Thematic modeling of digital literacy using language models: a qualitative-computational analysis of interviews on hearing health

Abstract

Objective: To describe thematic patterns of digital health literacy among adults with hearing loss and to evaluate the feasibility of using small-scale language models for the qualitative categorization of clinical interviews. Method: This is a cross-sectional, qualicomputational methodological proof-of-concept study. Transcripts of 15 semi-structured interviews with adults diagnosed with hearing loss were analyzed. A model based on MPNet (Masked and Permuted Pre-training for Language Understanding) was employed. Topic extraction was performed using the BERTopic method in zero-shot mode (automatic classification without prior training on the categories). The responses were organized into eight predefined thematic categories, derived directly from the domains addressed in the interview script. Results: The computational approach demonstrated feasibility in the extraction and structured clustering of the responses provided by the participants. The findings indicate a predominantly reactive trigger for digital literacy: information seeking begins mainly after diagnosis, with behaviors ranging from incidental to proactive and technologically proficient. Conclusion: The data indicate that family mediation contributes to these patients’ access to digital technologies and to the filtering of information. Audiological rehabilitation may benefit from training the family support network to also act as mediators of digital literacy. The findings suggest that small-scale language models constitute a feasible methodological tool to support qualitative research in speech-language pathology.