June 2026

Using AI to Listen: How Large Language Models Are Unlocking Health Care Stories from the African American Community

A team of researchers* at the University of Pittsburgh has developed a computational approach to extract meaningful insights from spoken health care narratives, which they tested on a set of stories that focus on the experiences of African American patients, caregivers, and healthcare providers.. We hope that this work may make it easier to learn from PaTH’s Story Booth narrative archive. Published in JMIR Medical Informatics, the study demonstrates how large language models (LLMs) can be combined with a classical topic modeling technique — latent Dirichlet allocation — to automatically transcribe, categorize, and summarize long-form audio stories. Rather than relying on traditional methods like post-visit surveys or social media analysis, the researchers drew on the MyPaTH Story Booth archive, a collection of first-person health care narratives shared by community members. Fifty stories from African American participants were analyzed, yielding 26 distinct topics — including health behaviors, doctor-patient relationships, caregiving, cancer treatment, and chronic pain management — that shed light on the lived realities and systemic gaps these individuals have encountered in their health care journeys.

A key innovation of the study was its use of a hierarchical summarization method to overcome the technical limitations of LLMs when processing lengthy narratives, paired with an evaluation framework that assessed whether AI-generated summaries were accurate, useful, and free from fabrication. When assessed by GPT-4 and validated against two domain experts, the summaries demonstrated moderate-to-high reliability, suggesting that LLM-based evaluation can serve as a scalable alternative to costly human review. The authors acknowledge limitations — including minor transcription errors that can propagate through the pipeline and the relatively small dataset — but emphasize the broader promise of the approach. By harnessing the communicative power of storytelling, this research offers a scalable pathway for health informatics researchers to surface nuanced, community-grounded insights that could potentially inform more equitable clinical care and health policy for underserved populations.

*Maneesh Bilalpur, Megan Hamm, Young Ji Lee, Natasha Norman, Kathleen Mctigue, and Yanshan Wang

«—- Back To News