Wearable Sensors and MET Features: The importance of Preprocessing and Understanding the Data

Abstract

Discerning user activities with wearable devices is important to monitor people's behavior and even for developing personalized strategies to prevent some diseases, which has significant implications for healthcare systems. The metabolic equivalent of the task (MET), a measure of the energy cost of physical activities, is a good indicator to discern different activities. However, there is no established formula for calculating the MET. Some works have used several features to approximate the MET, such as the filtered acceleration (ACCfil) or the percentage of the heart rate reserve (%HRR). However, none of them have studied the importance of each of these features in the accuracy of activity classification. This study investigates the impact of different preprocessing techniques on activity recognition accuracy, such as filtering the acceleration to obtain the feature ACCfil. To that end, we performed several experiments with different features and machine learning algorithms to detect the activities. Our results indicate that incorporating features such as ACCfil, HRR, and ratio of unfiltered to filtered acceleration (RUF) improves activity recognition accuracy in some cases up to 166.67%. These features are particularly useful for challenging activities such as stair climbing or household activities. These findings, with their direct implications for developing more accurate activity recognition models with wearable devices, underscore the importance of preprocessing techniques and feature integration.