Focusing on What Matters: Fine-grained Medical Activity Recognition for Trauma Resuscitation via Actor Tracking.

Area of Science:

• Medical technology
• Computer vision
• Artificial intelligence in healthcare

Background:

• Trauma is a major global cause of mortality, with a significant percentage of deaths being preventable.
• Errors during initial trauma resuscitation contribute to preventable deaths.
• Current decision support systems in trauma care are limited by manual data entry in time-critical settings.

Purpose of the Study:

• To address challenges in computer vision-based decision support for trauma resuscitation.
• To develop an AI system for accurate fine-grained activity recognition in complex medical scenarios.
• To improve the reliability of decision support during critical care.

Main Methods:

• An actor tracker was developed to focus on individual features in crowded scenes.
• Video Masked Autoencoder (Video-MAE) was employed for self-supervised learning with unlabeled data to enhance feature representation.
• An ensemble fusion method was utilized, combining predictions from video clips and actors for improved accuracy.

Main Results:

• The proposed system effectively identified fine-grained activities in trauma resuscitation settings.
• The computer vision approach addressed challenges like complex backgrounds and crowded scenes.
• Self-supervised learning with Video-MAE improved feature representation despite limited labeled data.

Conclusions:

• The developed system offers a viable solution for activity recognition in trauma resuscitation.
• This AI-driven approach has the potential to reduce errors and improve patient outcomes.
• The methods are applicable to other complex domains requiring detailed activity recognition.