Wearable sensors that accurately record head impacts experienced by athletes during play can enable a wide range of potential applications including equipment improvements, player education, and rule changes. One challenge for wearable systems is their ability to discriminate head impacts from recorded spurious signals. This study describes the development and evaluation of a head impact detection system consisting of a mouthguard sensor and machine learning model for distinguishing head impacts from spurious events in football games. Twenty-one collegiate football athletes participating in 11 games during the 2018 and 2019 seasons wore a custom-fit mouthguard instrumented with linear and angular accelerometers to collect kinematic data. Video was reviewed to classify sensor events, collected from instrumented players that sustained head impacts, as head impacts or spurious events. Data from 2018 games were used to train the ML model to classify head impacts using kinematic data features (127 head impacts; 305 non-head impacts). Performance of the mouthguard sensor and ML model were evaluated using an independent test dataset of 3 games from 2019 (58 head impacts; 74 non-head impacts). Based on the test dataset results, the mouthguard sensor alone detected 81.6% of video-confirmed head impacts while the ML classifier provided 98.3% precision and 100% recall, resulting in an overall head impact detection system that achieved 98.3% precision and 81.6% recall.

可穿戴传感器可准确记录运动员在比赛期间所经历的头部撞击，可以实现广泛的潜在应用，包括设备改进、球员教育和规则改变。可穿戴系统面临的一项挑战是它们能够从记录的杂散信号中区分头部撞击。本研究描述了头部撞击检测系统的开发和评估，该系统由护齿传感器和机器学习模型组成，用于区分足球比赛中的头部撞击和虚假事件。在 2018 和 2019 赛季参加了 11 场比赛的 21 名大学足球运动员佩戴了定制的护齿套，该护齿套装有线性和角加速度计，以收集运动学数据。视频经过审查以对传感器事件进行分类，这些事件是从持续头部撞击的仪表玩家那里收集的，作为头部撞击或虚假事件。来自 2018 年游戏的数据用于训练 ML 模型，以使用运动学数据特征对头部撞击进行分类（127 次头部撞击；305 次非头部撞击）。使用 2019 年 3 场比赛的独立测试数据集（58 次头部撞击；74 次非头部撞击）评估了护齿传感器和 ML 模型的性能。根据测试数据集的结果，护齿传感器单独检测到 81.6% 的视频确认头部撞击，而 ML 分类器提供了 98.3% 的准确率和 100% 的召回率，从而形成了一个整体头部撞击检测系统，达到了 98.3% 的准确率和 81.6% 的召回率. 使用 2019 年 3 场比赛的独立测试数据集（58 次头部撞击；74 次非头部撞击）评估了护齿传感器和 ML 模型的性能。根据测试数据集的结果，护齿传感器单独检测到 81.6% 的视频确认的头部撞击，而 ML 分类器提供了 98.3% 的精确度和 100% 的召回率，从而导致整体头部撞击检测系统实现了 98.3% 的精确度和 81.6% 的召回率. 使用 2019 年 3 场比赛的独立测试数据集（58 次头部撞击；74 次非头部撞击）评估了护齿传感器和 ML 模型的性能。根据测试数据集的结果，护齿传感器单独检测到 81.6% 的视频确认头部撞击，而 ML 分类器提供了 98.3% 的准确率和 100% 的召回率，从而形成了一个整体头部撞击检测系统，达到了 98.3% 的准确率和 81.6% 的召回率.