Introduction

There is mounting evidence that electronic health records (EHR) contain language reflecting the unconscious biases of clinicians, [1–7] and that this language may undermine the quality of care that patients receive.[8,9] Most of these studies have focused on stigmatizing language broadly, characterizing numerous ways that patients are negatively portrayed [1,2,10] and documenting racial disparities in the prevalence of negative or stigmatizing language [3–7].

One category of negative language that receives limited attention relates to assessments of patient credibility. Credibility assessments reflect a very specific type of bias, termed testimonial injustice, wherein a person (e.g., patient) may be less likely to be believed due to a prejudice on the part of the listener (e.g., clinician) [11]. Several qualitative studies suggest that Black Americans are more likely to have their concerns dismissed in healthcare settings [12,13], and two recent studies have documented this phenomenon by examining medical records. In our own earlier work [14], we found that notes from an academic ambulatory internal medicine clinic for patients identified as Black compared to White contained more markers of doubt in each of three linguistic categories: evidentials reporting the patient’s history as hearsay, patient quotes, and specific ‘doubt’ words. A recent study by Lee and colleagues [15] in an inpatient internal medicine setting similarly found greater use of evidentials in the notes of Black compared to White patients. While these early studies provide suggestive evidence of testimonial injustice in the EHR, they have been limited to the field of internal medicine and have used methods with limited specificity in identifying credibility-related language.

In the current study we sought to expand and strengthen the evidence related to bias in credibility assessments documented in the EHR in three important ways. First, we focused specifically on words used to undermine patient credibility using highly reliable natural language processing (NLP) models. Second, using these NLP models, we further distinguished different categories of terms undermining credibility, including those that cast doubt on patients’ sincerity as well as terms that undermine or support patients’ competence in providing accurate information (poor/good historian). Finally, we expanded beyond the field of internal medicine to include a wider range of clinical disciplines. Our overall goal was to assess possible racial differences in clinician use of language either supporting or undermining patient credibility within the EHR of patients from a large health system.

Methods

Results

Study sample

There were 13,065,081 notes written about 1,537,587 patients by 12,027 clinicians. Clinician-authors were attending physicians (44.9%), physicians in training (40.1%) and advanced practice providers (15.0%). Characteristics of patients and notes are shown in Table 1. The mean patient age was 43.3 years and 56% of patients were female. More than half (57.6%) were non-Hispanic White, 28.0% were non-Hispanic Black, 8.3% Hispanic/Latino, and 6.1% Asian. Most patients had commercial insurance (58.7%); 19.2% had Medicare, 14.9% had Medicaid and 7.2% were uninsured. Most notes were written for encounters in the Internal Medicine or Surgery departments (46.4 and 30.1% respectively), and most were written by attending physicians (65.7%). There were significant differences by patient race/ethnicity for all demographic and clinical variables, as outlined in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Participant characteristics.

https://doi.org/10.1371/journal.pone.0328134.t001

Language undermining patient credibility

Fewer than 1% of notes contained language undermining patient credibility (n = 106,523; 0.82%), with 62,480 (0.48%) containing language undermining patients’ sincerity and 52,243 (0.40%) containing language undermining competence.

In adjusted analyses (Table 2), notes written about non-Hispanic Black vs. White patients had overall higher odds of containing terms undermining credibility (aOR 1.29, 95% CI 1.27–1.32). This disparity was observed for terms undermining both patients’ sincerity (aOR 1.16; 95% CI 1.14–1.19) and patients’ competence (aOR 1.50; 95% 1.47–1.54). In terms of differences in specific words used in notes of Black vs. non-Black patients (Fig 1), there was a higher odds of each individual word undermining patient credibility, except for the terms related to narcotic or drug-seeking, which had a lower odds of appearing in the notes of Black vs. non-Black patients.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Racial/ethnic differences in language undermining and supporting credibility¹ in 13,681,829 notes across 4 clinical departments.

https://doi.org/10.1371/journal.pone.0328134.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Differences in specific terms undermining credibility for black vs. non-black patients.

https://doi.org/10.1371/journal.pone.0328134.g001

In fully adjusted models, there was no difference between Hispanic/Latino vs. White patients in language undermining patient credibility overall (aOR 0.99, 95%CI 0.95–1.03). When examining terms related to sincerity and competence separately, our findings were bidirectional. Notes written about Hispanic/Latino vs. White patients had lower odds of undermining sincerity (aOR 0.88, 95%CI 0.83–0.93). In unadjusted analyses, notes written about Hispanic/Latino vs. White patients also had lower odds of undermining competence, but this relationship was reversed in the fully adjusted model (aOR 1.20; 95% CI 1.13–1.27).

Notes written about Asian vs. White patients had lower odds of having language undermining credibility overall (aOR 0.85, 95% CI 0.81–0.89). When examining terms related to sincerity and competence separately, in the fully adjusted models notes written about Asian vs. White patients had lower odds of language undermining sincerity (aOR 0.74, 95% CI 0.69–0.79) but did not differ in the prevalence of language undermining competence (aOR 0.95, 95% CI 0.89–1.01).

Discussion

We found evidence of racial bias in clinician assessments of patient credibility, based on the language they used in documenting clinical encounters. Specifically, after accounting for multiple potential confounders, we found that clinicians were more likely to insinuate in the EHR that Black compared to White patients were not truthful and not competent in reporting their own experiences. This testimonial injustice can harm Black individuals in numerous ways. Having one’s concerns dismissed has an immediate negative impact on the individual’s experience of respect, which erodes trust and decreases likelihood of engaging in future care or following treatment recommendations [12–14]. Further, failure to attend seriously to patients’ own reported symptoms may lead to physical harms such as untreated pain, delayed diagnoses, incorrect treatments, serious medical errors, and even death [21–25]. Because notes with biased language can affect the attitudes and decision-making of subsequent clinicians reading those notes [8], language undermining Black patients’ credibility may contribute to healthcare inequities.

We might question, in light of the low prevalence of credibility language detected by our models, whether the racial differences we observed, even if statistically significant as a result of our large sample size, are clinically important. We believe the language detected by our NLP models, which relied upon a limited set of specific words that explicitly cast doubt on the patient’s testimony, represents the tip of an iceberg in terms of how words are used to undermine patients’ credibility, and that our findings accordingly signal larger underlying disparities in credibility assessments. In our own qualitative work examining medical records [1], we have found doubt conveyed by clinicians in the medical record in more nuanced and implicit ways, such as placing the patient’s words in quotes (e.g., he states his pain is “still a 10”), juxtaposing the patient’s subjective report with their own objective observation (e.g., reports adherence to meds but doesn’t know Lasix dosing), or using a greater number of “evidential” terms that treat patient-reported information as “hearsay,” allowing the clinician to be noncommittal about believing that information (e.g., she says she has a headache vs. she has a headache). Our own and another prior study found higher use of quotes [14] and evidentials in the notes of Black compared to White patients [14,15], adding to evidence of a credibility deficit experienced by Black patients [26].

It has been suggested that the term ‘poor historian’ (or any of its variants) may be used by clinicians who themselves failed to obtain a good history [27–31]. Thus, disproportionate use of this term in the notes of Black patients may reflect, in addition to credibility bias, ineffectiveness in communicating with Black patients [12,13,32,33]. Importantly, use of the term ‘poor historian’ has the potential to bias future clinicians, who may discount the patient’s symptoms or be less likely to solicit the patient’s story. There are times when a clinician may reasonably be concerned that they weren’t able to gather the information they needed because the patient was not able to communicate it clearly. In those instances, we suggest specifying that to be the case – “patient unable to provide a complete history” or “patient is uncertain of some details” – that would convey less blame or avoid a global discrediting label.

One could argue that it may also be dangerous to not question the accuracy of a patient’s history when that patient is not being truthful, is misremembering, or has a recognized mental health disorder such as malingering. In these cases, it is not our intent to prohibit or criticize the conveying of such doubt in a medical record. We cannot and do not intend to assert that any individual clinician is biased against any individual patient simply because they use one of these words. Our primary concern here is that the observed racial differences in the use of these words suggest an implicit racial bias, on a population-level, in assessments of credibility. It is this bias to which we encourage clinicians to become attentive, both in terms of their own assessments of patient credibility as well as in interpretation of what has been written in the record by other clinicians. When examining specific credibility-related terms, we found that, after adjusting for demographic and clinical characteristics including substance use disorder and mental illness, all terms were used more often among Black vs non-Black patients, except for those related to narcotic or drug seeking. While we cannot discern from our data the reasons underlying this finding, we suspect clinicians may avoid using those terms for Black patients out of awareness of the racial stereotypes they perpetuate. If so, raising clinicians’ awareness of the biases and stereotypes that likely underlie the disproportionate use of other terms undermining Black patients’ credibility may help to remedy racial inequities in credibility assessments and credibility-related documentation.

Although we found consistent evidence of credibility bias against Black compared to White patients, we did not find similar evidence for Hispanic/Latino or Asian patients. Indeed, we found the opposite, with fewer terms questioning sincerity among Hispanic/Latino or Asian patients compared to Whites. While our data do not provide clear explanations for these findings, it may be that the sociocultural distance between clinicians and patients perceived as “foreign” [34] result in less in-depth, more superficial interactions and therefore fewer opportunities to suspect patients of insincerity. Similarly, Asian patients’ being more likely than White patients to be explicitly called out as “good historians” may reflect a lower expectation for accurate history provision among patients perceived as foreign. This represents a form of “reverse linguistic stereotyping,” in which a person’s race, ethnicity, or appearance influences expectations and perceptions of that person’s communication and may lead a clinician to praise articulate histories in the face of anticipated language or cultural barriers.[35] “Poor historian” terminology had a lower prevalence in unadjusted analyses for Hispanic/Latino patients; however this association was inverted after adjustment for demographics (primarily patient age), which may reflect greater language barriers among older Hispanic patients.[36] Further studies exploring clinician perceptions of patient credibility by race and ethnicity could test these hypothetical explanations. Based on our findings, it appears that credibility bias is worst for Black Americans.

Our study has several limitations. We studied records from one health system, although the system does expand across the mid-Atlantic region and includes a broad array of clinical settings (5 different hospitals, both community-based and academic, as well as a network of ambulatory practices) and disciplines. Our health system includes a higher proportion of Black patients than the national average, and we do not know whether disparities might be larger or smaller in other settings. We did not have data on clinicians beyond their role/training and clinical department and could not examine the influence of other clinician characteristics (e.g., age, race, gender). We also did not analyze nursing notes because, in our health system, they are mostly templated. These are important areas for future study. We selected a limited set of terms representing credibility language; as noted above, our findings may therefore underestimate its true prevalence [14]. We used NLP models with high, but not perfect, accuracy in detecting credibility-related language [16]; these models may have therefore misclassified some notes and thereby under- or overestimated the prevalence of credibility-related language. Any resulting misclassification, however, was likely to be non-differential; i.e., it is unlikely that the NLP models were less accurate in classifying credibility-related terms for one racial/ethnic group vs another. The impact of such misclassification, therefore, would be to bias toward more conservative (less pronounced) estimates of racial disparities. In addition, we did not capture evidence for supporting credibility in the ‘sincerity’ category because, while clinicians do often document that a patient is a good historian, there is no parallel documentation practice for complimenting the patient on being truthful. Finally, we had a relatively small proportion of notes from patients who were not identified as Black or White, and documentation of demographic data in the medical record is not always accurate.

In conclusion, clinician documentation undermining patient credibility may disproportionately stigmatize Black individuals and favor Asian individuals. As stigmatizing language in medical records has been shown to negatively influence clinician attitudes and decision-making [8,9], these racial differences in documentation may negatively influence quality of care and outcomes and further exacerbate health inequities. This specific type of bias related to credibility assessment has been relatively overlooked and has therefore remained unchecked. Medical educators have taken up efforts to train future physicians to engage in more respectful, patient-centered documentation, [37,38] and to learn about the structural inequities and interpersonal attitudes [39] that underlie biased language use in the EHR. We recommend that these curricula include content on testimonial injustice and suggest eliminating phrases such as ‘poor historian,’ as well as veiled insinuations about patient insincerity. As artificial intelligence applications are increasingly adopted to aid in note writing, consideration should be given to programming those applications to also avoid such language. Finally, we recommend that health professionals continue to confront the racial biases underlying differential assessments of patient credibility in the first place.

Supporting information