Machine Learning Classification of Head Impact Sensor Data

Abstract

A shortcoming of using environmental sensors for the surveillance of potentially concussive events is substantial uncertainty regarding whether the event was caused by head acceleration (“head impacts”) or sensor motion (with no head acceleration). The goal of the present study is to develop a machine learning model to classify environmental sensor data obtained in the field and evaluate the performance of the model against the performance of the proprietary classification algorithm used by the environmental sensor. Data were collected from Soldiers attending sparring sessions conducted under a U.S. Army Combatives School course. Data from one sparring session were used to train a decision tree classification algorithm to identify good and bad signals. Data from the remaining sparring sessions were kept as an external validation set. The performance of the proprietary algorithm used by the sensor was also compared to the trained algorithm performance. The trained decision tree was able to correctly classify 95% of events for internal cross-validation and 88% of events for the external validation set. Comparatively, the proprietary algorithm was only able to correctly classify 61% of the events. In general, the trained algorithm was better able to predict when a signal was good or bad compared to the proprietary algorithm. The present study shows it is possible to train a decision tree algorithm using environmental sensor data collected in the field.

On-Field Deployment and Validation for Wearable Devices

“…While both guided and blinded video verification are often used to verify which wearable device-recorded events are HAEs, this method is not universally accessible due to both the high time and cost requirements. Recent advances in post-processing techniques, such as machine learning, have allowed for automated classification of HAEs, 21,24,28,46,62,78,81 but care must be taken with implementation of these techniques with acknowledgment of their limitations. Importantly, many commercial advanced post-pro-cessing techniques are proprietary, but their performance can and should be independently validated on the field in the setting of a deployment prior to extensive use.…”

“…The former relies on engineered features that are extracted from kinematic signals in order to differentiate HAEs. 28,62,78,81 Here, the support vector machine is commonly used and is often paired with frequency-based (i.e., peak frequency of acceleration signals) or biomechanics-based (i.e. biomechanical feasibility of events) features.…”

Phybrata Sensors and Machine Learning for Enhanced Neurophysiological Diagnosis and Treatment

“…Applications of machine learning to help automate concussion diagnoses and recovery monitoring have attracted significant research interest in recent years [30]. Data sources include structural and functional MRI [31], electroencephalography (EEG) [32][33][34][35][36], clinical scales [13,30,32,[37][38][39], balance and vestibular diagnostic data [37], gait analysis [30,40], eye tracking [41], blood biomarkers [42], analysis of head impact data [43][44][45], and a variety of wearable physiological sensors [33]. Wearable sensors are of particular interest because of their low cost, ease of use, and compatibility with remote patient monitoring.…”

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

“…Since the kinematic waveforms of true head impacts, e.g., soccer headers, and spurious acceleration events typically look distinct, within recent years, a modest number of studies has focused on the development of data-driven machine learning models to automatically classify sensor-recorded head impact events based on their characteristic acceleration or velocity profiles. In a non-purely sporting context, Rooks et al 25 trained a simple decision tree to distinguish head impacts from other acceleration events during routine sparring sessions of U.S. Army combat soldiers. While their algorithm was able to correctly classify 88% of all sensor-recorded events, only half (51%) of the predicted impacts actually corresponded to actual head impacts (precision) 25 .…”

“…In a non-purely sporting context, Rooks et al 25 trained a simple decision tree to distinguish head impacts from other acceleration events during routine sparring sessions of U.S. Army combat soldiers. While their algorithm was able to correctly classify 88% of all sensor-recorded events, only half (51%) of the predicted impacts actually corresponded to actual head impacts (precision) 25 . In actual sporting contexts, most approaches focused on the discrimination between true head impacts and non-impacts in American 26 , 27 or Australian rules football 28 .…”