Automated soccer head impact exposure tracking using video and deep learning

Immediate and Transient Perturbances in EEG Within Seconds Following Controlled Soccer Head Impact

Athletes in contact and collision sports can sustain frequent subconcussive head impacts. Although most impacts exhibit low kinematics around or below 10 g of head linear acceleration, there is growing concern regarding the cumulative effects of repetitive sports head impacts. Even mild impacts can lead to brain deformations as shown through neuroimaging and finite element modeling, and thus may result in mild and transient effects on the brain, prompting further investigations of the biomechanical dose-brain response relationship. Here we report findings from a novel laboratory study with continuous monitoring of brain activity through electroencephalography (EEG) during controlled soccer head impacts. Eight healthy participants performed simulated soccer headers at 2 mild levels (6 g, 4 rad/s and 10 g, 8 rad/s) and three directions (frontal, oblique left, oblique right). Participants were instrumented with an inertial measurement unit (IMU) bite bar and EEG electrodes for synchronized head kinematics and brain activity measurements throughout the experiment. After an impact, EEG exhibited statistically significant elevation of relative and absolute delta power that recovered within two seconds from the impact moment. These changes were statistically significantly higher for 10 g impacts compared with 6 g impacts in some topographical regions, and oblique impacts resulted in contralateral delta power increases. Post-session resting state measurements did not indicate any cumulative effects. Our findings suggest that even mild soccer head impacts could lead to immediate, transient neurophysiological changes. This study paves the way for further dose-response studies to investigate the cumulative effects of mild sports head impacts, with implications for long-term athlete brain health.

On-field Head Acceleration Exposure Measurements Using Instrumented Mouthguards: Multi-stage Screening to Optimize Data Quality

Instrumented mouthguards (iMGs) are widely applied to measure head acceleration event (HAE) exposure in sports. Despite laboratory validation, on-field factors including potential sensor skull-decoupling and spurious recordings limit data accuracy. Video analysis can provide complementary information to verify sensor data but lacks quantitative kinematics reference information and suffers from subjectivity. The purpose of this study was to develop a rigorous multi-stage screening procedure, combining iMG and video as independent measurements, aimed at improving the quality of on-field HAE exposure measurements. We deployed iMGs and gathered video recordings in a complete university men's ice hockey varsity season. We developed a four-stage process that involves independent video and sensor data collection (Stage I), general screening (Stage II), cross verification (Stage III), and coupling verification (Stage IV). Stage I yielded 24,596 iMG acceleration events (AEs) and 17,098 potential video HAEs from all games. Approximately 2.5% of iMG AEs were categorized as cross-verified and coupled iMG HAEs after Stage IV, and less than 1/5 of confirmed or probable video HAEs were cross-verified with iMG data during stage III. From Stage I to IV, we observed lower peak kinematics (median peak linear acceleration from 36.0 to 10.9 g; median peak angular acceleration from 3922 to 942 rad/s2) and reduced high-frequency signals, indicative of potential reduction in kinematic noise. Our study proposes a rigorous process for on-field data screening and provides quantitative evidence of data quality improvements using this process. Ensuring data quality is critical in further investigation of potential brain injury risk using HAE exposure data.

Artificial intelligence applications in the football codes: A systematic review

Artificial Intelligence (AI) is increasingly being adopted across many domains such as transport, healthcare, defence and sport, with football codes no exception. Though there is a range of potential benefits of AI, concern has also been expressed regarding potential risks. An important first step in ensuring that AI applications in football are usable, beneficial, safe and ethical is to understand the current range of applications, the AI models adopted and their proposed functions. This systematic review aimed to identify different applications of AI across football codes to synthesise current knowledge and determine whether potential risks are being considered. The systematic review included 190 peer-reviewed articles. Nine areas of application were found ranging from athlete evaluation and event detection to match outcome prediction and injury detection and prediction. In total, 27 different AI models were identified, with artificial neural networks the most frequently applied. Five AI assessment metrics were identified including specificity, recall, precision, accuracy and F1-score. Four potential risks were identified, concerning data security, usability, data biases and inappropriate athlete load management. It is concluded that, though a wide range of AI applications currently exist, further work is required to develop AI for football and identify and manage potential risks.

Individualized monitoring of longitudinal heading exposure in soccer

There is growing concern that repetitive soccer headers may have negative long-term consequences on brain health. However, inconsistent and low-quality heading exposure measurements limit past investigations of this effect. Here we conducted a comprehensive heading exposure analysis across all players on a university women's soccer team for over two calendar years (36 unique athletes), quantifying both game and practice exposure during all in-season and off-season periods, with over ten thousand video-confirmed headers. Despite an average of approximately 2 headers per day, players' daily exposures ranged from 0 to 45 headers, accumulating to highly variable total exposure of 2-223 headers over each in-season/off-season period. Overall, practices and off-season periods accounted for 70% and 45% of headers, respectively. Impact sensor data showed that heading kinematics fell within a tight distribution, but sensors could not capture full heading exposure due to factors such as compliance. With first-of-its-kind complete heading exposure data, we recommend exposure assessments be made on an individual level and include practice/off-season collection in addition to games and competitive seasons. Commonly used group statistics do not capture highly variable exposures, and individualized complete heading exposure tracking needs to be incorporated in future study designs for confirming the potential brain injury risk associated with soccer heading.

Identifying relevant diffusion MRI microstructure biomarkers relating to exposure to repeated head impacts in contact sport athletes

PURPOSE

Repeated head impacts (RHI) without concussion may cause long-term sequelae. A growing array of diffusion MRI metrics exist, both empiric and modeled and it is hard to know which are potentially important biomarkers. Common conventional statistical methods fail to consider interactions between metrics and rely on group-level comparisons. This study uses a classification pipeline as a means towards identifying important diffusion metrics associated with subconcussive RHI.

METHODS

36 collegiate contact sport athletes and 45 non-contact sport controls from FITBIR CARE were included. Regional/whole brain WM statistics were computed from 7 diffusion metrics. Wrapper-based feature selection was applied to 5 classifiers representing a range of learning capacities. Best 2 classifiers were interpreted to identify the most RHI-related diffusion metrics.

RESULTS

Mean diffusivity (MD) and mean kurtosis (MK) are found to be the most important metrics for discriminating between athletes with and without RHI exposure history. Regional features outperformed global statistics. Linear approaches outperformed non-linear approaches with good generalizability (test AUC 0.80-0.81).

CONCLUSION

Feature selection and classification identifies diffusion metrics that characterize subconcussive RHI. Linear classifiers yield the best performance and mean diffusion, tissue microstructure complexity, and radial extra-axonal compartment diffusion (MD, MK, De,⊥) are found to be the most influential metrics. This work provides proof of concept that applying such approach to small, multidimensional dataset can be successful given attention to optimizing learning capacity without overfitting and serves an example of methods that lead to better understanding of the myriad of diffusion metrics as they relate to injury and disease.

Automated player identification and indexing using two-stage deep learning network

American football games attract significant worldwide attention every year. Identifying players from videos in each play is also essential for the indexing of player participation. Processing football game video presents great challenges such as crowded settings, distorted objects, and imbalanced data for identifying players, especially jersey numbers. In this work, we propose a deep learning-based player tracking system to automatically track players and index their participation per play in American football games. It is a two-stage network design to highlight areas of interest and identify jersey number information with high accuracy. First, we utilize an object detection network, a detection transformer, to tackle the player detection problem in a crowded context. Second, we identify players using jersey number recognition with a secondary convolutional neural network, then synchronize it with a game clock subsystem. Finally, the system outputs a complete log in a database for play indexing. We demonstrate the effectiveness and reliability of player tracking system by analyzing the qualitative and quantitative results on football videos. The proposed system shows great potential for implementation in and analysis of football broadcast video.

A sensor-enabled cloud-based computing platform for computational brain biomechanics

BACKGROUND AND OBJECTIVES

Driven by the risk of repetitive head trauma, sensors have been integrated into mouthguards to measure head impacts in contact sports and military activities. These wearable devices, referred to as "instrumented" or "smart" mouthguards are being actively developed by various research groups and organizations. These instrumented mouthguards provide an opportunity to further study and understand the brain biomechanics due to impact. In this study, we present a brain modeling service that can use information from these sensors to predict brain injury metrics in an automated fashion.

METHODS

We have built a brain modeling platform using several of Amazon's Web Services (AWS) to enable cloud computing and scalability. We use a custom-built cloud-based finite element modeling code to compute the physics-based nonlinear response of the intracranial brain tissue and provide a frontend web application and an application programming interface for groups working on head impact sensor technology to include simulated injury predictions into their research pipeline.

RESULTS

The platform results have been validated against experimental data available in literature for brain-skull relative displacements, brain strains and intracranial pressure. The parallel processing capability of the platform has also been tested and verified. We also studied the accuracy of the custom head surfaces generated by Avatar 3D.

CONCLUSION

We present a validated cloud-based computational brain modeling platform that uses sensor data as input for numerical brain models and outputs a quantitative description of brain tissue strains and injury metrics. The platform is expected to generate transparent, reproducible, and traceable brain computing results.

A fully automatic method for segmentation of soccer playing fields

This paper proposes a strategy to segment the playing field in soccer images, suitable for integration in many soccer image analysis applications. The combination of a green chromaticity-based analysis and an analysis of the chromatic distortion using full-color information, both at the pixel-level, allows segmenting the green areas of the images. Then, a fully automatic post-processing block at the region-level discards the green areas that do not belong to the playing field. The strategy has been evaluated with hundreds of annotated images from matches in several stadiums with different grass shades and light conditions. The results obtained have been of great quality in all the images, even in those with the most complex lighting conditions (e.g., high contrast between sunlit and shadowed areas). In addition, these results have improved those obtained with leading state-of-the-art playing field segmentation strategies.

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

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.