EDA Signal Processing and Affective State Detection with the WESAD Dataset


  • Katarina Lodrant University of Ljubljana
  • Andrej Košir University of Ljubljana



Wrist-worn devices for physiological data acquisition offer a convenient solution for continuous monitoring with minimal disruption to the user. However, they often operate at low sampling frequencies, which could potentially compromise the accuracy of data analysis. The WESAD (WEarable Stress and Affect Detection) dataset [1], capturing data from 15 subjects using wrist-worn (4 Hz sampling) and chest-worn (700 Hz sampling) devices, serves as a valuable resource to study such impacts. Notably, it includes simultaneous measurement of electrodermal activity (EDA) from both the wrist and chest devices, associated with different affective states—baseline, stress, and amusement. EDA has been shown to be a valuable tool for detecting stress by measuring changes in skin conductance linked to the sympathetic nervous system's response to emotional arousal [2].

This study aims to (1) develop a framework for EDA signal processing and visualization, (2) analyze the impact of sampling rate on data integrity, and (3) use the processed data to train a machine learning (ML) classifier for affective states.


Utilizing the Python-based neurokit2 library [3], this research processed the raw EDA signals, focusing on their decomposition into tonic and phasic components. The study involved downsampling the high-frequency EDA signals to various lower frequencies and comparing the fidelity of the derived signal components. Furthermore, we segmented the signal into windows, from which we extracted features such as the mean, standard deviation, and peak count of the signal components. These features were then utilized to train a ML classifier for affective states. This approach allows for a direct comparison of the effects of different sampling rates on signal quality and subsequent ML performance.


Preliminary findings show substantial discrepancies in EDA data processed at 4 Hz, especially in peak detection accuracy, compared to higher frequencies. The study also suggests that a 700 Hz sampling rate is excessive; reducing the rate by half minimizes data loss while lowering computational demands. Variations in signal processing results across subjects indicate that personalized calibration might be necessary for optimal data use in practical applications.

Future Work

Future research will focus on implementing the affective state classifier, experimenting with various ML models. We will assess model performance across different sampling devices (chest vs. wrist) and problem complexities (binary vs. three-class), employing leave-one-subject-out cross-validation to validate the classifier's ability to generalize across unseen data, thus enhancing its real-world applicability.


This study establishes a framework for processing and visualizing raw EDA signals from wearable devices. It highlights the challenges of using low-frequency data for accurate stress and affect detection and advocates for a balanced approach in choosing appropriate sampling rates to ensure data reliability and operational efficiency in wearable technology applications.


[1] P. Schmidt, et. al., "Introducing WESAD, a multimodal dataset for wearable stress and affect detection," in Proc. 20th ACM Int. Conf. Multimodal Interaction (ICMI '18), New York, NY, USA, 2018, pp. 400-408. DOI: 10.1145/3242969.3242985.

[2] W. Boucsein, Electrodermal Activity, 2nd ed., New York, NY, USA: Springer Science + Business Media, 2012. DOI: 10.1007/978-1-4614-1126-0.

[3] D. Makowski, et. al., "NeuroKit2: A Python toolbox for neurophysiological signal processing," Behavior Research Methods, vol. 53, no. 4, pp. 1689-1696, 2021. DOI: 10.3758/s13428-020-01516-y.