Video-based Concurrent Activity Recognition for Trauma Resuscitation.

Area of Science:

• Computer Vision
• Medical Informatics
• Artificial Intelligence

Background:

• Teamwork in clinical settings, particularly during trauma resuscitation, involves complex concurrent activities.
• Accurate recognition of these activities is crucial for performance analysis and training.
• Existing methods often struggle with the dynamic and privacy-sensitive nature of clinical environments.

Purpose of the Study:

• To develop and evaluate a video-based system for concurrent activity recognition in clinical teamwork.
• To ensure patient and provider privacy during system development and operation.
• To benchmark the system's performance against established methods in a critical care scenario.

Main Methods:

• Developed a video-based system leveraging spatio-temporal features to preserve privacy.
• Extended the inflated 3D ConvNet (i3D) model for activity recognition.
• Tuned i3D model weights using back-propagated loss and applied prediction filtering.

Main Results:

• The system achieved 74% average precision (AP) for recognizing five key trauma resuscitation activities.
• The developed system demonstrated superior performance compared to previous domain-specific systems.
• Feature map visualization confirmed the model's focus on activity-relevant visual cues.

Conclusions:

• The proposed video-based system effectively recognizes concurrent activities in clinical teamwork while maintaining privacy.
• The extended i3D model shows promise for real-world healthcare applications.
• This technology can aid in analyzing and improving clinical team performance during critical events.