Automatic detection of procedural knowledge in robotic-assisted surgical texts

Area of Science:

• Natural Language Processing
• Machine Learning
• Surgical Informatics

Background:

• Automating knowledge extraction from surgical manuals is crucial for developing expert surgical systems.
• Identifying procedural information is a key subtask for extracting surgical workflows.

Purpose of the Study:

• To assess the feasibility of automatically identifying sentences with procedural information in surgical texts.
• To develop and evaluate machine learning models for this classification task.

Main Methods:

• Framed the problem as a binary classification task.
• Created and utilized a new public dataset of 1958 manually annotated sentences from robotic surgery texts.
• Applied classical machine learning, neural networks (FastText embeddings), and transformer-based models (BERT, ClinicalBERT).
• Analyzed the impact of dataset balancing techniques.

Main Results:

• Neural network architectures using FastText embeddings and ClinicalBERT demonstrated superior performance.
• These models empirically confirmed the feasibility of automatically identifying procedural sentences.
• Dataset balancing techniques did not yield substantial improvements in classification accuracy.

Conclusions:

• This work is the first to explore machine/deep learning for identifying procedural sentences in surgical texts.
• Introduced the first public dataset for benchmarking classification methods in this domain.
• Highlights the potential of AI in advancing surgical knowledge extraction and expert system development.