Building Inclusive Health Communication: A Collaborative Libras Datset for Deaf Health Equity

Introduction and Objectives Deaf communities continue to face barriers to accessing health information and services, while Latin American sign languages remain severely underrepresented in the datasets used to develop digital and artificial intelligence (AI) technologies. In Brazil, Brazilian Sign Language (Libras) is used by approximately 2.7 million Deaf and hard-of-hearing people, yet its representation in AI training data remains limited. This creates a risk that emerging technologies will reproduce existing inequalities in access to health. Data-Pop Alliance, with support from Mozilla, developed a multimodal dataset specifically focused on health and safety communication. This project aims to strengthen the foundations for more accessible health technologies while demonstrating how community participation and ethical data governance can be embedded in the development of health-related AI infrastructure. Methodology The dataset comprises 24.6 hours across 44 interviews, totaling 82.6GB of multimodal Libras data covering basic hygiene, disease prevention, accessing health services, emergency assistance, and understanding public health guidance. Data were collected in São Paulo and São José do Rio Preto, capturing variation across urban and periurban/rural contexts. Participants included Deaf community members, certified Libras educators, and healthcare interpreters. Each recording is accompanied by gloss-level Libras annotations, Brazilian Portuguese translations, and structured metadata. The project followed a participatory and community-led approach involving Deaf communities, Libras experts, interpreters, and an independent ethical body comprising Deaf representatives, language experts, and data ethics specialists. Results and Discussion The resulting dataset, freely accessible for members of the Deaf community, researchers, and interpreters, provides an infrastructure for research and development in sign-language recognition, Libras-to-Portuguese translation, multimodal AI, and health communication technologies for Deaf users. Beyond its technical contribution, the project demonstrates that addressing digital exclusion in health requires attention to who produces health-related data, whose language is represented, and who has a role in governing its use. Community participation shaped data collection, ethical review, consent procedures, and decisions around representation. Conclusion When sign languages are absent from training corpora, Deaf communities face additional barriers in accessing digital technologies, public communication systems, and emerging AI tools. These exclusions reinforce broader social inequalities and contribute to the invisibility of linguistic minorities within digital environments. By focusing on Libras and healthcare-related communication, the project demonstrates that inclusive AI development requires both technical innovation and representative data infrastructures grounded in accessibility, care, and linguistic justice. More broadly, the experience highlights the importance of treating linguistic representation as a component of equitable health communication and of embedding affected communities in the design and governance of emerging health technologies.