Multilingual Large Language Models in Resource-Constrained Healthcare: A Structured Review of RAG Integration and Deployment Challenges

Abstract

Artificial intelligence may help address healthcare workforce shortages and linguistic barriers, particularly in resource-constrained settings. However, the practical use of large language models in healthcare remains limited by computational requirements, English-language bias, scarcity of reliable regional-language medical data, and the risk of hallucinated information. This review examines multilingual healthcare large language models, focusing on architectural approaches, multilingual adaptation, retrieval-augmented generation, and cloud-edge deployment trade-offs. It also traces the development of healthcare chatbots from rule-based systems to transformer-based models and compares representative approaches according to language support, factual grounding, safety, and deployment feasibility. The reviewed evidence suggests that retrieval from curated medical sources can improve factual grounding, while native multilingual adaptation may preserve clinical context more effectively than translation-only pipelines. Nevertheless, reported results are difficult to compare because studies use different datasets, languages, hardware, and evaluation procedures. Future research should prioritize reproducible evaluation, real-world clinical validation, privacy protection, human oversight, and efficient offline deployment.