Machine understanding surgical actions from intervention procedure textbooks.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Automatic extraction of surgical knowledge is vital for developing clinical decision support systems and educational tools.
• Current methods require structured data, limiting the use of vast textual resources like surgery manuals and academic papers.

Purpose of the Study:

• To establish a benchmark for extracting detailed surgical actions from textual resources using Semantic Role Labeling.
• To evaluate the effectiveness of Transformer-based models, including a novel surgical language model (SurgicBERTa), for surgical knowledge extraction.

Main Methods:

• Framed surgical action extraction as a Semantic Role Labeling task.
• Utilized RoBERTa, BioMedRoBERTa, and a new domain-specific model, SurgicBERTa, comparing zero-shot, fine-tuning, and few-shot learning scenarios.
• Evaluated performance on in-domain and out-of-domain datasets across predicate and argument disambiguation sub-tasks.

Main Results:

• Fine-tuning a pre-trained, domain-specific language model (SurgicBERTa) achieved the highest performance across all evaluated splits and sub-tasks.
• The proposed approach demonstrated effectiveness in both in-domain and out-of-domain settings, as well as in few-shot learning scenarios.
• SurgicBERTa outperformed general and biomedical pre-trained models for surgical text analysis.

Conclusions:

• Fine-tuning domain-specific language models is the most effective approach for extracting procedural surgical knowledge from text.
• The developed benchmark and models, including SurgicBERTa, offer a significant advancement for AI-driven surgical knowledge systems.
• All models and datasets are publicly available to facilitate further research and development.