A Health-Focused Risk Taxonomy for AI: Assessing Unsafe Content Detection with Small Language Models (SLMs)

Large Language Models (LLMs) show promise in healthcare. To make the most of this technology, there is a need to address concerns about computational demands and privacy. Small Language Models (SLMs) offer a privacy-preserving alternative for specialised medical applications due to their lower resource needs and potential for local deployment. This paper examines existing LLM safeguarding frameworks and introduces a novel, health-focused risk taxonomy developed through literature review and co-design with healthcare professionals. Furthermore, the ability of 6 SLMs to detect unsafe content using 2 additional risk taxonomies is evaluated and compared. The 8b-parameter Granite Guardian model showed superior adaptation to the novel risk taxonomy (75% accuracy) even without fine-tuning, representing a promising direction for safe and reliable applications of SLMs in clinical settings.