Enhanced Trauma Video Review With Computer Vision: Trauma Resuscitation Phase Segmentation and Procedure Detection.

Area of Science:

• Medical imaging and computer vision applications in healthcare.
• Development of artificial intelligence for clinical workflow automation.

Background:

• Trauma video review (TVR) is crucial for assessing resuscitation quality but is time-consuming.
• Limited adoption of TVR hinders quality improvement initiatives in trauma care.

Purpose of the Study:

• To create and validate a computer vision model for automating trauma resuscitation phase and procedure identification.
• To overcome the labor-intensive limitations of manual trauma video review.

Main Methods:

• Analysis of 95 trauma resuscitation videos with manual annotation of 4 phases and key procedures.
• Development of an annotation framework guided by a multi-institutional research group.
• Evaluation of model performance using metrics like frame-wise accuracy, edit score, and F1 scores at various tIoU thresholds.

Main Results:

• High interrater reliability (mean tIoU: 0.89) achieved during manual annotation.
• The computer vision model demonstrated high accuracy: 98.3% frame-wise accuracy and 92.1% edit score.
• Excellent F1 scores (up to 94.5%) and over 66% average precision for detecting procedures like X-rays and central line placements.

Conclusions:

• Computer vision effectively automates trauma video review, accurately segmenting phases and detecting procedures.
• Automated TVR can streamline the review process, encourage wider adoption, and ultimately improve trauma care.
• This technology offers a scalable solution for enhancing trauma resuscitation quality assessment.