Continuous Inference of Psychological Stress from Sensory Measurements Collected in the Natural Environment Kurt Plarre  , Andrew Raij ✷ , Syed Monowar Hossain  , Amin Ahsan Ali  , Motohiro Nakajima ‡ , Mustafa al’Absi ‡ , Emre Ertin  , Thomas Kamarck ∨ , Santosh Kumar  , Marcia Scott  , Daniel Siewiorek † , Asim Smailagic † , Lorentz E. Wittmers, Jr. ‡ University of Memphis  , University of South Florida ✷ , University of Minnesota Medical School ‡ , The Ohio State University  , Carnegie Mellon University † , University of Pittsburgh ∨ , National Institute on Alcohol Abuse and Alcoholism  ABSTRACT Repeated exposures to psychological stress can lead to or worsen diseases of slow accumulation such as heart diseases and cancer. The main challenge in addressing the growing epidemic of stress is a lack of robust methods to measure a person’s exposure to stress in the natural environment. Pe- riodic self-reports collect only subjective aspects, often miss stress episodes, and impose significant burden on subjects. Physiological sensors provide objective and continuous mea- sures of stress response, but exhibit wide between-person dif- ferences and are easily confounded by daily activities (e.g., speaking, physical movements, coffee intake, etc.). In this paper, we propose, train, and test two models for continuous prediction of stress from physiological mea- surements captured by unobtrusive, wearable sensors. The first model is a physiological classifier that predicts whether changes in physiology represent stress. Since the effect of stress may persist in the mind longer than its acute effect on physiology, we propose a perceived stress model to predict perception of stress. It uses the output of the physiological classifier to model the accumulation and gradual decay of stress in the mind. To account for wide between-person dif- ferences, both models self-calibrate to each subject. Both models were trained using data collected from 21 subjects in a lab study, where they were exposed to cog- nitive, physical, and social stressors representative of that experienced in the natural environment. Our physiologi- cal classifier achieves 90% accuracy and our perceived stress model achieves a median correlation of 0.72 with self-reported rating. We also evaluate the perceived stress model on data collected from 17 participants in a two-day field study, and find that the average rating of stress obtained from our model has a correlation of 0.71 with that obtained from pe- riodic self-reports. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. IPSN’11, April 12–14, 2011, Chicago, Illinois. Copyright 2011 ACM 978-1-4503-0512-9/11/04 ...$10.00. Categories and Subject Descriptors C.3 [Special-Purpose and Application-Based Systems] General Terms Algorithms, Experimentation, Human Factors, Measurement Keywords Wearable sensors, physiological monitoring, stress inference 1. INTRODUCTION In moderation, stress can be a positive force in everyday life. It can motivate action (e.g., when in danger), improve performance, and increase excitement [24, 44]. However, ex- cessive, chronic, and repeated exposures to stress can lead to significant negative health consequences [38, 29]. Excessive stress can lead to headaches, trouble sleeping, and fatigue [30, 11, 3]. In the longer term, stress can be associated with risk for several chronic diseases including cardiovascu- lar diseases [37, 42]. Animal and human studies have shown that stress can also play a role in psychological or behavioral problems, such as depression, addiction, rage, and anxiety [22, 2, 13, 14]. The main challenge in addressing the nega- tive consequences of stress is a lack of robust methods that can continuously measure a person’s exposure to stress in the natural environment. In behavioral science, periodic self-reports are commonly used to measure perceived stress in natural environments. Self-reports allow collection of instantaneous measurements of perceived stress, often multiple times per day to reach a desired sampling of stress. However, self-reports only cap- ture subjective aspects of stress, may miss stress episodes, and impose significant burden on the subject. Since self-reports only capture perception of stress, they do not provide a proximal measure of the physical health consequences of stress, such as cardiovascular wear and tear, ulcer, and cancer. In addition, the episodic nature of stress means that discrete self-reports can miss stress episodes. To ensure capture of stress episodes, a continuous measure of stress is needed. Finally, the active participation required to provide self-reports means that self-reports are burden- some and obtrusive. To provide a self-report, a person must be willing to have their daily life interrupted to complete self-reports, sometimes as many as 20 times per day. A high 97