differences in disease presentation, comorbidities, and imaging equipment quality.
• The "Generalizability Gap": The lack of diverse training data means AI tools may not be reliable for the populations that need them most (e.g., for TB detection in Africa or Asia).
• Call to Action: The authors propose a roadmap for an equitable research ecosystem:
  1. Funding: Prioritize the creation of diverse, open-access CXR datasets from underrepresented populations.
  2. Collaboration: Establish partnerships where LMIC researchers are treated as equal leaders and knowledge producers, not just data sources.
  3. Policy: Publishers and conferences should incentivize diverse authorship and cross-income collaborations.
  4. Capacity Building: Invest in infrastructure and computational resources within LMICs to enable local research leadership.

Conclusion: The study concludes that without deliberate structural reforms to democratize data and authorship, medical AI risks becoming a tool that benefits the wealthy while failing the poor, thereby widening the global health divide.