Using locally-hosted Small Language Models (SLMs) to protect student, patient and research subject data in Health...

Area of Science:

• Medical Education Technology
• Artificial Intelligence in Healthcare
• Data Privacy in Research

Background:

• Cloud-based Large Language Models (LLMs) are increasingly adopted in Health Professions Education (HPE).
• Significant data privacy concerns arise from potential exposure of sensitive student, patient, and research participant information.
• Existing LLM solutions may not adequately address the stringent data privacy requirements in HPE.

Purpose of the Study:

• To propose and evaluate the use of locally-hosted Small Language Models (SLMs) as a privacy-preserving alternative to cloud-based LLMs in HPE.
• To demonstrate how SLMs can assist teachers and researchers while mitigating data exposure risks.
• To provide technical implementation details for effective SLM deployment in educational and research settings.

Main Methods:

• Implementation of locally-hosted Small Language Models (SLMs) on dedicated infrastructure.
• Task-based assessment of SLM performance in assisting HPE teaching and research activities.
• Qualitative analysis of data privacy implications and ethical considerations.

Main Results:

• Successful implementation of locally-hosted SLMs achieved the primary ethical objective of minimizing data exposure.
• SLMs demonstrated utility in supporting various HPE tasks for educators and researchers.
• Technical guidance for implementation was provided, highlighting practical considerations.

Conclusions:

• Locally-hosted SLMs present a viable and ethical solution for data privacy challenges in HPE.
• Further research is needed to explore SLM practicability across diverse environments and evolving technologies.
• The implications of this approach extend beyond HPE, offering broader potential for secure AI applications.