Suki Researchers Challenge Traditional AI Note Evaluation Methods in Healthcare

As the adoption of AI ambient scribe technology continues to grow in healthcare, researchers at Suki are raising concerns about the industry's standard method for evaluating clinical AI notes. According to their findings, the Physician Documentation Quality Instrument (PDQI-9), which has been the go-to tool since 2012, may be fundamentally flawed when it comes to assessing modern ambient AI scribes.

Suki, a leading provider of ambient clinical AI solutions, works with over 400 health systems. The company's researchers argue that the PDQI-9 is poorly suited for evaluating AI-generated notes because it focuses on overall note quality-such as organization and conciseness-rather than identifying specific errors unique to large language models (LLMs). These errors can include hallucinations, invented medication doses, and missing diagnoses.

The Limitations of Traditional Evaluation Rubrics

The PDQI-9 evaluation domains are up-to-date, accurate, thorough, useful, organized, concise, synthesized, internally consistent, and cohesive. However, these broad criteria may not be sufficient for detecting the nuanced errors that LLMs can introduce into clinical notes.

Kevin Wang, M.D., Suki's chief medical officer, emphasizes the need for a refresh in evaluation methods: "Base quality rubrics need a refresh in the post-LLM, ambient world." Suki's white paper analyzes the limitations of holistic Likert-based scoring instruments like PDQI-9 and highlights their inability to reliably evaluate AI-generated clinical notes.

In their study, Suki researchers examined 84 paired notes across four specialties. They found significant inconsistencies among reviewers, including:

• Disagreement over what constitutes a hallucination
• Unreliable assessments of accuracy
• Varying scores depending on which AI model generated the note

These issues, according to the researchers, reflect a fundamental mismatch between traditional note-quality rubrics and the types of errors LLMs can make. PDQI-9 scores notes holistically, focusing on factors like organization, conciseness, and cohesiveness-factors that are less relevant when it comes to detecting specific AI-generated errors.

Key Takeaways

• PDQI-9 Limitations: The traditional PDQI-9 rubric is ill-suited for evaluating modern ambient AI scribes, as it fails to detect LLM-specific errors.
• Inconsistencies in Reviewing: Suki's study reveals significant inconsistencies among reviewers when using the PDQI-9, highlighting the need for a more robust evaluation method.
• Need for Refreshed Rubrics: Healthcare providers and researchers should consider developing new evaluation tools that can better identify and address the unique challenges posed by AI-generated clinical notes.

Suki's research underscores the importance of updating evaluation methods to ensure that AI ambient scribes are accurately and reliably assessed. As AI continues to play a larger role in healthcare, it is crucial to have robust tools that can detect and mitigate potential errors, ultimately improving patient care and safety.