Date and Time:January 22, 2019 (Tuesday), 3.00 PM
Speaker:Ms. Anusha A S (EE14D016)
DC Members :
Dr. K. Sridharan
Dr. MohanasankarSivaprakasam (Guide)
Dr. V. Jagadeesh Kumar
Recent years have seen an increased interest in leveraging personal and ubiquitous sensing technologies towards mental health care, thereby, allowing unobtrusive, automatic and continuous monitoring of mental well-being. Our study focuses on using electrodermal activity (EDA) as a physiological biomarker for the detection of multiple levels of stress. EDA is an all-encompassing term used to indicate electrical properties of the skin. Unlike other target organs of the human body, which are connected to both the sympathetic and parasympathetic nervous system, skin with sweat glands and blood vessels are exclusively innervated by the sympathetic branch. This makes EDA an ideal and unperturbed measure of sympathetic activation, and therefore stress.
Through our study,we propose and validate the utility of wrist EDA for monitoring multilevel stress in a real-life setting. Accordingly,EDA data was collected from subjects who subsequently underwent different surgical procedures at the Sri Ramakrishna Hospital, Coimbatore, India. One of the major challenges during ambulatoryphysiological monitoring is that the data thus collected are prone to corruption due tomotion artifacts. To overcome this, a supervised machine learning algorithm that detects motion artifacts in the recorded EDA data was developed. The algorithm yielded an accuracy of 97.83% on a new user dataset. The clean EDA data were further analyzed to determine low, moderate and high levels of stress. A localized supervised learning scheme, based on the adaptive partitioning of the dataset was adopted for stress detection. Consequently, the inter-individual variability in the EDA due to person-specific factors like the sweat gland density and skin thickness, which may lead to erroneous classification, could be eliminated. The scheme yielded a classification accuracy of 85.06% on a new user dataset and proved to be more effective than the general supervised classification model.
We also attempted to extend the utility of EDA for monitoring sleep quality as stress is known to have an immediate and negative impact on sleep. Accordingly, EDA data were recorded overnight from subjects in their natural sleep environments. A non-crossing ordinal classifier based on the random forest was developed to automatically detect wake, light, rapid eye movement (REM), and deepstages of sleep. The scheme yielded an accuracyof 97.95% and exhibited good generalizability on a new user dataset yielding an average accuracy of 90.25%. The duration of wake, light, deep and REM stages was estimated as a percentage of total sleep duration and was validated against a reference device. The sleep estimates showed a positive correlation with the reference and the Bland-Altman analysis done on estimates indicated similarity with minimal bias.
All are cordially invited