Modelling stress levels based on physiological responses to web contents

Capturing data on user experience of web applications and browsing is important in many ways. For instance, web designers and developers may find such data quite useful in enhancing navigational features of web pages; rehabilitation therapists, mental-health specialists and other biomedical personnel regularly use computer simulations to monitor and control the behaviour of patients. Marketing and law enforcement agencies are probably two of the most common beneficiaries of such data - with the success of online marketing increasingly requiring a good understanding of customers' online behaviour. On the other hand, law enforcement agents have for long been using lie detection methods - typically relying on human physiological functions - to determine the likelihood of falsehood in interrogations. Quite often, online user experience is studied via tangible measures such as task completion time, surveys and comprehensive tests from which data attributes are generated. Prediction of users' stress level and behaviour in some of these cases depends mostly on task completion time and number of clicks per given time interval. However, such approaches are generally subjective and rely heavily on distributional assumptions making the results prone to recording errors. We propose a novel method - PHYCOB I - that addresses the foregoing issues. Primary data were obtained from laboratory experiments during which forty-four volunteers had their synchronized physiological readings - Skin Conductance Response, Skin Temperature, Eye tracker sensors and users activity attributes taken by a specially designed sensing device. PHYCOB I then collects secondary data attributes from these synchronized physiological readings and uses them for two purposes. Firstly, naturally arising structures in the data are detected via identifying optimal responses and high level tonic phases and secondly users are classified into three different stress levels. The method's novelty derives from its ability to integrate physiological readings and eye movement records to identify hidden correlates by simply computing the delay for each increase in amplitude in reaction to webpages contents. This addresses the problem of latency faced in most physiological readings. Performance comparisons are made with conventional predictive methods such as Neural Network and Logistic Regression whereas multiple runs of the Forward Search algorithm and Principal Component Analysis are used to cross-validate the performance. Results show that PHYCOB I outperforms the conventional models in terms of both accuracy and reliability - that is, the average recoverable natural structures for the three models with respect to accuracy and reliability are more consistent within the PHYCOB I environment than with the other two. There are two main advantages of the proposed method - its resistance to over-fitting and its ability to automatically assess human stress levels while dealing with specific web contents. The latter is particularly important in that it can be used to predict which contents of webpages cause stress-induced emotions to users when involved in online activities. There are numerous potential extensions of the model including, but not limited to, applications in law enforcement - detecting abnormal online behaviour; online shopping (marketing) - predicting what captures customers attention and palliative in biomedical application such as detecting levels of stress in patients during physiotherapy sessions.