Research Article
Stress Detection with Deep Learning Approaches Using Physiological Signals
@INPROCEEDINGS{10.1007/978-3-030-69963-5_7, author={Fabrizio Albertetti and Alena Simalastar and A\~{n}cha Rizzotti-Kaddouri}, title={Stress Detection with Deep Learning Approaches Using Physiological Signals}, proceedings={IoT Technologies for HealthCare. 7th EAI International Conference, HealthyIoT 2020, Viana do Castelo, Portugal, December 3, 2020, Proceedings}, proceedings_a={HEALTHYIOT}, year={2021}, month={7}, keywords={Physiological monitoring Stress prediction Sympathetic and parasympathetic activation Affective computing Telemonitoring Self-management systems}, doi={10.1007/978-3-030-69963-5_7} }- Fabrizio Albertetti
Alena Simalastar
Aïcha Rizzotti-Kaddouri
Year: 2021
Stress Detection with Deep Learning Approaches Using Physiological Signals
HEALTHYIOT
Springer
DOI: 10.1007/978-3-030-69963-5_7
Abstract
The problem of stress detection and classification has attracted a lot of attention in the past decade. It has been tackled with mainly two different approaches, where signals were either collected in ambulatory settings, which can be limited to the period of presence in the hospital, or in continuous mode in the field. A sensor-based continuous measurement of stress in daily life has a potential to increase awareness of patterns of stress occurrence. In this work, we first present a data-flow infrastructure suitable for two types of studies that conforms with the data protection requirements of the ethics committee monitoring the research on humans. The detection and binary classification of stress events is compared with three different machine learning models based on the features (meta-data) extracted from physiological signals acquired in laboratory conditions and ground-truth stress level information provided by the subjects themselves via questionnaires associated with these features. The main signals considered in current classification are electro-dermal activity (EDA) and blood volume pulse (BVP) signals. Different models are compared and the best configuration yields an score of 0.71 (random baseline: 0.48). The importance on prediction of phasic and tonic EDA components is also investigated. Our results also pave the way for further work on this topic with both machine learning approaches and signal processing directions.
