Jeudi 27 avril 2023

StressLink: Adaptive Forecast System with Time-Series Neural Networks
Jad Bakieh
Étudiant au doctorat sous la supervision des professeurs Marc Parizeau, Philippe Giguère et Cem Subakan.

Heure: 13h30
Local: PLT-1120

Résumé: Time is an essential element in machine learning systems, enabling the identification of patterns and trends within datasets. Moreover, the time-dependant variables can impact system operations incrementally, with both linear and non-linear effects based on the temporal dependencies between the variables and system functions. This principle applies across a broad range of fields, including mechanical, electrical, and biochemical systems. Incorporating time dimension in deep learning architectures is crucial to develop precise and predictive models that account for the dynamic changes in the system. In this way, wearable AI technology can collect real-time data about electrophysiological changes, enabling the analysis of temporal patterns and fluctuations in variables that impact human health. This research project focuses on the development of a wearable platform that can offer prognostic intelligence about one’s physiological and mental health by examining the impact of time-varying vital signs on future health outcomes. This study aims to develop a deep learning system that can be deployed on smart wearable devices, specifically smart wristbands, to continuously classify mental states and provide probabilistic forecasting of symptoms triggered by mental stress. To achieve stress level classification, time-series biosignals will be extracted from the wrist, namely, Electrodermal Activity (EDA), Photoplethysmography (PPG), and Skin Temperature (SKT) due to their collective ability in discriminating human stress levels. The prediction system will add a time dimension to enable a temporal classification system to learn hidden correlations between accumulative features and the probability of the target forecast.

Note: La présentation sera donnée en anglais et ne sera pas diffusée par visioconférence.