About

Mary Catherine Beach, MD, MPH, is a professor of medicine at the Johns Hopkins University School of Medicine and holds a joint appointment in the Department of Health, Behavior & Society at the Johns Hopkins Bloomberg School of Public Health. Her scholarship focuses on respect and relationships in healthcare, encompassing both empirical and conceptual dimensions. Her empirical work has primarily centered on respect and communication between patients and clinicians, with recent research concentrating on people living with HIV/AIDS and sickle cell disease. Dr. Beach's research has been funded by prominent institutions including the National Institutes of Health, the Agency for Healthcare Research and Quality, the Robert Wood Johnson Foundation, and the Greenwall Foundation. She serves on the editorial board for Patient Education and Counseling and on the Advisory Board for Communication in Medicine. At Johns Hopkins, she is co-chair of an Institutional Review Board (IRB), Course Director of the Scholarly Concentrations Program in the School of Medicine, and Director of the TL1 Predoctoral Clinical Research training program, contributing significantly to education and research in healthcare communication, respect, and patient-centered care.

Research topics

• Linguistics
• Family medicine
• Psychology
• Medicine

Selected publications

• Artificial Intolerance: Stigmatizing Language in Clinical Documentation Skews Large Language Model Decision-Making

  ArXiv.org · 2026-05-17

  articleOpen access

  Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as clinical decision support and medical documentation. However, the robustness of these models against subtle linguistic variations, specifically stigmatizing language (SL) commonly found in human-authored clinical notes, remains critically under-explored. In this work, we investigate whether frontier LLMs inherit and propagate this human bias when processing clinical text. We systematically evaluate nine frontier LLMs across four stigmatized medical conditions, utilizing clinical vignettes injected with varying intensities and phenotypes of SL (doubt, blame, and maligning). Our results demonstrate that all evaluated models exhibit substantial bias, with clinical decision-making significantly skewed towards less aggressive patient management. Notably, we observe a high sensitivity to linguistic framing, where a single SL sentence is sufficient to alter model outputs, revealing a clear dose-response relationship. Furthermore, we evaluate standard prompt-based mitigation strategies, including Chain-of-Thought (CoT) reasoning and model self-debiasing. These approaches show limited efficacy; models struggle to explicitly identify SL while remaining implicitly influenced by it. Our findings expose a critical vulnerability in current LLMs regarding fairness and robustness in clinical NLP, underscoring the need for rigorous algorithmic guardrails to prevent the automation of health disparities.

  PublisherOA PDF

• Physician Attitudes Embedded Within Electronic Medical Records of Persons With and Without Serious Mental Illness

  Schizophrenia Bulletin · 2026-03-01

  articleOpen accessSenior author

  BACKGROUND AND HYPOTHESIS: Negative descriptors and stigmatizing language of patients have been found in electronic medical records. Using electronic health records, this study sought to describe language patterns that reflected positive or negative attitudes for patients with and without a serious mental illness. STUDY DESIGN: Content analysis was performed on ambulatory internal medicine progress notes from patients with serious mental illness (schizophrenia, bipolar disorder, major depression with psychosis; n = 511) and a control population (n = 511), matched on age, sex, and race. Fisher's exact test was used to compare frequencies of identified themes across groups. STUDY RESULTS: Language reflecting negative attitudes appeared at greater frequency in notes of patients with serious mental illness than without (18.0% vs 11.7%, P = .006). Language expressing physicians' frustration (5.7% vs. 2.2%, P = .005) and questioning patient's credibility (5.1% vs. 1.8%, P = .0005) also appeared more frequently in notes for patients with serious mental illness (vs. without). Language reflecting positive attitudes appeared at similar rates (37.0% vs. 34.4%, P = .4). Compliments or approval of patients (16.8% vs. 8.0%, P < .001) or shared decision-making statements (8.2% vs. 4.7%; P = .03) were more frequent among patients with mental illness than without. CONCLUSIONS: Physicians may have differential emotional engagement-positive and negative-during visits with patients with serious mental illness. While not common, negative language and questioning patient credibility occurred more frequently in notes for patients with SMI than without. Given the stigma surrounding mental illness, work is needed to understand whether negative attitudes conveyed in notes are associated with poorer quality of care.

  PublisherDOI

• Artificial Intolerance: Stigmatizing Language in Clinical Documentation Skews Large Language Model Decision-Making

  arXiv (Cornell University) · 2026-05-17

  preprintOpen access

  Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as clinical decision support and medical documentation. However, the robustness of these models against subtle linguistic variations, specifically stigmatizing language (SL) commonly found in human-authored clinical notes, remains critically under-explored. In this work, we investigate whether frontier LLMs inherit and propagate this human bias when processing clinical text. We systematically evaluate nine frontier LLMs across four stigmatized medical conditions, utilizing clinical vignettes injected with varying intensities and phenotypes of SL (doubt, blame, and maligning). Our results demonstrate that all evaluated models exhibit substantial bias, with clinical decision-making significantly skewed towards less aggressive patient management. Notably, we observe a high sensitivity to linguistic framing, where a single SL sentence is sufficient to alter model outputs, revealing a clear dose-response relationship. Furthermore, we evaluate standard prompt-based mitigation strategies, including Chain-of-Thought (CoT) reasoning and model self-debiasing. These approaches show limited efficacy; models struggle to explicitly identify SL while remaining implicitly influenced by it. Our findings expose a critical vulnerability in current LLMs regarding fairness and robustness in clinical NLP, underscoring the need for rigorous algorithmic guardrails to prevent the automation of health disparities.

  PublisherDOI

• Exploring the Patient’s Lifeworld: A Qualitative Study of Personalizing Language in Electronic Health Records

  Journal of General Internal Medicine · 2026-03-31

  articleOpen accessCorresponding

  BACKGROUND: Physicians' understanding of patients as persons can bolster relationships and help patients feel seen. Including personal details in electronic health records (EHR) may enhance care, but the types of details physicians document about patients' lifeworlds remain largely uncharacterized. OBJECTIVE: To assess the prevalence and types of personalizing language written by physicians in clinical encounter notes. DESIGN: Cross-sectional qualitative content analysis of physician-written history and physical (H&P) notes. PATIENTS: Adult patients admitted to the internal medicine inpatient service at an urban academic tertiary care hospital. APPROACH: We performed an inductive content analysis of H&P notes. Two authors independently coded instances of personalizing language: details about patients' unique identities, social context, life experiences, and perspectives. The study team met periodically to discuss examples and categories and compare assessments until thematic saturation was reached. We summarized the prevalence of personalizing language and described its content. KEY RESULTS: We reviewed 570 H&P notes written about 437 unique patients (median age 57; 53% Black, 38% White; 54% male) by 186 physicians. More than half of notes (60%) contained at least one instance of personalizing language (median 2, range 1-7 per note). We identified 840 instances of personalizing language across seven domains: family (41%), work/education (32%), residence (21%), personal interests (2.9%), pets (1.8%), travel (0.6%), and goals/priorities (0.6%). Nearly half (47%) referred to general social context; 51% provided more detailed information. Only 1.9% described patients' thoughts, feelings, or goals. Personalizing language was more commonly used in notes for men vs women (70% vs 60%, p = 0.02). CONCLUSIONS: Most H&P notes included personalizing language, though many provided limited detail and very few described patients' interests, goals, and priorities. Future studies should evaluate patient perceptions of personalizing language in the medical record and its impact on physician attitudes towards and relationships with patients.

  PublisherOA PDFDOI

• 315. Estimating Therapeutic Alliance From Clinical Interview Sessions With Large Language Models: A PREDiCTOR Study

  Biological Psychiatry · 2026-04-25

  article
  PublisherDOI

• How words discredit: A taxonomy of stigmatizing language in the electronic health record

  Patient Education and Counseling · 2026-02-05

  articleOpen access

  OBJECTIVE: Language in electronic health records (EHRs) can transmit stigma, discrediting patients in ways that undermine the clinician-patient relationship and compromise future care. We sought to develop a taxonomy of stigmatizing language in EHRs to understand what patients are being stigmatized for, how that stigma is conveyed linguistically, and why. METHODS: We conducted a two-stage qualitative analysis of EHR notes from multiple clinical contexts in a large U.S. academic health system. For both stages, we drew enriched samples using natural language processing (NLP) to identify notes with at least one stigmatizing keyword from prior studies. First, we open coded 296 notes to generate categories of stigmatizing language and linguistic mechanisms, and to develop a preliminary taxonomy. We then applied and refined this framework by coding 400 additional notes. RESULTS: We identified six categories of stigmatizing sentiments characterizing patients as: (1) Socially Undesirable, (2) Difficult to Interact With, (3) Incompetent, (4) Manipulative, (5) Noncompliant, and (6) Not Credible. These were implied through negative descriptions of patient behavior portraying them as, e.g., Demanding, Adversarial, Deceptive, etc. Linguistic mechanisms extended beyond keywords, including practices for emphasizing the intensity of patient behavior (e.g., intensifiers), marking distance or divergence from the patient's perspective (e.g., skeptical evidentials), and casting the clinician as the neutral or rational party (e.g., euphemisms). CONCLUSION: Stigmatizing language in EHRs is not limited to discrete terms but is embedded in broader linguistic practices that shape how patients are represented and understood, particularly those describing how they fail to align with clinical expectations. This language may serve to document professional challenges, but it nonetheless reinforces paternalistic norms and compromises care. Understanding these dynamics is critical for moving toward patient-centered documentation and reducing harm in the EHR.

  PublisherDOI

• Stigmatizing Mothers: Qualitative Analysis of Languagein Prenatal Records

  Narrative Inquiry in Bioethics · 2025-01-01 · 4 citations

  articleSenior author

  Pregnant people experience moral judgment in healthcare settings that may be coded into clinical documentation. Stigmatizing language in medical records transmits bias between clinicians, potentially exacerbating disparities in maternal morbidity and mortality. We examined obstetrical records from 100 randomly selected patients who received prenatal and delivery care in an academic hospital system. Qualitative analysis sought to identify linguistic features conveying negative attitudes or moral judgment, revealing themes of epistemic injustice: (1) discrediting patient testimony as incompetent, unreliable, and hysterical; (2) unnecessary details that are objectifying, stigmatizing, or unprofessional; and (3) judgments of maternal fitness, where women are labeled “bad mothers” by emphasizing neglectful, selfish, and debauched characteristics. We conclude by advocating for further validation of our findings, revisiting medical education paradigms, and supporting the development of natural language processing (NLP) technologies to detect and intercept stigma.

  PublisherDOI

• Stigmatizing Mothers: Qualitative Analysis of Language in Prenatal Records

  Narrative Inquiry in Bioethics · 2025-01-01

  articleOpen accessSenior author

  Pregnant people experience moral judgment in healthcare settings that may be coded into clinical documentation. Stigmatizing language in medical records transmits bias between clinicians, potentially exacerbating disparities in maternal morbidity and mortality. We examined obstetrical records from 100 randomly selected patients who received prenatal and delivery care in an academic hospital system. Qualitative analysis sought to identify linguistic features conveying negative attitudes or moral judgment, revealing themes of epistemic injustice: (1) discrediting patient testimony as incompetent, unreliable, and hysterical; (2) unnecessary details that are objectifying, stigmatizing, or unprofessional; and (3) judgments of maternal fitness, where women are labeled "bad mothers" by emphasizing neglectful, selfish, and debauched characteristics. We conclude by advocating for further validation of our findings, revisiting medical education paradigms, and supporting the development of natural language processing (NLP) technologies to detect and intercept stigma.

  PublisherOA PDFDOI

• What changes minds? persuasive communication in decision-making for elective pediatric surgery

  Patient Education and Counseling · 2025-07-15

  article
  PublisherDOI

• A National Survey—Do Women 65 Years or Older Perceive Persuasion to Stop Getting Mammograms as Ethically Appropriate?

  Journal of General Internal Medicine · 2025-04-03 · 1 citations

  letterOpen access
  PublisherOA PDFDOI

Recent grants

• Communication Quality and Patient Outcomes in HIV/AIDS and Substance Abuse

  NIH · $923k · 2015–2022

• Maximizing respect and improving outcomes in HIV and substance abuse (MaRIPOHSA)

  NIH · $3.1M · 2014–2022

• NIH Grant K08HS013903

  NIH · $606k · 2008

• NIH Grant R34MH089279

  NIH · $707k · 2013

• Respect, Trust, and Patient Outcomes in Sickle Cell Disease

  NIH · $3.9M · 2009–2016

Frequent coauthors

• Somnath Saha

  204 shared

• Sophie Lanzkron

  Johns Hopkins Hospital

  157 shared

• Carlton Haywood

  Johns Hopkins Medicine

  153 shared

• Richard D. Moore

  Johns Hopkins University

  146 shared

• Lisa A. Cooper

  Johns Hopkins University

  111 shared

• Eric B Bass

  Johns Hopkins University

  86 shared

• P. Todd Korthuis

  Oregon Health & Science University

  80 shared

• John J. Strouse

  Duke University

  78 shared

Education

• MD

  Icahn School of Medicine at Mount Sinai

Awards & honors

• Greenwall fellow (Bioethics and Health Policy)
• Robert Wood Johnson Foundation’s Generalist Physician Facult…