Automatic Evaluation of Mental Health Stigma in Online Communication

While large language models (LLMs) are increasingly being explored for mental health applications, recent studies reveal that they can exhibit stigma toward individuals with psychological conditions. Existing evaluations of this stigma primarily rely on multiple-choice questions (MCQs), which fail to capture the biases embedded within the models' underlying logic. In this paper, we analyze the intermediate reasoning steps of LLMs to uncover hidden stigmatizing language and the internal rationales driving it. We leverage clinical expertise to categorize common patterns of stigmatizing language directed at individuals with psychological conditions and use this framework to identify and tag problematic statements in LLM reasoning. Furthermore, we rate the severity of these statements, distinguishing between overt prejudice and more subtle, less immediately harmful biases. To broaden the reasoning domain and capture a wider array of patterns, we also extend an existing mental health stigma benchmark by incorporating additional psychological conditions. Our findings demonstrate that evaluating model reasoning not only exposes substantially more stigma than traditional MCQ-based methods but also helps identify the flaws in the LLMs' logic and their understanding of mental health conditions.

Multilingual large language models (LLMs) are increasingly used in socially sensitive mental health contexts, including support chatbots, screening, and content moderation. This raises a reliability question: do semantically equivalent mental health inputs elicit comparable evaluations across languages, or systematic shifts consistent with language-associated social and cultural contexts? We examine this question in an English-Chinese setting with GPT-4o and Qwen3-32B using a two-level framework: construct-level evaluative orientation, measured by psychometric stigma instruments, and decision-level behavior, measured by binary stigma detection and four-class depression severity classification. Across instruments and models, Chinese prompts elicit higher stigma-related scores than English prompts. At the decision level, Chinese prompts reduce sensitivity to stigmatizing content and produce more conservative depression severity judgments, leading to more under-estimation errors. These findings show that prompt language can shift both evaluative orientation and downstream behavior in LLM-based mental health evaluation. They highlight the need to evaluate multilingual LLMs not only for aggregate performance, but also for whether they apply comparable evaluative standards across languages in socially sensitive domains.

Human immunodeficiency virus (HIV)-related stigma is a critical psychosocial determinant of health for people living with HIV (PLWH), influencing mental health, engagement in care, and treatment outcomes. Although stigma-related experiences are documented in clinical narratives, there is a lack of off-the-shelf tools to extract and categorize them. This study aims to develop a large language model (LLM)-based tool for identifying HIV stigma from clinical notes. We identified clinical notes from PLWH receiving care at the University of Florida (UF) Health between 2012 and 2022. Candidate sentences were identified using expert-curated stigma-related keywords and iteratively expanded via clinical word embeddings. A total of 1,332 sentences were manually annotated across four stigma subscales: Concern with Public Attitudes, Disclosure Concerns, Negative Self-Image, and Personalized Stigma. We compared GatorTron-large and BERT as encoder-based baselines, and GPT-OSS-20B, LLaMA-8B, and MedGemma-27B as generative LLMs, under zero-shot and few-shot prompting. GatorTron-large achieved the best overall performance (Micro F1 = 0.62). Few-shot prompting substantially improved generative model performance, with 5-shot GPT-OSS-20B and LLaMA-8B achieving Micro-F1 scores of 0.57 and 0.59, respectively. Performance varied by stigma subscale, with Negative Self-Image showing the highest predictability and Personalized Stigma remaining the most challenging. Zero-shot generative inference exhibited non-trivial failure rates (up to 32%). This study develops the first practical NLP tool for identifying HIV stigma in clinical notes.