Wearable Sensor Analysis

About the Project

The primary objective of Dr. Keadle’s study is to find the best interventions to reduce sedentary screen time (SST) for each user by at least 60 minutes a day for 16 weeks. Our work this quarter has been focused on understanding the goals of the project, becoming familiar with the datasets available, curating new estimates of activity from existing data, and our comparing estimates to ground truth sources of data. Once we are confident in our data pipelines and processes, we can build machine learning models that automatically classify whether the user’s activity is sedentary or non-sedentary.

Specifically, our 3 main objectives are:

1. Set up a pipeline for merging 3 different sources of data (ActivPal, FitBit, Wemo).
2. Produce code to compare the estimates of sedentary screen time (SST) from each data source.
3. Predict whether the reported SST is a valid estimate of SST based on other features in the datasets such as the time of day, whether the data was self-reported or not, etc.

According to our Ground Truth source of data, User 1001 spent over half their time doing leisure activities during a 3 hour time frame of interest. The majority of the remaining time in the 3 hour window was spent working by this particular user.

Each bar represents an instance where the mobile device recognized a user was sedentary. This can include a user watching TV or using a Tablet. The length of each bar represents how long the user was sedentary, with the longer the bar, the longer the duration. Visualizing each user’s sedentary time can help us evaluate which instances are “white noise” where a device picked up sedentary time when a user was not actually sedentary. For example, the short bars could be noise or mistakes picked up by the mobile device.