NaijaMed: A Multilingual Clinical Documentation Pipeline for Nigerian Healthcare

The medical records of millions of Nigerians who consult in Yoruba, Igbo, Hausa, or Pidgin are incomplete, inaccurate, or entirely absent due to communication mismatch between patient and medical personnel. Medical doctors in Nigerian hospital settings are compelled to code-switch between English, Yoruba, and Nigerian Pidgin in the same conversation just to properly ascertain the complaints of the patient. Considering the link between consultation and health outcome, the automation of patient communication is a necessary task to optimise healthcare. The existing clinical AI documentation tools are mostly Eurocentric as they are trained only on monolingual English data from Western countries which inherently renders them unethical or inadequate for the Nigerian context. The application of these systems to Nigeria may result in the omission of significant portions of the patient's account which could result in medication error or any other issue that could negatively impact the welfare of the patient.

NaijaMed is the first complete clinical documentation system built specifically for Nigeria's multilingual healthcare settings. It takes raw audio from doctor-patient conversations and turns it into structured SOAP notes (Subjective, Objective, Assessment, Plan). The system can identify who is speaking, detect which language is being used in each part of the conversation, and route English, Yoruba, and code-switched Pidgin-English through separate processing lanes. It also uses a specially tuned version of Google DeepMind's MedGemma model. The prototype has been tested on real Nigerian consultation recordings and performs well across the three most common language situations in Nigerian clinical practice. Preliminary evaluation using NaijaEval — an open-source evaluation toolkit purpose-built for African language AI systems — yields a code-switch rate of 0.004, a medical terminology preservation rate of 0.050, and a hallucination rate of 0.000, confirming the system's clinical safety on the current test corpus and identifying medical terminology coverage as the primary target for improvement. In this paper, we describe the full system design, the thinking behind it, a preliminary quantitative evaluation, current limitations, and a step-by-step data collection plan to eventually include Igbo, Hausa, and more complex forms of code-switching.

Keywords: clinical NLP · automatic speech recognition · code-switching · Nigerian Pidgin · Yoruba · multilingual healthcare AI · SOAP note generation · MedGemma · low-resource languages · African health informatics