Machine learning for depression and anxiety prediction in African contexts: a systematic review

Abstract

The underdiagnosis of depression and anxiety across Africa is partly due to the lack of adequate mental health care provision, institutionalized stigma and the use of expensive clinical assessment tools. While machine learning, deep learning, and natural language processing techniques have been suggested to aid with scaling prediction and screening, there is a lack of literature that is specific to Africa and similar low-resource contexts. This systematic review, carried out within the framework of PRISMA 2020 covering literature between 2018 and 2026, brings together 30 studies on various machine learning models for depression and anxiety prediction, with an emphasis on African-context studies. Studies were selected using structured keyword searches, checked for strict eligibility criteria and were double-checked for bibliographic accuracy. Tree-based ensemble methods and logistic regression are more prevalent in African-context studies, while language-based and multimodal approaches are still comparatively underdeveloped, with very little evidence about the use of natural language processing with African languages. There are also some cross-cutting issues such as inconsistent validation practices, lack of external validation, and reported algorithmic bias risks when models or sensing hardware trained using populations outside of Africa are applied in African contexts. The review ends with a research agenda that emphasizes the creation of paired multimodal datasets in Africa, NLP pipelines in African languages, frameworks for bias-auditing in African deployment environments, and reporting on outcomes to facilitate future synthesis through meta-analysis.