Anthology of Computers and the Humanities · Volume 3

Interrogating Racism in the Medical Literature Using Word Embeddings

Lauren D. Liao2 ORCID , Sajia Darwish3 , Caroline Figueroa4 ORCID , Erin Manalo-Pedro5 ORCID , Swetha Pola5 , Maithili Jha6 , Fernando De Maio7,8 ORCID , Claudia von Vacano1 ORCID , Chris J. Kennedy9,10 ORCID and Pratik S. Sachdeva1 ORCID

  • 1 D-Lab, University of California, Berkeley, Berkeley, USA
  • 10 Department of Psychiatry, Harvard Medical School, Boston, USA
  • 2 Division of Research, Kaiser Permanente, Oakland, USA
  • 3 Department of Biostatistics, Harvard University, Cambridge, USA
  • 4 Delft University of Technology, Delft, Netherlands
  • 5 Independent Scholar
  • 6 Rimtec Corporation, Addison, USA
  • 7 Department of Sociology, DePaul University, Chicago, USA
  • 8 Health Equity Research, American Medical Association, Chicago, USA
  • 9 Center for Precision Psychiatry, Massachusetts General Hospital, Boston, USA
Download PDF Download Citation

Permanent Link: https://doi.org/10.63744/voHr9u5XsC0n

Published: 21 November 2025

Keywords: racism, medicine, digital humanities, medical humanities, embeddings, natural language processing

Abstract

The medical literature has an important role to play in establishing anti-racist practice that may alleviate racial health inequities. Recent critical discourse analyses have demonstrated that medical literature often fails to explicitly name racism or discuss it through a structural lens, instead employing euphemistic language that obscures structural determinants of health inequities. Here, we build upon this work by using Word2Vec word embeddings to interrogate a corpus of 871 published articles containing the word “racism” sourced from top medical journals between 1999 and 2020. Our findings reveal distinct patterns in medical discourse around racism. First, hierarchical clustering of discrimination-, power-, and wealth-related words demonstrated clear separation between racism-related concepts and structural determinants, with racism showing minimal similarity to wealth-related words while clustering more closely with other forms of discrimination. Second, we found that qualifying language denoting uncertainty (e.g., “maybe”, “possibly”) showed higher similarity to racism-related words than more confident language, suggesting qualifying language serves as a hedge against direct assertions about racist processes. Finally, we conducted a network analysis revealing how concepts cluster within medical discourse, with bridging words between health inequities and racism clusters predominantly reflecting interpersonal rather than structural framings of racism. Specifically, words associated with health inequities, such as “stress” and “homelessness” connected to racism primarily through person-level gatekeepers such as “interpersonal,” “prejudice,” and “overt,” while structural concepts remained notably absent from common pathways. Overall, understanding these linguistic patterns is crucial as the medical community works to build anti-racist norms, while simultaneously relying on medical text to train artificial intelligence systems deployed in clinical settings.