Validation in Clinic & Field

Several user studies are being conducted both in the controlled laboratory environment and in the uncontrolled natural environment with human volunteers to evaluate the accuracy/concordance of measurements and inferences produced by the AutoSense system to that obtained from self-reports and to assess participant compliance with, and acceptance of the AutoSense data collection protocol. These studies are described below.

AutoSense Lab Study

This study (conducted at the University of Minnesota Medical School) aims to evaluate the quality of data collected by the AutoSense chestband and alcohol sensor in a controlled lab setting. Participants are young healthy males and females (21-50 yrs). The screening protocol includes assessment of medical history, a medical exam, and completion of questionnaires on alcohol use and stress. The clinical study design includes repeated measurement of alcohol levels in blood and in interstitial fluids, and measurement of cardiovascular variables, cortisol concentrations, and mood reports during two sessions (alcohol and placebo). Specifically:

• Alcohol content – blood concentration, breathalyzer (eight blood and breathalyzer samples)
• ISF alcohol sensing (one minute intervals throughout lab session)
• Salivary cortisol samples (total of eight times during rest and in response to acute stressor)
• Blood cortisol samples (total of eight times during rest and in response to acute stressor)
• Cardiovascular: Systolic blood pressure, diastolic blood pressure, and heart rate (every 3 minutes before, during and after acute stressor (public speaking exercise)
• Mood states: Positive affect; distress; physical symptoms; Biphasic Alcohol Effects Scale (BAES)

Participants are exposed to three types of acute stressors:

• Public speaking to simulate social, evaluative challenge
• Mental arithmetic problem to simulate cognitive challenge
• Cold pressor test to simulate acute physical challenge

Out of 76 participants screened on-site, 42 have completed the study and more subjects are in progress. Preliminary analysis of data from 34 subjects shows

visible gender differences in breath alcohol levels.

In August 2009, we began using the chestband. Analysis for chestband data in progress. We plan to complete greater than 25 subjects with the chestband.

The following figure shows the effect of alcohol on participant stress responses.

Personalized Stress Inferencing Study

—This study (conducted at the University of Minnesota) aims to validate a real-time machine learning algorithm to infer stress from physiological measurements in the field. This algorithm produces inferences that are:

• —Indicative of perceived stress
• —Personalized to each individual
• —Adaptive to changes in context (e.g., posture, speech)

The study is divided into three stages:

1. Lab Study: This stage aims to capture training data for the personalized stress inferencing algorithm. The participant is exposed to various stressors (public speaking, mental arithmetic, and cold pressor) while they wear the AutoSense chestband. This enables capturing physiological responses to stressors. In between stressors, the participant provides self-reports of their perceived stress via questionnaire. The self-report and physiological measures are then used to develop an initial stress inferencing algorithm personalized to the participant.
2. Field Study Day One: This stage aims to improve the initial stress inferencing algorithm using new training dagta captured in the field. The participant is sent into the field wearing the AutoSense chestband. He/she also carries a mobile phone loaded with the initial stress inferencing algorithm personalized to his/her specific stress signature (measured in the lab study). As the participant goes about the day, the mobile phone periodically prompts the participant to answer questionnaires assessing his/her stress levels (ecological momentary assessments). Questionnaires appear when the inferencing algorithm believes the participant is stressed, as well as at fixed intervals. This enables capturing data to verify and improve the initial stress inferencing algorithm, as well as capture baseline measures of stress throughout the day. To ensure the participant is exposed to stress in the field, the participant will take a college exam during field study day one.
3. Field Study Day Two: This stage aims to validate the participant's final stress inferencing algorithm (derived from data gathered in the previous stage). The procedure is identical to that in field study day one, except that the participant does not take an exam or undergo any type of designated stressor. The questionnaire data is used to validate real-time stress inferences.

The study population consists of 16 students. The study begins in February 2010.

AutoSense Field Study

This study aims to improve and validate the AutoSense system for use in future field studies measuring personal exposure to alcohol, smoking, and psychosocial stress. 30 university students (15 male) who are smokers will be recruited from the University of Memphis. As in the personalized stress inferencing study, each participant will be asked to participate in three stages: Lab Study, Field Study I, and Field Study II. During the Lab Study, physiological and self-reported measures of stress will be collected from the participant while they watch movies in a lab setting. This data will be used to generate an initial version of a software detector of stress (positive and negative) personalized to that participant. Approximately four weeks later, when the initial stress detector is ready, Field Study I will begin. In Field Study I, physiological and self-report data will be collected from the participant in his/her natural environment for seven consecutive days. These data will be used to test the initial stress detector and build an improved personalized stress detector. Field Study II begins approximately four weeks later, when the improved stress detector is ready. In Field Study II, physiological and self-report data will be collected from the participant in his/her natural environment for an additional seven consecutive days. This data will be used to validate the improved stress detector. The first and second field studies will be timed to coincide with exam weeks, to improve the chances of capturing both low and high stress and positive and negative stress.

In all three phases of the study, participants will wear AutoSense. In the field studies, participants will also carry a mobile smart phone customized to communicate with the AutoSense sensor suite. The mobile phone has three roles. First, it will use signals captured by AutoSense and sensors on the phone (phone signals: 3-axis acceleration, names of nearby Bluetooth and WiFi devices, GPS traces, and user interaction) to infer participant stress levels and the timing and characteristics of situational factors associated with stress. These factors include alcohol consumption, conversation episodes, smoking episodes, exposure to smoking hotspots, physical activity levels, posture, physical activity, places visited, and commuting episodes. Second, participants will use the phone to complete questionnaires in the field. The field questionnaire will ask participants to assess their subjective stress level as well as indicate if they are experiencing any of the previously described situational factors associated with stress. The questionnaire responses will be used to improve and validate the stress and situational factors detection algorithms used in Field Study I and II, respectively. Third, participants will indicate the beginning of drinking and smoking episodes in the field by pushing a button on the mobile phone.