One-dimensional convolutional neural networks for low/high arousal classification from electrodermal activity

The rapid identification of arousal is of great interest in various applications such as health care for the elderly, athletes, drivers and students, among others. Therefore, advanced methods are needed to classify the level of activation autonomously. In this paper, three architectures based on one...

Full description

Bibliographic Details
Authors: Sánchez Reolid, Roberto, López de la Rosa, Francisco, López Bonal, María Teresa, Fernández Caballero, Antonio
Format: article
Publication Date:2021
Country:España
Institution:Universidad de Castilla-La Mancha
Repository:RUIdeRA. Repositorio Institucional de la UCLM
OAI Identifier:oai:ruidera.uclm.es:10578/32233
Online Access:http://hdl.handle.net/10578/32233
Access Level:Open access
Keyword:One-dimensional convolutional neural networks
Electrodermal activity
Arousal classification
Description
Summary:The rapid identification of arousal is of great interest in various applications such as health care for the elderly, athletes, drivers and students, among others. Therefore, advanced methods are needed to classify the level of activation autonomously. In this paper, three architectures based on one-dimensional convolutional networks (1D-CNN) using electrodermal activity as physiological input are proposed. These have been designed for low and high arousal discrimination, elicited through video clips. The first architecture, based on a purely convolutional architecture, has yielded an F1-score of 81.95%. Two other architectures (hybrid), based on 1D-CNNLSTM (long short-term memory) and 1D-CNN-BiLSTM (bidirectional LSTM), have outperformed the first one, obtaining 88.95% and 91.02% F1-score, respectively. Furthermore, a comparison of these methods has been performed with widely used network architectures such as AlexNet, GoogLeNet, VGG16, VGG19 and ResNet-50, which have obtained F1-scores 82.09%, 83.14%, 82.69%, 83.95% and 82.00%, respectively. Our architectures offer good performance with shorter training time compared to pretrained architectures