Audio-based event detection in the operating room

Area of Science:

• Medical Informatics
• Signal Processing
• Machine Learning

Background:

• Current operating room (OR) workflow analysis systems are primarily research-based and lack universal application.
• The potential of audio data for OR workflow analysis remains largely unexplored, despite advantages like low cost and independence from visual or location data.

Purpose of the Study:

• To develop and evaluate an audio-based event detection system for operating rooms.
• To explore the feasibility of using sound analysis for comprehensive workflow recognition in surgical environments.

Main Methods:

• Utilized a novel dataset of over 63 hours of annotated audio from an operating room.
• Processed audio data into log-mel-spectrograms for input into pretrained convolutional neural networks.
• Employed data augmentation techniques to improve classification accuracy.

Main Results:

• Achieved high performance with lightweight models like MobileNet, demonstrating computational efficiency.
• Successfully classified 11 distinct audio event classes, including surgical procedures and equipment movements.
• Attained an overall accuracy of 90%, precision of 91%, and F1-score of 91% on the custom dataset.

Conclusions:

• Audio event recognition is a viable and valuable addition to OR workflow analysis, extending beyond speech.
• The proposed method is computationally inexpensive and can be integrated into existing or future recognition pipelines.
• This proof of concept highlights audio as a significant, underutilized data source for surgical environments.