776 BO01 – Can a novel computer vision-based framework detect head-on-head impacts during a rugby league tackle?

Background Given that head-on-head impacts have the highest risk for Head Injury Assessments (HIAs), the capability of computer vision to automatically identify when a head-on-head impact occurs between two players would be beneficial. This poses a practical solution for rugby league and other sports which could allow stakeholders (e.g., match officials, doctors) to evaluate the frequency, timestamp, and incidence of head-on-head events that occur during a match using only video-based data. Objective Develop and evaluate a novel computer vision framework to automatically classify head-on-head impacts using video footage from elite-level rugby league matches. Design Validity study. Setting European Super League Rugby and Australian National Rugby League. Participants 625 non-televised standard-definition (SDTV) and 396 televised high-definition (HDTV) video clips. The inclusion criteria for a head-on-head impact were defined as direct contact by one player's head on another player's head (opponent or teammate), resulting in a visible change in the trajectory of the head. Assessment of Risk Factors Head-on-Head Impacts manually coded by experieenced video analyst was the criterion to evaluate the validity of the automated framework. Main Outcome Measurements The main outcome measurements were the Sensitivity, Specificity, and Positive Predictive Value (PPV) of the computer vision framework. Results The overall performance evaluation of the framework to classify head-on-head impacts against manual coding had a sensitivity, specificity, and PPV (95% confidence intervals) of 68(66–71)%, 84(80–86)%, and 0.61(0.59–0.63) in SDTV, and 65(63–67)%, 84(81–88)%, and 0.61(0.59–0.64) in HDTV. Conclusions The study introduces a novel computer vision framework for head-on-head impact detection which has the potential to detect head-on-head impacts in real-time and could streamline manual video verification of head-on-head impact monitoring and/or wearable sensor impacts, given the similar performance to other wearable sensors. Future work should explore how such a framework can be expanded into head-on-other (e.g., shoulder, hip, elbow, knee, ground) impacts.