Background: Health research funding and output are misaligned with global disease burden, favouring diseases of affluent populations. Medical AI is the fastest-growing segment of health research, but whether it replicates or worsens this misalignment has not been quantified. AI research depends on labelled benchmark datasets that exist for some diseases but not others, creating a potential systematic bias in disease selection. Methods and Findings: We mapped 197,844 medical AI publications (2015–2025) from OpenAlex to 115 Global Burden of Disease (GBD) 2023 Level 3 causes covering 94.1% of global disability-adjusted life years (DALYs). A Research Attention Index (RAI) quantified each disease's publication share relative to its DALY share. AI research was moderately aligned with burden (Spearman *r*~s~ = 0.477; 95% CI 0.318–0.615), with extreme concentration (Gini 0.718): the top 10 diseases received 54.4% of publications. Skin melanoma (RAI 53.0), brain cancer (14.2), and breast cancer (11.4) were over-studied; road injuries (RAI 0.025; 74.7 million DALYs, 57 publications), diarrhoeal diseases (0.034), and anxiety disorders (0.036) were under-studied. AI was far better aligned with high-income-country burden (*r*~s~ = 0.619) than low-income-country burden (0.239). Multivariable regression (*n* = 38 diseases; R² = 0.71) identified dataset availability, research community size, and high-income-country burden share as predictors; burden itself was not significant. Simulation showed that strategic dataset creation for under-studied diseases could make AI a net corrective force. Limitations include title-based disease mapping (recall 75.8%), the small regression sample, and a linear mixing assumption in the corrective model. Conclusions: Medical AI research amplifies the research–burden mismatch — driven not by disease burden but by benchmark dataset availability. Strategic investment in datasets for high-burden, under-studied diseases could transform AI into a corrective force for global health equity.
Hayden Farquhar (2026) studied this question.