A Decision-tree Model For Classifying Physical Activity Types Using A Three-axis Accelerometer In Japanese Adults

Abstract

Accelerometry is a valid method for assessing physical activity in free-living conditions. Although there are numerous studies on estimating physical activity-induced energy expenditure (PAEE), classification of physical activity types using an accelerometer is not well established. PURPOSE: To develop a physical activity classification model using a three-axis accelerometer in Japanese subjects. METHODS: Eleven healthy adults (age, 33.3±10.4 years) stayed in a human calorimeter for 23 hours and performed activities (sitting, standing, walking, lying down) while wearing a three-axis accelerometer (Actigraph-wGT3X-BT). Raw three-axis acceleration values were obtained at 30 Hz. Raw acceleration values were processed into five feature quantities: filtered acceleration values (FAC), raw acceleration values (RAC), absolute value of filtered accelerations (ABS), standard deviation of FAC (SD), and peak-to-peak of FAC (PP). A classification model was proposed using a decision tree and an 11-fold cross-validation method with time resolutions of 1, 5, or 10 seconds. RESULTS: The classification model achieved the following accuracy levels for each time segment for all activities: 75.1% with 1-second segments, 76.1% with 5-second segments, and 77.7% with 10-second segments. Selected feature quantities in these models were RAC, ABS, and PP in 1-second segments; FAC, RAC, ABS, SD, and PP in 5-second segments, and FAC, RAC, and PP in 10-second segments. The classification model with a 10-second time resolution had better accuracy than with 5- or 1-second time resolution. The model’s accuracy for each type of physical activity with 10-second segments was 80.6% for sitting, 59.6% for standing, 99.7% for walking, and 84.1% for lying down. CONCLUSIONS: These results suggest that the decision tree is a valid method for physical activity classification in Japanese subjects. The addition of other features and classification methods, e.g., random forest, needs to be discussed in future studies.