AI Scribe Safety: Measuring What Happens After Signing

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Documentation

Background:

• Current evaluations of AI scribes primarily focus on the initial draft quality.
• Signed clinical notes are dynamic and undergo further processing by downstream AI tools.
• Existing metrics may not fully capture the real-world impact of AI-generated documentation.

Purpose of the Study:

• To question the adequacy of current AI scribe evaluation metrics.
• To advocate for downstream safety measurements in AI scribe assessments.
• To highlight the potential for compounding errors in electronic health records.

Main Methods:

• Conceptual analysis of AI scribe evaluation frameworks.
• Discussion of the dynamic nature of clinical notes.
• Argumentation for a shift in safety assessment paradigms.

Main Results:

• Initial draft quality is an insufficient metric for AI scribe safety.
• Downstream applications of clinical notes can amplify initial errors.
• A new framework for evaluating AI scribe safety is proposed.

Conclusions:

• AI scribe evaluations must extend beyond initial draft quality.
• Safety must be assessed based on how errors propagate through the electronic health record.
• A downstream-focused approach is crucial for ensuring patient safety with AI scribes.