Survey maps small language models in healthcare

What the study found: The authors present a comprehensive survey of small language models (SLMs) in healthcare and organize them with a taxonomic framework. They describe SLMs as a scalable and clinically viable solution for efficient performance in resource-constrained environments.
Why the authors say this matters: The authors say the survey is meant to help healthcare professionals and informaticians understand recent innovations in model optimization and support future research and development. They also frame SLMs as relevant because of data privacy concerns and limited resources in healthcare.
What the researchers tested: The article is a comprehensive survey rather than an experimental study of one model. It includes a timeline of healthcare SLM contributions and a framework organized across three dimensions: NLP tasks, stakeholder roles, and the continuum of care.
What worked and what didn't: The survey covers architectural foundations for building models from scratch, adapting SLMs through prompting, instruction fine-tuning, and reasoning, and improving accessibility and sustainability through compression techniques. It also compiles experimental results across widely studied NLP tasks in healthcare, but the abstract does not state which approaches performed best or poorly.
What to keep in mind: The abstract does not describe specific limitations of the survey or the included studies. It also does not provide detailed quantitative findings in the summary text available here.

Key points

• The article is a comprehensive survey of small language models in healthcare.
• The authors present a taxonomic framework for classifying SLMs across NLP tasks, stakeholder roles, and the continuum of care.
• The abstract says SLMs may be a scalable and clinically viable solution in resource-constrained environments.
• The survey discusses model building from scratch, prompting, instruction fine-tuning, reasoning, and compression techniques.
• The abstract does not identify specific best-performing or weakest approaches.