Cultural and Language Perspectives on AI Adoption in Diabetes Management

Abstract

Diabetes mellitus affects over 800 million people worldwide, yet most AI-based diabetes management applications target English-speaking populations in high-resource settings. This limits access to culturally and linguistically appropriate support for Amharic-speaking users. This thesis investigates barriers to AI-based diabetes management for Amharic-speaking users, evaluates the performance of LLM-based assistants in Amharic compared with English, and develops requirements for a culturally adapted system.

This research used a mixed-method exploratory sequential design with four stages. These included a gap analysis of four diabetes apps, evaluation of 72 responses generated by ChatGPT, Gemini, and Microsoft Copilot, interviews with five Amharic speakers, and requirements synthesis.

The findings showed that existing diabetes applications lacked Amharic language support, script compatibility, and culturally relevant food databases. The evaluated AI assistants performed less effectively in Amharic than in English, particularly in language quality and cultural adaptation. Many Amharic responses also contained mixed English and Amharic text, reducing readability and usability. Responses to culturally specific queries generally performed better than generic queries, suggesting that cultural context may improve AI response quality.

Overall, 37 requirements in five dimensions were identified through this study, where 16 are classified as MUST. The findings show that language and cultural adaptation contribute more to how well an AI-based diabetes system performs for Amharic-speaking users