A neural network for the detection of soccer headers from wearable sensor data

Abstract

To investigate the proposed association between soccer heading and deleterious brain changes, an accurate quantification of heading exposure is crucial. While wearable sensors constitute a popular means for this task, available systems typically overestimate the number of headers by poorly discriminating true impacts from spurious recordings. This study investigated the utility of a neural network for automatically detecting soccer headers from kinematic time series data obtained by wearable sensors. During 26 matches, 27 female soccer players wore head impacts sensors to register on-field impact events (> 8 g), which were labelled as valid headers (VH) or non-headers (NH) upon video review. Of these ground truth data, subsets of 49% and 21% each were used to train and validate a Long Short-Term Memory (LSTM) neural network in order to classify sensor recordings as either VH or NH based on their characteristic linear acceleration features. When tested on a balanced dataset comprising 271 VHs and NHs (which corresponds to 30% and 1.4% of ground truth VHs and NHs, respectively), the network showed very good overall classification performance by reaching scores of more than 90% across all metrics. When testing was performed on an unbalanced dataset comprising 271 VHs and 5743 NHs (i.e., 30% of ground truth VHs and NHs, respectively), as typically obtained in real-life settings, the model still achieved over 90% sensitivity and specificity, but only 42% precision, which would result in an overestimation of soccer players’ true heading exposure. Although classification performance suffered from the considerable class imbalance between actual headers and non-headers, this study demonstrates the general ability of a data-driven deep learning network to automatically classify soccer headers based on their linear acceleration profiles.