Smartphone TBI Sensing using Deep Embedded Clustering and Extreme Boosted Outlier Detection

Traumatic Brain Injury (TBI), caused by a severe impact to the head, can have long-lasting and possibly life-long disability of patients. This ultimately creates a huge economic and social burden on patients and the healthcare system. Many TBI patients do not get early and adequate medical care. Sensor-rich, ubiquitously owned smartphones can now be used to passively sense a wide range of ailments, facilitating continuous monitoring of patients and high-risk groups in the real world. In this paper, we propose a deep learning approach for distinguishing smartphone users with TBI from healthy controls based on smartphone-sensed behaviors within 24-hours of the injury. Our method analyzes smartphone sensor data by first utilizing Deep Embedded Clustering (DEC) to identify clusters of users with similar smartphone-sensed behaviors. Extreme Gradient Boosted Outlier Detection (XGBOD) is then employed on each of the identified clusters to predict users with TBI. In rigorous evaluation, our method achieved a balanced accuracy of 88 % and a sensitivity of 74 %. Our proposed method can flag smartphone users with TBI, enabling them to receive early medical attention and improve their prognostic outlook.