Ethical Implications of Small Language Models (SLMs) in Healthcare Applications.

Area of Science:

• Medical Informatics
• Artificial Intelligence Ethics
• Digital Health

Background:

• Small Language Models (SLMs) offer potential for healthcare applications.
• Ethical considerations are paramount for the responsible deployment of SLMs in medicine.

Purpose of the Study:

• To assess the ethical risks associated with Small Language Models (SLMs) in healthcare.
• To evaluate SLM applications in clinical documentation and decision support using the Digital Ethics Canvas.

Main Methods:

• A group activity involving 13 participants during the SLM4Health workshop.
• Utilized the Digital Ethics Canvas framework, examining five ethical lenses: welfare, autonomy, privacy, fairness, and sustainability.
• Assessed two use-cases: SLMs for clinical documentation and SLMs for clinical decision support.

Main Results:

• Identified risks for clinical documentation SLMs include documentation errors (hallucinations, omissions), potential medical mistakes, social/data biases affecting underrepresented groups, and clinician skill erosion.
• Identified risks for clinical decision support SLMs include biases harming vulnerable populations, unequal data quality exacerbating health disparities, and privacy concerns from sensitive data exposure.
• SLMs demonstrated a smaller environmental footprint than large models, but still require energy and resources.

Conclusions:

• SLMs present significant ethical challenges in patient welfare, fairness, autonomy, and privacy within clinical documentation and decision support.
• While SLMs enhance efficiency, proactive development of mitigation strategies is essential to address identified ethical risks.
• Further research is needed to ensure the safe and equitable integration of SLMs into healthcare ecosystems.