Using mobile phones and sensors to identify physical activity type in real time

Accelerometer logging devices, such as Actigraphs, can measure gross overall activity intensity, but knowledge of activity type might prove valuable for both measurement and intervention studies. Some recent work has demonstrated that postures and certain ambulatory activities can be recognized using statistical machine learning algorithms. These algorithms work by analyzing features computed from raw accelerometer data, given a set of example activities, and then computing models for each target activity that can be automatically and quickly matched to new incoming data streams using statistical pattern matching. Funded by the NIH's Genes and Environment Initiative, we are developing new wireless accelerometers that communicate with the phone and permit detection of a wide variety of physical activities at the moment that they are happening without requiring a person to wear bulky or wired equipment. We will demonstrate the prototype system and discuss pilot results of testing the algorithm in the lab and in free-living conditions. In one experiment with 20 subjects, an overall accuracy of 75% on 51 different physical activities was obtained with 2 min of training data per activity. If activity types of different intensities are clustered (i.e., “walking 2mph” and “walking 3mph” are clustered as “walking”), overall activity type detection accuracy increased to 91%. We will discuss how this system might be used for both physical activity measurement research and enabling novel physical activity health promotion interventions that provide just-in-time, tailored, physical activity feedback using the mobile phones.