Statistical analysis of wearable sensor data
Presenter: Julian Wolfson
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Predicting human activity using smartphone GPS and accelerometer data
The near-ubiquity of smartphones in developed countries means that a sizable fraction of the adult population in those countries is carrying a remarkably powerful computer with advanced sensing capabilities. Modern smartphones are equipped with a GPS receiver, accelerometer, and gyroscope, all of which receive and record a constant stream of data throughout the day. Data from these sensors provide clues to a user�s daily activities: a 12-hour overnight stay is likely to be at home, a 20-minute trip at an average speed of 35 mph is likely to have been made by car or bus, etc. Taken together, particularly over time, the sensor data can be used to form a �snapshot� of the user�s behavior patterns. Smartphones are therefore potentially a very powerful tool for researchers eager to collect more objective behavioral data and to understand how behaviors relate to health and well-being. Another advantage in the research context is that user burden is relatively low; by using statistical machine learning methods, daily activities can be inferred directly from the sensor data without requiring constant user input. This low burden (and resulting decrease in respondent fatigue) allows smartphones to be deployed as a data collection tool in larger studies, over longer time periods, for a lower cost. Smartphones are also being used to deliver interventions which may be personalized according to a user�s stated preferences, previous behavior, or current location. In this talk, I will provide an overview of the data available from smartphone sensors, and describe some of the techniques available for analyzing them. Examples will come from my own experience helping to develop a smartphone application, Daynamica, which uses GPS and acclerometer data to automaticaly detect and predict a user's daily activities. I will also discuss potential applications of this technology in health research, and highlight some of the key statistical challenges for prediction and inference using smartphone sensor data.
Statistical analysis of wearable sensor data
Presenter: Vadim Zipunnikov
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Multilevel Functional Methods for modeling actigraphy data and its application to predicting mortality in the US population
The crashing wave of activity tracking �wearables� opens up an opportunity to unveil previously hidden but pivotal signatures of disability and disease. To achieve this promise, the understanding, interpretation and analysis of complex multimodal and multilevel data produced by such devices becomes crucial. The first part of my talk will provide an overview of the instruments that are available for real-time measurement of physical activity as well as a quick review of the strengths and limitations of current analytical approaches for modeling physical activity data. In the second part, I will talk about analysis of physical activity data collected on 10000+ subjects in National Health and Nutrition Examination Survey (NHANES), a nationally representative sample of the US population. I will present recent multilevel functional data approaches that separate and quantify the systematic and random circadian patterns of physical activity, model them as functions of age, gender, and dominant comorbidities and demonstrate that these patterns are powerful predictors of mortality.
Statistical analysis of wearable sensor data
Presenter: Tiago Barreira
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
24 hour accelerometry data collection, challenges and opportunities
In recent times, researchers have begun to collect accelerometer data over the entire day (24 hours) as opposed to during waking hours only. This methodological shift serves two purposes: 1) to increase compliance to objective monitoring protocols and consequently wear time, and 2) to assess sleep time. This paradigm shift is not without problems, as the separation of sleep, non-wear, and sedentary behavior can be an arduous task complicated by the similarities exhibited in accelerometer data from these behaviors. Together with my colleagues, we created an automated algorithm (publically available) to identify different activity types (e.g., sleep, non-wear, sedentary behavior, and physical activity) from minute-by-minute accelerometer data. I will discuss the data collection procedures, the challenges with data processing and data analysis and the opportunities that gathering 24-hour accelerometer data presents.
Statistical analysis of wearable sensor data
Presenter: Jeff Goldsmith
When: Monday, July 11, 2016 Time: 4:00 PM - 5:30 PM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Discussion
Discussion of session talks on physical activity measurement and quantification.
