Background: Sedentarism is associated with the development of noncommunicable diseases (NCD) such as cardiovascular diseases (CVD), type 2 diabetes, and cancer. Therefore, the identification of specific sedentary behaviors (TV viewing, sitting at work, driving, relaxing, etc.) is especially relevant for planning personalized prevention programs.
Objective: To build and evaluate a public a dataset for the automatic recognition (classification) of sedentary behaviors.
Results: The dataset included data from 30 subjects, who performed 23 sedentary behaviors while wearing a commercial wearable on the wrist, a smartphone on the hip and another in the thigh. Bluetooth Low Energy (BLE) beacons were used in order to improve the automatic classification of different sedentary behaviors. The study also compared six well know data mining classification techniques in order to identify the more precise method of solving the classification problem of the 23 defined behaviors.
Conclusions: A better classification accuracy was obtained using the Random Forest algorithm and when data were collected from the phone on the hip. Furthermore, the use of beacons as a reference for obtaining the symbolic location of the individual improved the precision of the classification.
Keywords: Machine learning; activities of daily living; dataset; sedentary behavior.