Abstract:

Access to digital healthcare in Nigeria remains limited by language barriers, as most medical platforms are designed primarily in English, while a large portion of the population communicates in indigenous languages. This gap often prevents users from accurately describing symptoms, restricting effective communication and access to appropriate healthcare services. This study presents a multilingual medical diagnosis system capable of accepting symptom descriptions in English, Yoruba, Igbo, Hausa, and Nigerian Pidgin. The system generates ranked differential diagnoses with clinically grounded descriptions and treatment recommendations delivered in the user’s preferred language. Using advances in Natural Language Processing, the framework employs the Aya Expanse 8B model, deployed through Ollama, to translate indigenous inputs into English for downstream analysis. Treatment recommendations are generated through a Retrieval-Augmented Generation (RAG) architecture integrating WHO Essential Medicines guidelines, Federal Ministry of Health (FMOH) protocols, NIH MedlinePlus, and OpenFDA resources. Output reliability is improved through deterministic description generation, a three-tier language refinement pipeline, and hallucination detection algorithms, while Supabase manages authentication and interaction history. Evaluation results show strong performance, achieving 89% Top-3 accuracy, low hallucination rates, and high linguistic reliability, demonstrating the effectiveness of the proposed hybrid deterministic-RAG framework for multilingual healthcare support in low-resource African languages.