Skip to main content

WPI - Computer Science Department , MS Thesis Presentation - Florina Asani "Traumatic Brain Injury (TBI) detection and Bioscore generation for ailment monitoring using smartphone sensor data"

Wednesday, May 05, 2021
2:00 pm to 3:00 pm

Florina Asani
MS Student
WPI - Computer Science Department

Date: Wednesday, May 5th, 2021
Time: 2:00 pm - 3:00 pm

Advisor :  Prof. Emmanuel O.Agu

Reader : Prof.  Elke Rundensteiner

TBI causes millions of individuals distress and it can lead to significant motor, cognitive and emotional deficit. However, TBI patients are currently assessed infrequently especially between scheduled appointments. Emerging research in the field of mobile health has proven that smartphone sensor data can be used to help self-monitor chronic diseases, decrease healthcare visits, and encourage healthy behavior. Methods for passive monitoring of remote patients can facilitate early ailment detection and prevention as well as detecting health deterioration and preventing complications in recovering patients, such as after a sudden injury causing TBI.


The goal of the thesis is two-fold: (1) Proposing a framework that detects whether a subject has Traumatic Brain Injury (TBI), using smartphone sensor mobility data (accelerometer, magnetometer, gyroscope, pedometer, pressure, altitude, accessibility) after an injury occurred. (2) Proposing the concept of Bioscore, which facilitates monitoring recovery trajectories of subjects in the recovery phase. Bioscore is generated by estimating the probability(certainty) (0 - 1) that a subject has a given ailment based on their smartphone-sensed behavior and mobility patterns.

The model uses self-attention mechanism for multimodality feature fusion and a stacked LSTM for TBI prediction, and the Bioscore is generated using Monte-Carlo Dropout uncertainty estimation, converted to estimate certainty.

The classification model achieves a balanced accuracy of 90.2% and a true positive rate of 83.3% in correctly identifying TBI instances. When Bioscore is accurately generated and visualized throughout days, it steadily decreases for users who have normal recoveries, as expected.

The proposed framework could facilitate population-level passive and remote ailment screening and monitoring. A dashboard consisting of Bioscores of a certain community can help preemptively alert and inform individuals of that community to seek medical help and acquire needful medications.