# Artificial intelligence and hearing health: a global evidence review of biases and equity implications for Africa

## TL;DR

This paper reviews how AI in hearing health can worsen inequities in Africa due to biases in data and design.

## Contribution

It provides the first global synthesis of AI biases in audiology and proposes a framework for equitable AI integration in low-resource settings.

## Key findings

- Six types of AI biases were identified, with representation bias being most common due to language limitations.
- Lack of African audiogram and speech data leads to poor AI performance in local contexts.
- Equitable AI requires local datasets, fairness-aware models, and regionally led governance.

## Abstract

Artificial intelligence (AI) is increasingly integrated into audiology and hearing health, yet evidence from across the health sciences shows that AI systems routinely embed structural biases that can exacerbate inequities, particularly for African and other low- and middle-income country (LMIC) populations. This review identified and analysed bias types in AI applications relevant to audiology and examined their ethical, cultural, and linguistic implications for LMIC settings. A narrative review design was adopted to accommodate the heterogeneity of available evidence, where thematic saturation was more appropriate than effect-size aggregation. Peer-reviewed articles published between 2015 and 2025 were retrieved from PubMed, Scopus, Web of Science, and IEEE Xplore, with inclusion requiring explicit engagement with AI and bias or equity. Rigour was assessed using a six-domain quality rubric, and data were extracted into structured evidence tables for thematic synthesis. Thirty-three studies met inclusion criteria: six were audiology-specific empirical studies (all small scale), and the remainder were reviews or conceptual analyses. No study presented empirical African audiogram, auditory brainstem response (ABR), or speech data. Six recurrent bias types were identified; representation, measurement, algorithmic, evaluation, deployment, and intersectional, with representation bias most frequent, exemplified by English-only corpora that underperform on tonal or indigenous languages. These biases manifest as misclassified hearing loss, reduced ABR accuracy, inequitable hearing-aid personalisation, and poor cochlear-implant algorithm transferability. Advancing equitable AI in audiology requires multilingual, paediatric-inclusive, locally governed datasets; fairness-aware model design with stratified reporting; and African-led governance and capacity-building to support future validation and implementation research.

Main findings: This article provides the first global synthesis of how biases in artificial intelligence translate into clinical risks for hearing assessment, diagnosis, and rehabilitation.Added knowledge: It highlights how linguistic diversity, infrastructural limitations, and externally driven technology create distinct vulnerabilities in African and other low-resource contexts.Global health impact for policy and action: The article advances a framework for equitable integration of artificial intelligence in hearing health through locally owned datasets, fairness-aware design, and regionally led governance.

Main findings: This article provides the first global synthesis of how biases in artificial intelligence translate into clinical risks for hearing assessment, diagnosis, and rehabilitation.

Added knowledge: It highlights how linguistic diversity, infrastructural limitations, and externally driven technology create distinct vulnerabilities in African and other low-resource contexts.

Global health impact for policy and action: The article advances a framework for equitable integration of artificial intelligence in hearing health through locally owned datasets, fairness-aware design, and regionally led governance.