Exploring Wrist Pulse Signals using Empirical Mode Decomposition: Emotions

Abstract

Emotion recognition is attracting considerable interest among the research community. In this work, Empirical Mode Decomposition has been implemented to derive both statistical and nonlinear features from Wrist Pulse Signal to classifying emotions namely anxiety and boredom. Wrist Pulse signals were extracted from 24 subjects using TETRIS game as a stimulus using Fission and Fusion approach. The acquired signals were pre-processed to remove unwanted noise and artefacts present within the signal. In addition, various classifiers namely Naiive Byes, Support Vector Machine, K-Nearest Neighbour, Logistic Regression, Linear Discriminant Analysis, Quadratic Discriminant Analysis were considered. Results from these classifiers indicate that both Logistic Regression and Quadratic Discriminant Analysis gave an indistinguishable accuracy of 99.71% (fission) and 77.08% (fusion) for anxiety state. Moreover, for boredom state, the highest classification accuracy was 66.67 % for Naiive Bayes using fission and 64.58% for fusion. Results highlight the impact of empirical mode decomposition with hilbert transform for the recognition of emotion from wrist pulse signals.