The Evolution of Small Language Models in Healthcare: A Narrative-Evolutionary Literature Review

Abstract

Small Language Models (SLMs) are emerging as transformative tools in healthcare AI, offering an optimal balance between task performance, privacy, and computational efficiency. Unlike Large Language Models (LLMs), which require substantial infrastructure and raise privacy concerns, SLMs are designed for domain-specific applications in real-time, resource-constrained clinical settings. This narrative-evolutionary review traces the development, capabilities, and deployment of SLMs across key healthcare domains, including clinical documentation, decision support, mental health, and telehealth. It outlines enabling technologies such as model compression, federated learning, and edge deployment that facilitate secure and efficient operation within sensitive healthcare infrastructures. To assess their clinical utility, we introduce a domain-relevance matrix and evaluate SLMs using both conventional and deployment-aware metrics, such as latency, memory usage, and compliance with data regulations. The review also identifies critical research gaps in benchmarking practices, data diversity, and explainability. We emphasize the importance of collaborative data ecosystems, standardized evaluation protocols, and socio-technical governance to ensure safe, multimodal, and regulation-compliant integration of SLMs in healthcare workflows. Overall, our findings position SLMs as a scalable, transparent, and context-aware solution for advancing real-world medical NLP, especially in scenarios demanding adaptability, trustworthiness, and decentralized AI performance.