Automatic artifact recognition and correction for electrodermal activity based on LSTM-CNN models

Abstract

Researchers increasingly use electrodermal activity (EDA) to assess emotional states, developing novel applications that include disorder recognition, adaptive therapy, and mental health monitoring systems. However, movement can produce major artifacts that affect EDA signals, especially in uncontrolled environments where users can freely walk and move their hands. This work develops a fully automatic pipeline for recognizing and correcting motion EDA artifacts, exploring the suitability of long short-term memory (LSTM) and convolutional neural networks (CNN). First, we constructed the EDABE dataset, collecting 74h EDA signals from 43 subjects collected during an immersive virtual reality task and manually corrected by two experts to provide a ground truth. The LSTM-1D CNN model produces the best performance recognizing 72% of artifacts with 88% accuracy, outperforming two state-of-the-art methods in sensitivity, AUC and kappa, in the test set. Subsequently, we developed a polynomial regression model to correct the detected artifacts automatically. Evaluation of the complete pipeline demonstrates that the automatically and manually corrected signals do not present differences in the phasic components, supporting their use in place of expert manual correction. In addition, the EDABE dataset represents the first public benchmark to compare the performance of EDA correction models. This work provides a pipeline to automatically correct EDA artifacts that can be used in uncontrolled conditions. This tool will allow to development of intelligent devices that recognize human emotional states without human intervention.