Enhanced Trauma Video Review With Computer Vision

This methodological study evaluated whether computer vision can automate key components of trauma video review (TVR), a powerful but labor-intensive quality improvement tool. Investigators analyzed 95 de-identified trauma resuscitation videos from a Level I trauma center and, with guidance from a multi-institutional TVR research group, created a standardized framework for four visually discernible phases of resuscitation: pre-arrival, paramedic handover, acute resuscitation, and pre-departure. They also defined an annotation scheme for common bedside procedures including X-rays, FAST examinations, central line placement, and IV access. An expert annotation team labeled a subset of 30 videos, achieving high interrater reliability using temporal intersection over union (tIoU), which served as the ground truth for model training and evaluation.

The authors then built a two-stage computer vision pipeline using multiview trauma bay video. Spatiotemporal features were first extracted with an Inflated 3D Convolutional Network (I3D) to capture both visual content and motion across up to four synchronized camera angles. These features were passed to two downstream models: a Multi-Stage Temporal Convolutional Network (MS-TCN) for continuous trauma phase segmentation, and ActionFormer, a transformer-based temporal action localization model, for detection and temporal localization of discrete procedures. Model performance was assessed with frame-wise accuracy, edit score, F1 scores at multiple tIoU thresholds for phase segmentation, and average precision/recall for procedure detection.