The AutoSense project is part of the Genes Environment and Health Initiative (GEI) at the National Institutes of Health (NIH). Within GEI, Autosense is one of five cooperative agreement programs under the Exposure Biology Program. The AutoSense project is supported by the National Institute on Drug Abuse (NIDA). The project involves more than ten researchers from Computer Science, Electrical Engineering, Behavioral Science, Physiology, and Biochemistry, spread across the Carnegie Mellon University, Guided Therapeutics, The Ohio State University, the University of Memphis, the University of Minnesota, and the University of Pittsburgh. The principal investigator is Dr. Santosh Kumar from the University of Memphis, and the project scientist is Dr. Marcia Scott from the National Institute on Alcohol Abuse and Alcoholism (NIAAA).

Project Description

AutoSense is an unobtrusively wearable wireless sensor system for continuous assessment of personal exposures to addictive substances and psychosocial stress as experienced by human participants in their natural environments. Currently, AutoSense consists of an arm band with four wireless sensors and a chestband with six wireless sensors. All the ten sensors are integrated onto an embedded platform called “mote,” a tiny self-contained, battery-powered computer with a wireless radio that can host multiple sensors, collect and process data from them using customized algorithms, and communicate on secure wireless channels. More details are available in an ACM SenSys'11 paper.

AutoSense Sensor Suite

System Description: The chestband consists of 2-lead Electrocardiogram (ECG), galvanic skin response (GSR), respiratory inductive plethysmograph (RIP) band for robust measurement of respiration, skin temperature, ambient temperature, and a 3-axis accelerometer. Accelerometers help classify physical activities, estimate their intensities, and help remove motion artifacts from the measurements of ECG, RIP, and GSR. The armband consists of WrisTAS alcohol sensor from Giner Inc., accelerometers, GSR, and temperature sensors. All sensors communicate wirelessly with a smart phone. Sensors on the phone (e.g., GPS, microphone) complement those on the body. The phone also acts as a local server for heavier computation and storage. Additionally, the phone is used for Ecological Momentary Assessment (EMA). The entire system is designed for unobtrusive wearing and robust data collection from the natural environment of human subjects.

AutoSense overcomes several challenges in the design of wearable sensor systems for use in the field. First, it is unobtrusively wearable because it integrates six sensors in a small form factor. Second, it demonstrates a low power design; with a lifetime exceeding ten days while continuously sampling and transmitting sensor measurements. Third, sensor measurements are robust to several sources of errors and confounds inherent in field usage. Fourth, it integrates an ANT radio for low power and integrated quality of service guarantees, even in crowded environments. The AutoSense suite is complemented with a software framework on a smart phone that processes sensor measurements received from AutoSense to infer stress, conversation, activity, and other rich human behaviors.

On-body Computation. Signals from the wireless sensors are processed locally on the mote to filter out noises and artifacts. On the accompanying smart phone, more than 60 features (e.g., heart rate, heart rate variability, etc.) are derived from ten sensors on the body and two sensors on the phone, for estimation of psychosocial stress & blood alcohol content, identification of smoking episodes, identification of place (home, work, etc.), identification and characterization of conversations, classification of physical activities, identification of commuting modalities, etc. Simpler features are computed on the mote, while the computationally intensive ones are computed on the phone. Machine-learning methods are used to personalize the inferencing of various human states to account for wide between subject differences. Training for personalization is achieved by eliciting self-report on the phone, when significant changes in the human state are suspected by the system (e.g., increase in the stress level). Micropayments are used to encourage the wearing of the wireless sensors and timely responses to the EMA interview.

Measures Collected by AutoSense. AutoSense collects continuous measures from ten sensors (ECG, respiration, alcohol, GSR at chest and upper arm, skin and ambient temperature at chest and upper arm, and accelerometry of the torso and upper arm). These sensor measurements are used to obtain continuous measures of physiological stress, psychological stress level, conversation status and duration, posture and physical activity level, location, and commuting status. Significant changes in any of these measures can be used to trigger a self-report on a smart phone in real-time.

Use in Lab and Field Studies: Several user studies are ongoing and scheduled to evaluate the sensitivity, specificity, and usability of the AutoSense system.

The first study (16 subjects – 8 men, 8 women), completed in Summer’10, integrated a lab session with three acute stressors (public speaking, cold pressor, and mental arithmetic) with two sessions of 24 hours in the wild. Self-reports on stress levels were collected 20 times on each day in the field. Data collected from this study was used to develop and evaluate models for inferring physiological stress and psychological stress from ECG and respiration measurements. These models were found to detect lab stress with 90% accuracy and had a correlation of 0.71 with daily stress levels self-reported in the field. It was also found that respiration measurements have similar discriminatory power for detecting stress as ECG. These results are summarized in a 12-page ACM IPSN'11 paper that was nominated for best paper award.

In the second study, 36 participants wore AutoSense for 3 continuous days in the wild. They were divided into three categories to evaluate the effect of variation in compensation structure. As reported in an ACM UbiComp'11 paper, it was found that adaptive micro-payment structure for compensating participants led to greater quality and quantity of data. In addition, the participants filled out a survey on their privacy concerns after completing the study and after being show the behaviors (stress, activity, conversation, and commuting) that was inferred automatically by AutoSense from the data collected on them over 3 days. As reported in an ACM CHI'11 paper, people are largely unaware of the privacy risks that emerge from the use of mobile health devices such as AutoSense.

The third study (30 subjects – 15 men, 15 women, daily smokers and social drinkers), in currently ongoing. In this study, participants wear the entire suite of sensors for a week at a time in the wild. The data collected from the study will be used to develop models for automatically classifying emotion into two classes (positive and negative), identifying smoking episodes, and identifying drinking episodes.