Knowledge transfer for surgical activity prediction.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Surgery
• Neurosurgery

Background:

• Automatic recognition of surgical activities is crucial for situation-aware operating rooms.
• A significant challenge is the lack of annotated training data for machine learning models.
• Knowledge transfer offers a promising solution to overcome data deficits.

Purpose of the Study:

• To enhance the prediction of surgical activities by leveraging knowledge transfer techniques.
• To compensate for limited annotated training data in surgical datasets.
• To improve the development of situation-aware operating room systems.

Main Methods:

• Utilized word embedding to encode semantic information of surgical terms.
• Applied transfer learning to share knowledge between different clinical neurosurgical datasets.
• Combined word embedding and transfer learning for improved predictive performance.

Main Results:

• Achieved a 22% improvement in surgical activity prediction accuracy.
• Demonstrated the effectiveness of knowledge transfer in addressing data scarcity.
• Identified key insights into surgical practices through the analysis of transfer learning outcomes.

Conclusions:

• Word embedding significantly enhances the machine learning process for surgical activity recognition.
• Transfer learning proves more effective than simple data aggregation, particularly for dissimilar procedures.
• The proposed methods contribute to more robust and accurate prediction systems for operating rooms.