The purpose of this study was to develop and validate an algorithm for classifying physical activity (PA) classes and sedentary behavior (SED) from raw acceleration signal measured from hip. Twenty-two adult volunteers completed a pre-defined set of controlled and supervised activities. The activities included nine daily PAs. The participants performed PA trials while wearing a hip-worn 3D accelerometer. Indirect calorimetry was used for measuring energy expenditure. The raw acceleration data were used for training and testing a prediction model in MATLAB environment. The prediction model was built using bagged trees classifier and the most suitable extracted features (mean, maximum, minimum, zero crossing rate, and mean amplitude deviation) were selected using a sequential forward selection method. Leave-one-out cross validation was used for validation. Activities were classified as lying, sitting, light PA (standing, table wiping, floor cleaning, slow walking), moderate PA (fast walking) and vigorous PA (soccer and jogging). The oxygen consumption data were used for estimating the intensity of measured PA. Total accuracy of the prediction model was 96.5%. Mean sensitivity of the model was 95.5% (SD 3.5) and mean specificity 99.1% (SD 0.5). Based on the results PA types can be classified from raw data of the hip-worn 3D accelerometer using supervised machine learning techniques with a high sensitivity and specificity. The developed algorithm has a potential for objective evaluations of PA and SED.
