Designing a Swahili-Speaking Medical Chatbot for Oncology, Dermatology, and Otorhinolaryngology Care in Low-Resource Settings

Abstract

Access to specialized medical care in fields such as otorhinolaryngology, dermatology, and oncology remains critically limited in low-resource settings, particularly among Swahili-speaking populations in East Africa. This paper presents Lueji, a Swahili-speaking medical chatbot designed to assist in providing accessible health information in these domains. A major challenge in building such systems lies in the scarcity of domain-specific Swahili-language medical corpora. To address this, we constructed a Swahili dataset by automatically translating the Huatuo-26M Chinese medical corpus using the GPT-4o-mini model, followed by dataset cleaning, normalization, and human validation. The translated dataset was used to fine-tune the UlizaLlama model, resulting in a specialized chatbot capable of generating medically relevant responses in Swahili. Evaluation using BLEU, ROUGE, and GLEU metrics demonstrated significant performance gains over a baseline model, with Lueji achieving BLEU-1 of 22.17, ROUGE-1 of 31.68, and GLEU of 10.77. These results highlight the effectiveness of combining automatic translation with supervised fine-tuning for low-resource language adaptation. Future work will focus on expanding the dataset, incorporating real-time retrieval-augmented generation (RAG), and conducting field testing with healthcare professionals and patients.