Gregory J Norman*, University of California, San Diego, San Diego, United States Wanmin Wu, University of California, San Diego, San Diego, United States Ernesto Ramirez, University of California, San Diego, San Diego, United States Carlyn Peterson, University of California, San Diego, San Diego, United States Sanjoy Dasgupta, University of California, San Diego, San Diego, United States Track: Research Presentation Topic: Persuasive communication and technology Presentation Type: Oral presentation Submission Type: Single Presentation Building: Joseph B. Martin Conference Center at Harvard Medical School Room: A-Pechet Room Date: 2012-09-15 11:45 AM – 12:30 PM Last modified: 2012-09-12

If you are the presenter of this abstract (or if you cite this abstract in a talk or on a poster), please show the QR code in your slide or poster (QR code contains this URL).

Background:
Lack of physical activity (PA) and continual sedentary behavior (SB) over time can contribute to obesity and morbidity. In this study we are developing integrated tools to measure and change PA and SB. The first step was to create a valid activity classification tool that uses sensors (e.g., accelerometer and gyroscope) currently onboard most smart mobile phones. Data from these sensors were processed using machine learning to classify activity types.

Objective:
We examined the validity of the iPod Touch (Apple, Inc.) for classifying types of physical activity, with the goal of creating a measurement and feedback system that easily integrates into individuals’ daily living.

Methods:
Sixteen adults (8 males, 8 females; age = 41.19, BMI = 28.82 kg/m2) completed between 9 and 16 activities in a laboratory and in a simulated free-living condition. In the laboratory 3-minute bouts on the treadmill included walking at a slow pace (1.5 mph), a normal pace (3.0 mph), a brisk pace (4.0 mph), and jogging (5.5 mph). Participants then recorded time sitting and bouts of walking up and down stairs at a brisk and normal pace. This was followed by completing 400m self-paced walking bouts at slow, normal, and brisk paces as well as one 400m jog on an outdoor track. The iPod Touch was carried on the participant in the front shorts pocket or on the arm.

Results:
To create training data for machine learning, the start and end time of each activity file was auto-trimmed based on the running variance sum of the x, y, and z axis accelerations. Mean and variance of the vector magnitude of the 3-axis acceleration values served as the extracted features for classification. Different size sliding windows of data segments were tested and a 1-second window of approximately 30 data points was found to work best. Decision tree, Naïve Bayes and K-nearest neighbor algorithms were tested. Using a leave-one-out training and validating protocol, classification accuracies for the individual activities were poor for the three algorithms (17.14 % to 21.48%). However, when activities were grouped into three broad categories of walking, jogging and sitting, the average accuracies improved substantially to 80.3% for decision tree, 80.6% for KNN, and 98.4% for Naïve Bayes.

Conclusions:
Using minimally extracted features from onboard smart phone sensors general categories of activity (walking, jogging, and sitting) were classified with high accuracy using Naïve Bayes. This first step in our study gave an important indication of the possibilities and limitations of using a smart phone as an activity data collector. This system has potential high ecological validity because it requires people to only carry one device that they commonly carry with them already. The next step in our research is to test an onboard classifier application on the phone that can prompt users when needed for annotations in order to learn and classify individual activity patterns with high accuracy. The final step will be testing the feedback component that can offer individually-tailored prompts and suggestions to increase PA and decrease SB time.