Impact Identification With Machine Learning From Low Frequency Wearable Sensor Data Among National Ice-hockey Players

作者：Aaron Pilotti-Riley, Davor Stojanov, Muhammad Sohaib Arif, Erik M. Bollt, Stephen J. McGregor · 发表于：Medicine & Science in Sports & Exercise · 年份：2019 · DOI：10.1249/01.mss.0000560594.57047.db · 研究领域：Sports injuries and prevention、Information and Cyber Security、Winter Sports Injuries and Performance

PURPOSE: To determine if machine learning approaches could be used to improve impact identification from low frequency data collected from wearable sensors (WS) among national ice-hockey team members. METHODS: 23 members of the U.S. National (NTDP) U18 team consented to procedures approved by EMU Human Subjects Committee. Using previously validated impact events identified from data collected at 100 Hz (Impact Processor, Zephyr MD), we used two neural network approaches, autoencoder (AE), neural network that can reconstruct inputs of large data sets and multi-layer percptron (MLP), a neural network that uses non-linear activation of multiple layers of interconnecting nodes as well as support vector machine (SVM) to attempt to improve the identification of impacts from summary data recorded at 1 Hz. A dataset, selected from 8 players with the highest ice time in one game, was comprised of 86 impacts and 88602 samples were used to train the AE. Variables from 1 Hz data used to train the AE included Activity and Peak resultant acceleration as well as Peak and Min accelerations in the Vertical, Lateral and Sagittal planes. The trained AE was then applied to the validated test set from 8 players for 3 games consisting of 409 impacts and 462138 datapoints. Accuracy was determined by F1 score (F1 = 2 * (precision * recall) / (precision + recall), where precision = true positives / (true positives + false positives) and recall = true positives / (true positives + false negatives)). R...