Standardization of surgical gesture taxonomy: a SAGES Delphi consensus study

Area of Science:

• Surgical workflow analysis
• Medical AI
• Computer vision in surgery

Background:

• Current AI for surgical analysis lacks generalization due to non-standardized action representations.
• Gesture-level tokenization offers specificity but is hindered by inconsistent terminology, limiting dataset use.
• A standardized language is crucial for advancing AI in surgical data science.

Purpose of the Study:

• To establish a standardized, hierarchical taxonomy for surgical gestures.
• To overcome fragmentation in surgical gesture terminology.
• To enable improved AI model development and cross-study comparisons.

Main Methods:

• A SAGES-led Delphi consensus process was used, starting with 270 literature terms.
• A hybrid pipeline combined large language model (LLM)-assisted clustering with expert review.
• Methods included surveys, video-based validation, and an in-person consensus meeting.

Main Results:

• A hierarchical taxonomy with 10 clusters, 24 gestures, and 46 sub-gestures was established.
• Multi-instrument annotation was supported over dominant-instrument-only labeling.
• Video validation showed high agreement for many gestures but highlighted ambiguities for refinement.

Conclusions:

• A standardized surgical gesture taxonomy provides a foundational language for surgical data science.
• This framework aims to reduce annotation variability and improve AI development.
• Defining temporal boundaries for gestures is the next essential step.