Decoding sensors data using machine learning algorithms to detect an individual’s stress levels

Abstract

Digital health is currently on the rise. Through the use of modern technologies and machine learning algorithms, it is possible to significantly increase an individual’s quality of life. Given the impact psychological stress has on human lives, the aim of this thesis is to provide further information about classifying the relationship between this stress and the human body’s vital signs, all through the use of physiological data collected by sensors which is later cleansed and processed for use with a variety of machine learning algorithms. A good relationship between psychological stress and vital signs has already been established. Once this correlation is efficiently reproduced through machine learning algorithms, a lot of progress can be made, fundamentally changing the way people go about their daily lives. When considering research on the matter, there clearly exist a plethora of studies making use of a similar setup. By making use of laboratory or ambulatory stressors, a protocol is designed in which labelled data can be generated, discriminating between an individual's stressed or baseline conditions. By simply recording these labels and using sensors to monitor and/or collect the individual’s vital signs throughout such phases, a relationship between the two can be found. A common occurrence with these studies was the use of sensors collecting data from the cardiovascular system, due to its strong correlation with stress. Skin Temperature, on the other hand, is not as effective when employed for stress classification. The motivation behind this project was therefore to use a similar approach as these studies and further the research available on the use of ECG and skin temperature in such scenarios, whilst comparing results of various variables such as different normalization techniques and popular machine learning algorithms. The initial phase of the project consisted in getting to know the sensors and the state of the art. For, it is only when the technology is understood that it is possible to truly harness its potential. An Artifact was developed which prepares and classifies the physiological data as required. Throughout the building phase, the WESAD dataset was used as the main source of physiological data, consisting of both sensor data as well as their labels. For pre-processing, a Butterworth filter with varying cut-off frequencies was implemented to cleanse the data, after which the Pan-Tompkins algorithm was used to extract the RR-intervals from the cleansed ECG signal. Twenty well-known features were then extracted from these intervals. For skin temperature on the other hand, the readings are very simple and therefore only 6 statistical features could be extracted. Once extracted, these features are then checked for correlation with their labels, normalised and passed on to the 4 implemented classifiers: SVM, Knn, DT and GNB. One of the issues present with an approach making use of sensors is that data gathered from such devices is susceptible to noise which, unless processed properly, can skew the results achieved by great margins. When testing the artifact developed with the testing data from the WESAD dataset, great results were obtained, achieving nearly 90% binary classification accuracy with SVM, Knn and DT. An experiment was then done following a similar procedure as used in the WESAD dataset in which 5 individuals were chosen to take part in the protocol, using a skin temperature sensor and ECG sensor. Readings were recorded in a baseline state and in a stressed state, stress was induced using the Colour Word Stroop Test. When this data was used for stress classification one could truly see how simple variables such as a change in device, individual and stressor can affect the results achieved.