A Topic Modeling Analysis of Stigma Dimensions, Social, and Related Behavioral Circumstances in Clinical Notes Among Patients with HIV

Ziyi Chen, Yiyang Liu, Mattia Prosperi, Krishna Vaddiparti, Robert L. Cook, Jiang Bian, Yi Guo, Yonghui Wu

Abstract

Objective: To characterize stigma dimensions, social, and related behavioral circumstances in people living with HIV (PLWH) seeking care, using NLP methods applied to a large collection of EHR clinical notes from a large integrated health system in the southeast United States.

Methods: We identified a cohort of PLWH from the UF Health IDR and performed topic modeling analysis using Latent Dirichlet Allocation (LDA) to uncover stigma-related dimensions and related social and behavioral contexts. Domain experts created a seed list of HIV-related stigma keywords, then applied a snowball strategy to review notes for additional terms until saturation was reached iteratively. To identify more target topics, we tested three keyword-based filtering strategies. The detected topics were evaluated using three widely used metrics and manually reviewed by specialists. In addition, we conducted word frequency analysis and topic variation analysis among subgroups to examine differences across age and sex-specific demographics.

Results: We identified 9,140 PLWH at UF Health and collected 2.9 million clinical notes. Through the iterative keyword approach, we generated a list of 91 keywords associated with HIV-related stigma. Topic modeling on sentences containing at least one keyword uncovered a wide range of topic themes, such as "Mental Health Concern, Stigma," "Treatment Refusal, Isolation," and "Substance Abuse." Topic variation analysis across age subgroups revealed substantial differences.

Conclusion: Extracting and understanding the HIV-related stigma and associated social and behavioral circumstances from EHR clinical notes enables scalable, time-efficient assessment and overcoming the limitations of traditional questionnaires. Findings from this research provide actionable insights to inform patient care and interventions to improve HIV-care outcomes.

Introduction

Stigma remains a significant barrier to effective care for people living with HIV (PLWH), often leading to decreased treatment adherence, poor mental health outcomes, and social isolation. Despite its impact, measuring stigma in a clinical setting is notoriously difficult. Traditional methods, such as the Berger Stigma Scale, rely on patient surveys which are often time-consuming, expensive to administer at scale, and prone to low response rates.

However, electronic health records (EHRs) contain a vast repository of unstructured data in the form of clinical notes. These notes, written by physicians, nurses, and social workers, often capture the nuanced social and behavioral circumstances of a patient’s life that structured data (like ICD-10 codes) miss. This study explores the use of Natural Language Processing (NLP) to systematically mine these notes for dimensions of HIV-related stigma.

Methods

The research team utilized Latent Dirichlet Allocation (LDA), an unsupervised machine learning technique, to discover latent themes within the clinical text. A primary challenge in this work was the "needle in a haystack" problem: stigma-related information is often buried within millions of routine medical observations.

To address this, the researchers compared three distinct filtering strategies:

  1. Strategy 1: Modeling the entire corpus of clinical notes.
  2. Strategy 2: Modeling only the notes that contained at least one stigma-related keyword.
  3. Strategy 3: Modeling only the specific sentences containing those keywords.

To build a robust keyword list, the team employed a "snowball" strategy. Starting with a seed list of expert-defined terms, they iteratively reviewed notes for new terms. They further expanded this list using GatorTron, a large language model trained on clinical text, to identify relevant synonyms through word embeddings. This process resulted in a final list of 91 keywords.

Data and Cohort

The study drew data from the University of Florida (UF) Health Integrated Data Repository (IDR), covering the period from 2012 to 2022.

  • Cohort: 9,140 PLWH identified via a validated computable phenotype algorithm (98.9% sensitivity).
  • Corpus: Approximately 2.9 million clinical notes, including progress notes, discharge summaries, and telephone encounters.
  • Demographics: The cohort was 61.7% male and 60.1% Non-Hispanic Black, reflecting the diversity of the patient population in the Southeast United States.

Key Results and Insights

The Power of Sentence-Level Filtering

The study found that Strategy 3 (sentence-level modeling) was the most effective. When modeling entire notes, stigma-related themes were frequently diluted by routine medical jargon. By narrowing the focus to specific sentences, the proportion of stigma-related topics identified by the model rose from 20% to 82%, significantly improving the granularity of the results.

Multidimensional Stigma Themes

The LDA model identified 15 topics, which were synthesized into 11 overarching themes. These themes demonstrate that HIV stigma is inextricably linked to broader social determinants of health. Key themes included:

  • Food Insecurity and Financial Strain
  • Intimate Partner Violence and Safety
  • Treatment Refusal and Isolation
  • Substance Abuse and Mental Health Concerns
  • Missed Appointments and Care Monitoring

Demographic Variations

While the study found no significant differences in stigma documentation between male and female patients, age played a major role in the types of stigma recorded. Younger patients (ages 18–39) showed a higher probability of topics related to substance abuse and missed appointments. In contrast, older patients (60+) were more likely to have documentation regarding physical health concerns and the "fear of falling," suggesting that for older PLWH, stigma may be compounded by the challenges of aging.

Discussion and Future Work

This research demonstrates that NLP can provide a scalable, time-efficient alternative to traditional stigma assessments. By uncovering the complex web of social and behavioral circumstances documented in clinical notes, healthcare providers can gain a more holistic view of the challenges facing PLWH.

There are limitations to this approach, including the potential for "high-utilizer bias," where patients who visit the clinic more frequently have more notes and thus more documented stigma. Additionally, as a single-site study at UF Health, the specific keyword distributions may vary in other geographic regions.

Future work will focus on applying these NLP tools across multiple healthcare systems and investigating the direct correlation between these identified stigma dimensions and long-term clinical outcomes, such as viral suppression and retention in care.

Summary updated

Abstract

Objective: To characterize stigma dimensions, social, and related behavioral circumstances in people living with HIV(PLWHs) seeking care, using NLP methods applied to a large collection of EHR clinical notes from a large integrated health system in the southeast United States. Methods: We identified a cohort of PLWHs from the UF Health IDR and performed topic modeling analysis using Latent Dirichlet Allocation to uncover stigma-related dimensions and related social and behavioral contexts. Domain experts created a seed list of HIV-related stigma keywords, then applied a snowball strategy to review notes for additional terms until saturation was reached iteratively. To identify more target topics, we tested three keyword-based filtering strategies. The detected topics were evaluated using three widely used metrics and manually reviewed by specialists. In addition, we conducted word frequency analysis and topic variation analysis among subgroups to examine differences across age and sex-specific demographics. Results: We identified 9140 PLWHs at UF Health and collected 2.9 million clinical notes. Through the iterative keyword approach, we generated a list of 91 keywords associated with HIV-related stigma. Topic modeling on sentences containing at least one keyword uncovered a wide range of topic themes, such as "Mental Health Concern, Stigma", "Treatment Refusal, Isolation", and "Substance Abuse". Topic variation analysis across age subgroups revealed substantial differences. Conclusion: Extracting and understanding the HIV-related stigma and associated social and behavioral circumstances from EHR clinical notes enables scalable, time-efficient assessment and overcoming the limitations of traditional questionnaires. Findings from this research provide actionable insights to inform patient care and interventions to improve HIV-care outcomes.
BibTeX
@article{chen2026topic,
  author = {Ziyi Chen and Yiyang Liu and Mattia Prosperi and Krishna Vaddiparti and Robert L. Cook and Jiang Bian and Yi Guo and Yonghui Wu},
  title = {{A Topic Modeling Analysis of Stigma Dimensions, Social, and Related Behavioral Circumstances in Clinical Notes Among Patients with HIV}},
  journal = {International Journal of Medical Informatics},
  year = {2026},
  url = {https://openreview.net/forum?id=EEaPVD1Npv},
}
Figure