Fully-Automated Human Activity Recognition with Transition Awareness from Wearable Sensor Data for mHealth

Wearable sensor-based activity trackers often suffer from data during transitions, which are very different in nature from the actual activities of interest. For increased reliability of activity classification, it is important to classify transitions accurately. This study presents a set of activity recognition algorithm based on decision trees with ensemble approach while autonomously deals with the transition values necessary to separate from normal activity data in the real-time environment. For this purpose, we have used Mobile Health (mHealth) open-access dataset from the UCI Machine Learning Repository. In this study, we investigate ensemble method bagging tree algorithm with the leave-one-subject-out approach to determine the best technique to deal with the null values while detecting regular activities of interest for a fully-automated system. We have also focused on the usage of minimum data and sensors for allowing real-time applications. The data were collected from accelerometer, gyroscope, and magnetometer located on chest, right-lower arm and left ankle. We have measured the sensitivity and specificity to determine the efficacy of our approach. Based on all of the performance metrics, Bagged trees, an ensemble method, has performed better than previously reported algorithm and needed fewer data and fewer sensors. Our approach has a weighted sensitivity of 95.2% and a weighted specificity of 94.9%. The results show that transitions can be efficiently detected while recognizing other activities from mHealth wearable data for health and well-being monitoring of smart and connected communities (S&CC).