Krishna Rathi
Vol. 14, Jul-Dec 2022
Abstract:
In any data preprocessing pipeline for physiological time series data, artefact detection and removal is an essential step, particularly when the data are obtained outside of controlled experimental circumstances. Given that these artefacts are frequently easily recognised with the naked eye, unsupervised machine learning methods seem like a viable alternative to manually labelled training datasets. Current techniques are frequently heuristic-based, non-generalizable, or designed for less artifact-prone, controlled experimental environments. In this work, we evaluate three such unsupervised learning techniques: K-nearest neighbour distance, isolation forests, and 1-class support vector machines. These algorithms are used to analyse electrodermal activity (EDA) data obtained during surgery in order to identify heavy cautery-related artefact. In order to provide inputs for the unsupervised learning techniques, we first defined 12 characteristics for every half-second frame. We contrasted the top-performing unsupervised learning technique for each subject with four other current approaches for the elimination of EDA artefacts. The only learning strategy that was successful in completely eliminating the artefact for each of the six subjects was unsupervised learning. In complex circumstances, this technique can be readily extended to different modalities of physiological data.
DOI: http://doi.org/10.37648/ijrmst.v14i01.022