Predicting TBI by using Smartphone-sensed mobility patterns, gait and balance


Downloadable Content

open in viewer

In the United States, Traumatic Brain Injury (TBI) has become a major cause of death and disability. 56,800 deaths were reported due to TBI in the year 2014. Violent blow, sudden jerk to the head or the body are some causes of TBI. Methods to detect TBI at an early stage can help reduce emergency visits and even create life-saving experiences. Imaging tests such as Computerized Tomography (CT), Magnetic resonance imaging (MRI) and Glasgow Coma Scale (GCS) have been widely utilized by doctors and physicians to detect TBI. However, these tests often mis-diagnose the injury and are costly as well. Moreover, these tests require active user involvement and frequent clinic visits. Smartphones are now ubiquitously owned with powerful in-built sensors, making them useful for continuous health monitoring. This thesis focuses of using smartphone sensors for detecting TBI at the onset of injury. A lot of previous work has focused on understanding TBI by extracting the patterns obtained from smartphone sensors. In this thesis, three approaches to understand how the patterns of TBI differ from that of Non-TBI users are compared, namely; i) computing hand-crafted features on raw sensor data; ii) computing hand-crafted features on pre-processed sensor data; iii) using auto-encoder based approach using location, gait and balance. The location patterns extracted have been taken from the work of Mirco Musolesi. 6 location features, 9 gait and 4 balance statistical features were extracted from the location and accelerometer sensor data using different segmentation methods. These features were then normalized and classified using machine learning algorithms. Hand-crafted feature extraction on raw-sensor data gave the best results on 3rd day - 24 hours window size with XGBoost having Sensitivity as 0.889 and Specificity of 1. For the second approach, the best results were obtained using 50% overlap with Random Forest having Sensitivity as 0.667 and Specificity of 1. For the auto-encoder based approach, Random Forest performed the best on 2nd day with 12 hours of window-size having Sensitivity as 0.778 and Specificity of 0.959.

  • etd-23191
Defense date
  • 2021
Date created
  • 2021-05-06
Resource type
Rights statement


In Collection:



Permanent link to this page: