Employability Of Unsupervised Machine Learning Techniques For Elimination Of Electrodermal Activity Artefacts During And Post Surgeries

Abstract

Detecting and removing artefacts is crucial in data preprocessing pipelines for physiological time series data, mainly when collected outside controlled experimental settings. The fact that such artefacts are often easily identifiable visually suggests that unsupervised machine learning algorithms could be practical without requiring manually labelled training datasets. Current methods are often heuristic-based, not generalizable, or designed for controlled experimental settings with fewer artefacts. In this study, we evaluate the effectiveness of three unsupervised learning algorithms—Isolation Forests, One-Class Support Vector Machine, and K-Nearest Neighbor Distance—in removing heavy cautery-related artefacts from electrodermal activity (EDA) data collected during surgeries involving six subjects. We defined 12 features for each half-second window as inputs to the unsupervised learning methods. We compared each subject's best-performing unsupervised learning method to four existing EDA artefact removal methods. The unsupervised learning method was the only approach that removed the artefacts across all six subjects. This approach can be easily extended to other types of physiological data in complex settings.