A Data-Driven Detection System for Predicting Stress Levels from Autonomic Signals

Abstract

This paper introduces and details a detection system for continuous monitoring of psychological stress. The formulation of classes relating to varying levels of stress intensity is described. This is necessary for determining the therapy needed to alleviate the associated effects, particularly in population groups suffering from chronic and mental illnesses such as diabetes and depression. The data-driven detection system mainly comprises kernel principal component analysis for dimensionality reduction, and nearest neighbour classifier for supervised learning to determine the associated stress intensity. We evaluate the generalised stress detection system using a 3-fold cross validation and a test set comprising an independent subject. We obtain a 0.66 F1-score with a precision of 0.70 and a recall of 0.67 over 4 classes of stress: no stress, low stress, moderate stress, and high stress.