Classifying posture and activity using a wearable device

Abstract

For a patient monitoring system to work, it is important to keep track of a patient’s physiological signals, such as heart rate and blood pressure. Equally important is the availability and analysis of data that can add context to the physiological signals. In this thesis, accelerometer and gyroscope data collected from a medical wearable device are used to infer the posture and activity of the wearer. A simple, threshold based method is presented, that distinguishes between activity and rest and can distinguish between some postures. Additionally, a method is presented that can detect postural transitions, and these transitions are classified using different neural network architectures and other machine learning methods. A fully connected neural network with regularization achieves the best classification accuracy. The same methods are applied to classify postures, activities and postural transitions. The best overall classification accuracy is achieved with a window-based random forest classifier, combined with an LSTM network over those windows. The thesis is concluded with a discussion about the usability of the methods in a real life application.