Machine Learning Prediction of TBI from Mobility, Gait and Balance Patterns

Abstract

Traumatic Brain Injury (TBI)¹, caused by a violent blow to the head, has become a major cause of death and disability with 56,800 deaths reported in 2014. Early detection of TBI can reduce emergency visits and save lives. Prior methods were clinical tests such as Computerized Tomography (CT) that are expensive and required hospital visits, or mobile health applications that required active involvement that is burdensome. Sensor-rich smartphones have emerged as viable platforms for continuous health monitoring. In this paper, we explore using smartphone sensors for detecting TBI at the onset of injury. Mobility patterns, gait and balance are three human attributes that distinguish subjects with TBI from non-TBI controls. Our work is the first to use machine learning approaches to detect TBI using mobility, gait and balance features extracted from passively gathered smartphone sensor data. Three machine learning pipelines were investigated and compared, namely; (1a) computing hand-crafted features on raw sensor data; (1b) computing hand-crafted features on pre-processed sensor data; (2) using an auto-encoder based approach to encode raw mobility patterns that were then combined with handcrafted gait and balance features. In total, we explored 6 location features, 9 gait and 4 balance statistical features extracted from the smartphone's location and accelerometer sensor data using different segmentation methods. These features were then normalized and classified using machine learning algorithms. Of the three approaches, the best results were achieved using classification of handcrafted gait, balance and mobility features using tree-based classifiers. However, encoding mobility features using an autoencoder that were classified using the Random Forest classifier achieved the best results in terms of early detection. Hand-crafted feature extraction on raw-sensor data gave the best results on the 3rd day after injury using a 24-hour window size with XGBoost achieving Sensitivity of 0.889 and Specificity of 1. For handcrafted features with pre-processing approach, the best results were obtained using 50% overlap with Random Forest having Sensitivity as 0.667 and Specificity of 1. For the autoencoder based approach, Random Forest performed the best on the 2nd day after injury using a 12-hour window-size achieving Sensitivity of 0.778 and Specificity of 0.959.