Towards the Use of Machine Learning Classifiers for Human Activity Recognition Using Accelerometer and Heart Rate Data from ActiGraph

Human Activity Recognition (HAR) aims at detecting human physical activities such as eating, running, laying down, sitting, etc., through sensor-generated data. With the ubiquitous nature and utilization of sensor-enabled devices such as smartphones, smartwatches, and wristbands in daily life, numerous modern applications have been developed and implemented in HAR around the world. In this study, rather than using only accelerometry data generated from smartphones which are more commonly adopted in recent literature, we aim to predict human activities using an accelerometer and heart rate (HR) data generated by Actigraph, as they can accurately measure moderate-to-vigorous intensity physical which is mostly affected by body composition and also better suited for self-monitoring. For this purpose, we explored the effectiveness of these features through the application of machine learning classifiers. A very recently publicly available Actigraph-generated data (MMASH) that contains accelerometer and HR recordings were used in the experiments. To evaluate the effectiveness of different indicators for recognising human activities, we performed a series of four experiments.
In working towards recognising four activities, the best-performing machine learning models achieved an averaged accuracy value of 67±11% through using HR as a significant feature. The result shows that HR provides more information that can be used to predict better human activity recognition.