Research and Application of Denoising Head Impact Kinematics Measurement Based on Convolutional Neural Networks

Research Background

Mild Traumatic Brain Injury (MTBI) is a global health threat. Humans often face the risk of MTBI in situations such as falls, traffic accidents, and sports. According to statistics, there were over 27 million brain injury cases worldwide in 2016, of which 80% were classified as “mild” brain injuries. Although the symptoms are relatively mild, they can lead to long-term pathological changes. MTBI is typically classified using the Glasgow Coma Scale score, with patients scoring above 12 being classified as mild brain injury patients. While symptoms may quickly recover after the acute phase, patients may develop complications such as Chronic Traumatic Encephalopathy (CTE) in the long run.

To quantify the impact of head injuries on the brain, researchers have developed various wearable sensor technologies to measure head kinematics parameters. These systems include the Head Impact Telemetry System (HITS), XPatch, headband/headscarf sensors, and sensors mounted on mouthguards. However, these sensors inherently have noise issues due to the imperfect interface with the human body, thus requiring higher precision measurement methods.

Source of the Paper

The paper was written by researchers Xianghao Zhan, Yuzhe Liu, Nicholas J. Cecchi, et al., from Stanford University’s Department of Bioengineering and Beihang University’s School of Biological Science and Medical Engineering. This paper was published in the IEEE Transactions on Biomedical Engineering journal in 2024. The aim is to provide a one-dimensional convolutional neural network (1D-CNN) model based on deep learning to denoise the head impact kinematics data measured by mouthguards, thereby improving the accuracy of Traumatic Brain Injury (TBI) risk monitoring.

Research Workflow

The paper describes an original research that mainly includes the following stages:

Experimental Setup and Data Collection

Researchers used a jaw-loaded headform with a Hybrid III neck in a laboratory environment to simulate head impacts. Impacts were applied by an air pressure-accelerated impactor and reference head kinematics data were read using high-precision accelerometers and gyroscopes installed at the headform’s center of gravity (COG). The experiments collected 163 independent impact tests using standard protocol and non-repetitive protocol. The mouthguard data were filtered and time-aligned with reference data.

Development of 1D Convolutional Neural Network Denoising Model

The research team designed a 1D-CNN model containing six convolutional layers to extract temporal features of kinematic signals and compare them with reference signals. Specifically, the dataset was divided into a training set (113 impacts, 70%), validation set (25 impacts, 15%), and test set (25 impacts, 15%), augmented via sliding windows.

Calculation of Brain Injury Criteria

To evaluate the denoising effect, researchers calculated six different Brain Injury Criteria (BIC), including Head Injury Criterion (HIC), Head Impact Power (HIP), the Generalized Acceleration Model for Brain Injury Threshold (GAMBIT), Severity Index (SI), Brain Injury Criterion (BRIC), and Combined Probability of Concussion (CP).

Brain Strain and Strain Rate Calculation Based on Finite Element Model

A verified KTH finite element model was adopted to compute brain tissue-level strain (MPS) and strain rate (MPSR), further applying logistic regression injury risk functions to predict concussion risk.

Main Research Results

Model Performance at the Kinematic Level

Results showed higher correlation between the denoised mouthguard measurement data and reference data using the 1D-CNN model, with absolute peak error and Root Mean Squared Error (RMSE) reduced by 36% and 56% respectively. RMSE for z-axis linear acceleration reduced by 56%, and peak absolute error for x-axis angular velocity reduced by 86%.

Model Performance at the Brain Injury Criteria Level

Among the six brain injury criteria, error reductions in four criteria were significant post-denoising, with an average error reduction of 82%. These brain injury criteria are effective methods for quickly quantifying brain injury risk after head impacts.

Model Performance at the Tissue Level Strain and Strain Rate

Regarding brain strain and strain rate calculation, the 1D-CNN model similarly demonstrated efficacy. Error in MPS and MPSR estimation was significantly reduced (p<0.001) post-denoising, indicating better accuracy in detecting high strain and high strain rate regions.

Blind Testing with Field Soccer Impacts

Blind testing on 118 field soccer impact data showed that denoising not only reduced peak kinematics data but also significantly decreased the whole-brain MPS and MPSR.

Model Performance with Human Impact Data

Further testing on 413 human impact cases showed that the denoising model significantly improved measurement accuracy for the noisiest 10% and 5% of impact data, reducing peak angular velocity error by up to 75.6% and 82.3%.

Research Significance and Value

This paper introduces a method of using deep learning based one-dimensional convolutional neural network for denoising head kinematics data, aiding in increasing the accuracy of mouthguard measurement data. This can provide more precise head kinematics data for further TBI diagnosis and research. The proposed method has substantial application value in addressing the noise problem in head impact data, especially in complex scenarios like field sports, warranting further validation and promotion.

The study showcased good denoising performance in test data and provided code for model development and data processing for other researchers. However, in practical applications, it requires further validation using real human subjects data to ensure model accuracy and reliability across various real-world scenarios.