Abstract:
|
Passively collected smartphone sensor data (e.g. GPS, accelerometer, gyroscope) offer a promising solution to the phenotyping problem in the biomedical sciences. These data may provide a basis for refined social and behavioral markers, and can inform novel strategies for monitoring patients and delivering interventions. Identifying the proximity of a smartphone to its user is a key component of any analysis: Only when the phone is "on-person" do we know that sensor data reflect the geographic location and behavior of the user. Our approach to classifying "on-person" and "off-person" times is based on a comparison of observed triaxial accelerometer signal variability to the baseline variability associated with sensor noise. Our method requires minimal training data, and scales readily to studies with large numbers of participants. In our preliminary study, it attained sensitivity of 97% and specificity of 93%. We present the details of our method and an application to smartphone accelerometer data from several studies, including a study involving patients with cancer at a Harvard Medical School teaching
|