Language serves as the fundamental bridge between health information and health action. The manner in which medical guidance is communicated its vocabulary, tone, cultural framing, and emotional resonance profoundly influences whether individuals understand, trust, and ultimately act upon health advice. In Nigeria, Africa's most populous nation, a significant linguistic paradox exists while English functions as the official language of government, education, and formal healthcare, Nigerian Pidgin operates as the true lingua franca, spoken by an estimated 75-100 million people across ethnic, regional, and socioeconomic boundaries. This linguistic reality creates a critical gap in digital health communication, where life-saving information about HIV prevention, treatment adherence, and stigma reduction often fails to reach those who need it most not because the information is unavailable, but because it is delivered in a language and register that feels foreign, clinical, and culturally distant.The emergence of large language models (LLMs) presents an unprecedented opportunity to address this communication gap. Models such as GPT, LLaMA, and others have demonstrated remarkable capabilities in generating human-like text, answering complex questions, and adapting to specialised domains. However, these capabilities remain overwhelmingly concentrated in high-resource languages, perpetuating what scholars have termed "systematic linguistic inequality" in artificial intelligence. Nigerian Pidgin, despite its millions of speakers and vital role in everyday communication, remains critically underrepresented in AI development. When existing models attempt to process Pidgin, they frequently produce outputs that are grammatically awkward, culturally tone-deaf, or entirely incomprehensible failing to capture the pragmatic markers, politeness conventions, and culturally grounded expressions that give Pidgin its communicative power.

This research addresses this gap through the development and evaluation of the Pidgin Health Language Assistant (PHLA), a fine-tuned large language model designed specifically to generate culturally appropriate and linguistically authentic health information in Nigerian Pidgin. Adopting the Design Science Research (DSR) paradigm, the study follows a systematic artefact development process encompassing problem identification, solution design, implementation, and rigorous evaluation. The research is guided by one main research question - How can large language models be fine-tuned to generate health information in Nigerian Pidgin that is both linguistically accurate and culturally appropriate? supported by five sub-research questions aligned with the DSR process phases and the Business Process Management (BPM) lifecycle.The implementation involved three core technical contributions. First, a culturally annotated dataset of 396 HIV-related question-answer pairs was developed, extending the foundational work of Martin et al. (2024) on embedding-based health information retrieval. Each entry was translated into Nigerian Pidgin using a hybrid methodology combining AI-assisted drafting with comprehensive native speaker verification, and annotated for cultural features including tone, politeness markers, and stigma sensitivity. Second, the LLaMA-3 8B model was fine-tuned using parameter-efficient techniques specifically Low-Rank Adaptation (LoRA) with 4-bit quantisation (QLoRA) enabling effective adaptation within the computational constraints of accessible platforms such as Google Colab. This approach trained only 8.4 million parameters (0.104% of the base model), demonstrating that cultural AI adaptation need not require massive computational resources. Third, the fine-tuned model was deployed through an accessible Gradio-based web interface on Hugging Face Spaces, implementing a hybrid architecture that combines rule-based safety mechanisms with FAQ matching and LLM generation capabilities.The evaluation centred on human judgment, recognising that automated metrics cannot adequately assess the cultural and pragmatic dimensions essential to effective health communication. A cross-sectional survey was administered to 64 Nigerian Pidgin speakers over six weeks, assessing PHLA's responses across four theoretically grounded dimensions: clarity (comprehensibility), politeness (respectful and culturally appropriate tone), naturalness (authentic Pidgin expression), and trustworthiness (credibility of health information). The findings demonstrate strong positive reception: PHLA achieved a grand mean rating of 4.51 on a 5-point scale, with 91.5% of all ratings falling in the positive range (4 or 5). Clarity received the highest ratings (M=4.61), confirming that PHLA successfully makes health information accessible. Trustworthiness (M=4.54) indicated that users find the health content credible a critical factor for behaviour change. Politeness (M=4.49) confirmed appropriate navigation of cultural sensitivities, while naturalness (M=4.41), though slightly lower, still demonstrated authentic Pidgin expression with identified opportunities for regional adaptation. Qualitative feedback revealed that participants particularly valued PHLA's emotional resonance and practical relevance, while suggesting enhancements including regional variant support and more conversational register.This research makes contributions across theoretical, methodological, and practical domains. Theoretically, it demonstrates how constructs from cultural linguistics, politeness theory, and health communication can be operationalised in AI development, providing a framework for embedding cultural considerations in training data. Methodologically, it offers a replicable human-centred evaluation approach for assessing culturally adapted AI systems. Practically, it delivers PHLA as a working proof-of-concept, an annotated dataset as a resource for future research, and design guidelines for culturally adapted health AI. The findings carry implications for health communication practice in Nigeria and similar multilingual contexts, for AI developers working on low-resource languages, and for policy discussions regarding linguistic equity in digital health.

In conclusion, this research demonstrates that large language models can be effectively adapted to serve speakers of low-resource languages through culturally grounded dataset development, parameter-efficient fine-tuning, and human-centred evaluation. PHLA represents a meaningful step toward ensuring that the benefits of AI-powered health communication extend beyond speakers of dominant global languages to reach communities in their own voices, idioms, and cultural frameworks. As AI increasingly shapes how health information reaches communities worldwide, ensuring that these technologies speak all languages including Nigerian Pidgin is not merely a technical challenge but a matter of health equity and linguistic justice.