Selection of the Most Relevant Physiological Features for Classifying Emotion

Abstract

Emotion classification using physiological sensors can be used for a wide range of applications in the areas of wellness, medicine, entertainment, sport, learning, advertising, human computer interfaces among other. The associated technologies need to be improved in order to be really efficient in real life applications. The sensors should be less obtrusive as possible, and the algorithms that estimate emotion most accurate as possible. The knowledge of the most relevant features for classifying emotions is crucial for both objectives by allowing to select a reduced set of sensors and by optimizing classification algorithms performances. In this paper we analyze the relevance of several features extracted from peripheral physiological sensors by using two databases freely available to the research community. The use of two separate databases allow to analyze if some features are relevant independently from the way the emotions are elicited and from the material used for experiments. We find features extracted from galvanic skin response (GSR) to be relevant for both databases. Eye closing rate and variance of zygomatic electromyography (EMG), only available in one database, are relevant for respectively arousal and valence. The hearth rate variability (HRV) is relevant but only for one of the databases, using an electrocardiogram (ECG) whereas for the other database photopletismography (PPG) was used. Only with a few set of well selected feature and sensors we reach classification performances similar to literature classifiers using more features.

Using Deep Convolutional Neural Network for Emotion Detection on a Physiological Signals Dataset (AMIGOS)

“…The affect detection requires an adequate features extraction of the signals, which correlate with the emotional states recorded by the participants in the self-assessment. That is, the relationship between features and emotions determines the physiological reaction [43] and is taken as input to the predictor. Parametric measurements of the ECG signals in the time domain quantify the variability of interbeat intervals (IBI) measurements successive.…”