From word lists to validated classifiers
Earlier work established that EHR notes can transmit stigma through two distinct linguistic forms: stigmatizing labels, which ascribe negative behavior to patients, and doubt markers, which discredit patient testimony [1]. Foundational taxonomy work has argued that these forms discredit patients in ways that undermine the clinician-patient relationship [2]. But most NLP studies before CARE-SD relied on regular expression or exact term matching in relatively small note samples, which limits recall and misses paraphrased stigma [1]. CARE-SD addresses that gap by expanding an initial 18-word stigmatizing label list and 6-word doubt marker list using contextual embeddings and GPT-3.5, pruning with expert annotators, and training supervised classifiers on human-annotated MIMIC-III sentences [1]. The stigmatizing label classifier reached 81% accuracy, 75% precision, 84% recall, and 0.79 macro-F1; the RoBERTa doubt marker classifier reached 86% accuracy, 86% precision, 71% recall, and 0.84 macro-F1, both comparable to human annotator agreement [1].
The methodological advance is not just higher performance. It is that sentence-level classifications are aggregated to note-level outcomes, making it possible to test distributional hypotheses at scale [1]. That design choice is what allows the paper to move from detection to epidemiology.
Who receives more stigmatizing language
In MIMIC-III, Black or African American patients had an estimated 1.16 times higher rate of stigmatizing labels per chart than White patients, while Asian, Hispanic/Latino, Native American/Alaskan Native, and Other racial groups showed lower rates [1]. Insurance showed larger associations: government-run insurance was associated with 2.46 times the rate of stigmatizing labels, and self-pay with 2.12 times, compared with private insurance [1]. Diagnoses carried the strongest signals, with rate ratios from 1.50 to 4.90 for conditions including symptomatic HIV, obesity, opioid use disorder, substance use disorder, mood disorder, anxiety, PTSD, suicide attempts, and suicidal ideation [1]. Doubt markers followed similar patterns, though male patients had 1.25 times higher rates than female patients, a finding that contrasts with prior work reporting more pain disbelief among female patients [1].
These are associations, not causal estimates. The authors note that sickle cell disease and schizophrenia showed no significant differences, likely because of very small sample sizes in this ICU cohort [1]. The unknown or declined race category, about 10% of patients, also complicates interpretation [1].
Provider type and the clustering of stigma
Stigmatizing language was not evenly distributed across provider types. Nurses used stigmatizing labels at 1.40 times the rate of physicians, and social workers at 2.25 times the rate, with similar patterns for doubt markers [1]. The authors interpret this partly through exposure: nurses spend more time with patients and may document more negative behaviors, while social workers may address psychosocial contexts that elicit evaluative language [1]. Median incidence rate ratios showed much stronger clustering by patient than by provider, which the authors read as support for the Social Categories and Stereotypes Communication Framework: stereotypes operate as shared group-level processes rather than purely individual quirks [1].
This is an interpretation, not a direct test of the framework. The clustering result is consistent with collective stereotyping, but it could also reflect repeated admissions, note duplication, or documentation templates that carry language forward [1].
How CARE-SD compares with adjacent detection work
A precursor study on gender-expansive patients in MIMIC-IV built an annotated corpus of 754 notes and found that 43.37% of notes contained stigma, with misgendering present in 74.88% of descriptor-labeled gender-expansive notes [3]. That work also showed that transformer models had subgroup fairness gaps, with ClinicalBERT false-positive-rate disparity of 13.48 percentage points, reduced to 2.03 points after fairness-aware thresholding [3]. CARE-SD differs in population, dataset version, and outcome definition: it uses MIMIC-III rather than MIMIC-IV, focuses on race, insurance, diagnosis, and provider type rather than gender-expansive identity, and does not report fairness-constrained model evaluation [1][3]. Both studies agree that stigmatizing language is detectable at scale and unevenly distributed, but they measure different constructs and should not be treated as replication of one another.
A competing perspective from provider-patient communication research emphasizes that communication quality and racial concordance shape disparities through interactional pathways, not only through written documentation [4]. CARE-SD cannot adjudicate that debate because it analyzes notes, not encounters. Its contribution is to show that one measurable trace of those interactions, note language, carries a structural signature.
What the evidence does not yet show
The strongest boundary is the dataset. MIMIC-III is a single-center, de-identified ICU database from Beth Israel Deaconess Medical Center covering 2001 to 2012, with 11,630 patients in this analysis [1]. Associations between stigmatizing language and patient race, insurance, or diagnosis do not prove that language causes health disparities, and the authors explicitly avoid that causal claim [1]. The results also cannot be generalized to all EHR systems, outpatient settings, or non-ICU documentation [1]. Validation work on de-identification shows that clinical text processing is highly context-dependent, and performance in one institution or language does not automatically transfer [5]. Limitation evidence from a lung cancer diagnostic study is not directly about NLP stigma, but it illustrates the broader principle that single-cohort diagnostic performance requires external validation before clinical adoption [6].
Several open questions remain. The gender reversal in doubt markers needs targeted qualitative work to identify which clinical contexts drive it [1]. The null findings for sickle cell disease and schizophrenia need larger samples [1]. And the field still lacks prospective evidence that detecting stigmatizing language changes documentation or patient outcomes; adjacent NLP reviews report that fairness considerations and implementation guidance are often absent from clinical NLP studies [7]. CARE-SD provides a scalable measurement tool, but measurement is not yet intervention.
About These Sources
This research page is built on 7 peer-reviewed studies — published from 2014 to 2026, 5 from 2024 or later — selected as the most relevant from 13 studies that passed quality screening, drawn from 78 papers retrieved from a database of over 500 million.
Sources used in this answer
Applying Natural Language Processing to Address Healthcare Bias: Using CARE-SD to Evaluate the Distribution of Stigmatizing Language in Clinical Notes
CARE-SD combines expanded lexicons and supervised classifiers to detect stigmatizing labels and doubt markers in MIMIC-III, finding higher rates among Black patients, government-insured and self-pay patients, stigmatized diagnoses, and notes by nurses and social workers.
How words discredit: A taxonomy of stigmatizing language in the electronic health record
This foundational taxonomy work argues that stigmatizing language in EHRs discredits patients in ways that undermine the clinician-patient relationship, establishing the conceptual basis for detecting multiple forms of stigma.
Stigmatizing Language in Gender-Expansive Patient Records: Corpus Development, Disparity Analysis, and Natural Language Processing-Based Detection Study.
This precursor study built a 754-note annotated MIMIC-IV corpus for gender-expansive patients, found stigma in 43.37% of notes with misgendering dominant, and showed that fairness-aware thresholding reduced subgroup false-positive-rate disparities in transformer classifiers.
The contribution of provider–patient communication to health disparities
This competing perspective emphasizes that provider-patient communication, racial concordance, and implicit bias shape health disparities through interactional pathways, offering an alternative explanation that note-language analysis alone cannot resolve.
Harnessing Moderate-Sized Language Models for Reliable Patient Data Deidentification in Emergency Department Records: Algorithm Development …
This validation study fine-tuned moderate-sized generative models for de-identification of emergency department notes, achieving an F1 of 0.9673 with Mistral 7B, and illustrates that clinical text processing performance is context-dependent and requires careful validation.
Tumor specific DNA in bronchial lavage as a new diagnostic tool in lung cancer.
This limitation evidence from lung cancer diagnostics shows that single-cohort diagnostic performance, even with high sensitivity and specificity, requires validation before clinical adoption, reinforcing the general principle that CARE-SD's single-center findings need external confirmation.
Natural language processing techniques to detect delirium in hospitalized patients from clinical notes: a systematic review
This systematic review of delirium detection from clinical notes found that transformer models achieved high performance but that 61.5% of studies had high risk of bias, none conducted external validation, and fairness considerations were absent, defining the validation gap that CARE-SD also faces.