Concussions are mild traumatic brain injuries and a common occurrence in contact sports. It is now a major public health problem in the USA. The figure below demonstrated that >200 concussion cases occur per season, which is equivalent to about one case per game.
A topic modeling using natural language processing has demonstrated that sports concussion currently is and will remain a hot research topic in the next 5 years.
With the advances of computer science, machine learning technique has been adopted in the research field of sports concussion. In this article, I will introduce some representative machine learning applications in this research field.
Concussion classification – Logistic Regression
My first study using machine learning was published in the BMMB Journal (Zhao et al., 2017). The study combines neuroimage processing, computational modeling, and machine learning technique to classify concussions in reconstructed NFL head impacts. In this study, the Worcester Head Injury Model (WHIM) was used to simulate 58 reconstructed NFL head impacts (Newman et al., 2000), which include 25 concussion cases and 33 non-injuries. The white matter tractography of the individual whose brain geometry was used to create the WHIM and the ICBM-DTI-81 deep white matter (WM) atlas were co-registered with the WHIM so that resulting strain responses can be sampled either on the 50 deep WM regions or on the tractography traversing each region. Nine injury metrics were defined based on either head kinematics, regional brain strains or strain along WM tractography.
For each injury metric, a logistic regression was performed at each of the 50 deep WM regions along with the injury labels. We compared the AUC, accuracy, sensitivity, and specificity across all the nine injury metrics and deep WM regions. For example, two of the tractography-based metrics were more injury-predictive than all other metrics, in terms of average sensitivity across all the deep WM regions, as shown below.
Concussion classification — Deep Learning
Instead of using a single scalar injury metric to classify injury that may lead to a loss of features, my colleagues and I adopted a feature-based deep learning technique to perform a concussion classification. As illustrated by the figure below, WM fiber strain sampled on all the WM voxels were used as the input of the deep neural network.
The classification performance of the deep learning technique outperformed other machine learning techniques, such as SVM, random forest, and logistic regression based on a single feature, as shown below.
Regional Brain Strain Estimation — CNN
In the context sports concussion, we also applied deep learning technique in a study of brain strain estimation (Wu et al., 2019). In this study, we augmented the head impact kinematics measured from reconstructed NFL head impacts (Newman et al., 2000) and collegiate football head impacts (Hernandez et al, 2015) by randomly scaling the magnitude and rotating the axis of rotational velocity.
A convolutional neural network was established using 2D images converted from augmented rotational velocity as the input and regional peak maximum principal or fiber strain as the output.
The CNN architecture predicted vary similar regional brain strains to the directly simulated results for both in-range and out-of-range predictions, as shown below. This study established a CNN architecture that allows us to instantaneously estimate regional brain strains.
Whole-Brain Strain Estimation — CNN
While regional brain strains can be instantaneously estimated in Wu et al. (2019), the strains are single values representing the deformation of brain regions of interest and do not inform strain distribution throughout the whole brain. Following that study, we used both rotational velocity and acceleration which were converted to 2D image as the input of a pre-trained CNN architecture to estimate the whole-brain strains. By optimizing the pre-trained CNN architecture, we were able to estimate the brain strain distribution. This study was recently published in Journal of Neurotrauma (Ghazi et al., 2020). The CNN architecture is shown below.
Finally, the optimized CNN architecture predicted very similar whole-brain strains to the directly simulated counterpart, as shown below.
Following the idea of our study, another research group replicated our methodology to estimate the whole-brain strains using the KTH head model (Zhan et al, (2020)). The difference of this study from ours is that a different dataset where the head impact kinematics measured from collegiate football head impacts was employed.
Computational head models serve as a bridge to connect head impact kinematics and the resulting brain mechanical responses. However, one shortcoming of computational models is the heavy computational cost. With the machine learning technique, this shortcoming can be addressed as introduced in this article. The brain mechanical responses can be instantaneously estimated without significant loss of accuracy.
This article was written at Metis Data Science Bootcamp as an investigation program.