Results
Automated Detection of Sensor Detachments for Physiological Sensing in the Wild
Body area sensor networks measure biomedical signals from subjects continuously, as they go about their daily lives. Signals measured in these conditions are affected by anomalies, such as artifacts and noise. Some anomalies can be corrected, if detected in real-time, for example, ECG electrode detachment. We have identified energy and computationally efficient algorithms for the detection of sensor detachment for the AutoSense system.
mStress: Supporting Continuous Collection of Objective and Subjective Measures of Psychosocial Stress on Mobile Devices
Excessive, chronic, and repeated exposure to psychological stress can lead to significant health problems. However, new methods for better coping with stress that could significantly improve health and quality of life, cannot be developed and evaluated without scientifically valid datasets describing the experience of stress in everyday life. In prior research, scientifically valid datasets have been difficult to capture from natural environments. Sensors, which continuously capture objective information about physiology and behavior, are prone to noise and failure. In addition, aspects of everyday life (e.g., conversation, exercise, etc.) interfere with the physiological response to stress, making it difficult to tease out the effect of stress from changes in physiology. To overcome the challenges of assessing both exposures and responses to stressful events, new wireless sensing systems are needed to capture scientifically valid datasets describing the experience of stress in natural environments.
We designed and evaluated mStress, a smartphone (Android G1) based system that continuously collects and processes multi-modal measurements from six body-worn wireless sensors to infer in real-time whether the subject wearing the sensors is stressed. mStress generates prompts for timely collection of self-reports, triggered
by real-time changes in stress level inferred by the system, to collect the subjective experience of stress when it is fresh in the participant’s mind. To improve the quality of data, mStress incorporates several features including paying micro-incentives for timely completion of self-reports, realtime detection of and response to confounding factors that affect physiological signals, and real-time detection of sensor detachments so the participant can rectify themselves.
All of this functionality occurs entirely on the mobile phone without any help from the back-end cloud.
mStress was used by 23 human volunteers in a scientific study, in which each participant wore the sensors and provided self-reports during their wake hours for one full day in their natural environment. The phone lasted over 14 hours. Over 200 million samples of sensor measurements were collected, 19,000 stress predictions were made, and 803 prompts for self-report were answered, 98% of which were completed within 7 minutes of the prompt.
Architectural Diagram of the mStress Framework
Privacy Risks Emerging from the Adoption of Innocuous Wearable Sensors in the Mobile Environment
Wearable sensors are revolutionizing healthcare and scientific research by enabling capture of physiological, psychological, and behavioral measurements in natural environments. However, these seemingly innocuous sensor measurements can be used to infer potentially private behaviors such as stress, conversation, smoking, drinking, illicit drug usage, among several others. We conducted a user study to assess how concerned people are about disclosure of a variety of behaviors and contexts that are embedded in wearable sensor data. Our results show participants are most concerned about disclosures of conversation episodes and stress — inferences that are not yet widely publicized. These concerns are mediated by both temporal and physical context associated with the data and the participant’s personal stake in the data. Our results provide key guidance on the extent to which people understand the potential for harm and data characteristics privacy researchers should focus on to reduce the perceived harm from such datasets.
A conceptual framework for reasoning about privacy
issues in sharing personal sensory data.
Participant concern level with respect to selected behaviors and contexts.
Group NS, S-Pre, and S-Post represent increasing levels of personal stake in the
data. The height of the bars represent the mean concern level, and the error
bars depict ± one standard deviation away from the mean.