A computational model of observer stress

Abstract

Stress is a major growing concern in our age, adversely impacting individuals and society. Stress research has a wide range of benefits with the potential to improve health and wellbeing, personal day-to-day activities, increase work productivity and benefit society as a whole. This makes it an interesting and socially beneficial area of research. It motivates objective understanding of how average individuals respond to events they observe in typical environments they encounter, which this thesis investigates through artificial intelligence particularly bio-inspired computing and data mining. This thesis presents a review of the sensors that show symptoms which have been used to detect stress and computational modelling of stress. It discusses non-invasive and unobtrusive sensors for measuring computed stress. The focus is on sensors that do not impede everyday activities which could be used to monitor stress levels on a regular basis. Several computational techniques have been developed previously by others to model stress based on techniques including machine learning techniques but these are quite simplistic and inadequate. This thesis presents novel enhanced methods for modelling stress for classification and prediction using real-world stress data sets. The main aims for this thesis are to propose and define the concept of observer stress and develop computational models of observer stress for typical environments. The environments considered in this thesis are abstract virtual environments (text), virtual environments (films) and real environments (real-life settings). The research comprised stress data capture for the environments, multi-sensor signal processing and fusion, and knowledge discovery methods for the computational models to recognise and predict observer stress. Experiments were designed and conducted to acquire real-world observer stress data sets for the different environments. The data sets contain physiological and physical sensor signals of observers and survey reports that validate stress in the environments. The physiological stress signals in the data sets include electroencephalogram (EEG), electrocardiogram (ECG), galvanic skin response, blood pressure and the physical signals include eye gaze, pupil dilation and videos of faces in visible and thermal spectrums. Observer stress modelling systems were developed using analytics on the stress data sets. The systems generated stress features from the data and used these features to develop computational models based on techniques such as support vector machines and artificial neural networks to capture stress patterns. Some systems also optimised features using techniques such as genetic algorithm or correlation based techniques for developing models to capture better stress patterns for observer stress recognition. Additionally, a computational stress signal predictor system was developed to model temporal stress. This system was based on a novel combination of support vector machine, genetic algorithm and an artificial neural network. This thesis contributes a significant dimension to computational stress research. It investigates observer stress, proposes novel computational methods for stress, models stress with novel stress feature sets, and proposes a model for a temporal stress measure. The research outcomes provide an objective understanding on stress levels of observers, and environments based on observer perceptions. Further research suggested includes investigating models to manage stress conditions and observer behaviours.