Clinicians are more likely to indicate the doubt or disbelief in the medical records of black patients than in those of white patients – a model that could contribute to the racial disparities in progress in health care. This is the conclusion of a new study, analyzing more than 13 million clinical notes, publishing on August 13, 2025 in the outdoor access magazine Plos a By Mary Catherine Beach of Johns Hopkins University, United States
There is increasing evidence that electronic health files (DSE) contain a language reflecting the unconscious biases of clinicians, and that this language can undermine the quality of the care that patients receive.
In the new study, the researchers analyzed 13,065,081 DSE notes written between 2016 and 2023 approximately 1,537,587 patients by 12,027 clinicians of a large health system in the middle of the Atlantic. They used artificial intelligence tools (AI) to find which notes had a language suggesting that the clinician doubted the patient’s sincerity or narrative competence for the example indicating that the patient “claims”, “insists” or is “categorical about” their symptoms, or is a “poor historian”.
Overall, less than 1%(n = 106,523; 0.82%) of medical tickets contained a language undermining the patient’s credibility – of which about half has undermined sincerity (n = 62,480; 0.48%) and half of the competence (n = 52,243; 0.40%). However, the notes written on non -Hispanic black patients, compared to those written on white patients, were likely to contain terms undermining the credibility of patients (AOR 1.29, 95% CI 1.27–1.32), sincerity (AOR 1.16; 95% CI 1.14–1.19) 1.47–1.54). In addition, the notes written on black patients were less likely to have a language supporting credibility (AOR 0.82; 95% CI 0.79–0.85) than those written on white or Asian patients.
The study was limited by the fact that it used only one health system and did not examine the influence of the characteristics of clinicians such as race, age or gender. In addition, as the NLP models used had a high precision, but not perfect, in the detection of the language linked to credibility, they may have poorly classified certain notes and thus overestimate the prevalence of language linked to credibility.
However, the authors conclude that the documentation of the clinician undergoing the patient’s credibility can stigmatize black people disproportionately, and that the results are probably “the tip of an iceberg”. They say that medical training should help future clinicians to become more aware of unconscious prejudice and that AI tools used to help write medical notes should be programmed to avoid biased language.
The authors add: ” For years, many patients – especially black patients – said their concerns were rejected by health professionals. By isolating words and sentences suggesting that a patient cannot be believed or taken seriously, we hope to raise awareness of this type of credibility bias in order to eliminate it.«