Unsupervised Machine Learning Methods for Artifact Removal in Electrodermal Activity.

Area of Science:

• Physiological data analysis
• Machine learning in healthcare
• Biomedical signal processing

Background:

• Artifacts in physiological time series data complicate analysis, especially in non-controlled clinical environments.
• Existing artifact removal methods are often heuristic, lack generalizability, or are designed for less artifact-prone data.
• Unsupervised learning offers a potential solution by avoiding the need for manually labeled datasets.

Purpose of the Study:

• To evaluate unsupervised machine learning algorithms for artifact detection and removal in electrodermal activity (EDA) data.
• To assess the efficacy of isolation forests, 1-class SVM, and KNN distance in removing surgical cautery artifacts.
• To compare unsupervised methods against existing artifact removal techniques.

Main Methods:

• Twelve features were extracted from half-second windows of EDA data.
• Isolation forests, 1-class SVM, and KNN distance algorithms were applied to identify and remove artifacts.
• The best performing unsupervised method was compared with four other established artifact removal techniques for each subject.

Main Results:

• Unsupervised learning methods demonstrated superior performance in artifact removal compared to existing methods.
• The chosen unsupervised learning approach was the only method successful in fully removing cautery-related artifacts across all six subjects.
• The developed method shows promise for application to other physiological data modalities.

Conclusions:

• Unsupervised machine learning provides a highly effective and generalizable approach for artifact removal in physiological data.
• This technique is particularly valuable for processing data collected in complex clinical settings, enhancing data utility.
• Robust artifact detection enables the use of diverse physiological signals for improved diagnostic and therapeutic decisions.