Abstract
OBJECTIVE: To develop and evaluate a computer vision model for automating the identification of trauma resuscitation phases and procedures during trauma video review (TVR). BACKGROUND: TVR is a valuable tool for assessing trauma resuscitation quality and identifying improvement opportunities. However, its labor-intensive nature limits widespread adoption. METHODS: Ninety-five de-identified trauma resuscitation videos from a Level I trauma center were analyzed. Thirty videos (32%) were manually annotated to define 4 trauma phases-pre-arrival, paramedic handover, acute resuscitation, and pre-departure-and procedures, including X-rays, ultrasound, and intravenous access. A multi-institutional research group guided the annotation framework development. Interrater reliability was assessed using temporal intersection over union (tIoU). Model performance metrics included frame-wise accuracy, edit score, F1 scores (tIoU thresholds: 0.1, 0.25, 0.5), average precision (AP), and average recall. RESULTS: The cohort included 65 (68.4%) male patients, median [interquartile range (IQR)] age 31 (23-44.5) years, with 75 (78.9%) blunt injuries and a median (IQR) injury severity score of 22 (12-29). Annotators achieved a high interrater reliability [mean (standard deviation) tIoU: 0.89 (0.19)]. The model achieved a frame-wise accuracy of 98.3%, edit score of 92.1%, and F1 scores of 94.5%, 94.1%, and 86.3% at tIoU thresholds of 0.1, 0.25, and 0.5, respectively. Procedure detection AP exceeded 66% for X-rays and central line placements. CONCLUSIONS: Computer vision can effectively automate TVR, enabling accurate phase segmentation and procedure detection. This approach has the potential to streamline TVR, promote adoption, and improve trauma care quality.