Detection of emotional stress from physiological data using wearables

Research contained in this thesis suggests that wearable devices (wearables) can be used beyond traditional sports and fitness applications and can be used to detect changes in human emotional states relating to stress. Studies into stress have pointed out that Australia alone loses $11bn annually because of people suffering from stress and the resulting low productivity/absenteeism associated with this condition. This research has several applications ranging from social awareness to business productivity. Our research enquiry asks the question; how can a person’s stress level be reliably and uniquely detected using physiological data from wearable sensors? To answer the central question, this research covers diverse areas such as emotional psychology, physiology, affective computing, and data mining. Whilst wearables pack a vast array of physiological sensors that detect the heart rate, electrodermal activity, temperature, and respiration, making sense of this raw data required further investigation.
Three key research contributions in this work make up a framework that maximises interpretable machine learning and knowledge discovery. Firstly, a pre-processing method for time series physiological data which maximises interpretable machine learning. Secondly, a classification method using trees and forest algorithms employing the fusion of multiple physiological modalities for stress detection. Thirdly, a knowledge discovery method for selecting high-quality logic rules representative of stress. In addition to these research contributions, to validate this framework, an experiment was conducted involving collecting physiological data from 27 participants over three days of their train travel and work routine. The resultant anonymised dataset, contains over 700 hours of raw physiological data, including heart, temperature, electrodermal activity and accelerometer data. In addition to the raw data, the labelled classification datasets for each participant will be published for the benefit of the research community.