3/10/16: Using a Continuous Measurement of Stress (cStress) to Trigger Just-in-Time Stress Intervention on Mobile Phones
March 10, 2016
Karen Hovsepian, PhD
Assistant Professor of CS
University of Memphis
About the Webinar:
This webinar is composed of two major components. In the first half, Hovsepian will present a computational model for continuous measurement of stress from physiological responses, called cStress.
The model (published in ACM UbiComp 2015) is trained using data collected in a lab stress protocol. By leveraging highly precise wearable biosensors, careful data pre-processing, and advanced machine-learning techniques, the model addresses key challenges associated with performing inference in the wild. Researchers carried out extensive model validation on an independent lab study dataset with 26 participants, where the final model obtained a recall rate of 89% and a false positive rate of 5%. We also validated cStress against self-report in a field study dataset with 20 participants, where cStress predicted each instantaneous self-report with an accuracy of 72%.
In the second half, Sarker will present a time-series pattern mining method that mines time series of cStress outputs to devise an optimal strategy for triggering stress interventions in the mobile environment. Markers obtained from cStress constitute a continuous time series with intermittent losses and rapid variability. In addition, sensor-triggered interventions represent an interruption to the user and hence should be delivered only when there is high confidence in the accuracy of the marker. Finally, given that the data collection, computation of marker, and decision of triggering an intervention all happen continuously at the smart phone, battery life of the phone can be depleted quickly. The proposed approach (published in ACM CHI 2016) for triggering just-in-time intervention addresses each of these challenges.
About Karen Hovsepian and Hillol Sarker:
Dr. Karen Hovsepian is an Assistant Professor of Computer Science at Troy University, specializing in Machine Learning and Data Mining. Dr. Hovsepian received his M.S. and Ph.D. in Computer Science from New Mexico Tech, and, since then, has focused on utilizing the predictive and explanatory power of Machine Learning to address key problems in mobile health, bioinformatics, and finance. At MD2K, Dr. Hovsepian leads the team tasked with building an accurate inference model of mental stress — a key factor in conditions and illnesses of primary interest to MD2K researchers. Among his other current research interests is the application of mobile computing to road safety and comfort via development/implementation of Computer Vision and Pattern Recognition algorithms on mobile sensing platforms attached to cars and drivers. More about Karen Hovsepian.
Hillol Sarker’s work aims to identify the optimal timing to proactively trigger just-in-time stress interventions with the help of wearable on-body sensors (e.g., respiration, ECG, and accelerometer) and mobile sensors (e.g., GPS). He defines the timing of intervention as having two components and worked in both parts. First, from the stress likelihood time-series, his work was able to identify the precise timing for just-in-time stress interventions and developed a machine learning model to predict significant stress episodes before their occurrences. Second, his work was able to detect whether a person is available to engage in a task-like intervention which requires significant user involvement. Mr. Sarker believes that such prediction and prevention can improve health and quality of life for everyone, given the ubiquity of stress in daily life and its wide-ranging adverse impact on physical, psychological, behavioral, and social health. His research also focuses on the detection of smoking episodes from wrist sensors and identification of the predictors of smoking relapse. More about Hillol Sarker.