Abstract
Background: Extracellular microelectrode recording (MER) is a prominent technique for studies of extra cellular single-unit neuronal activity. In order to achieve robust results in more complex analysis pipelines, it is necessary to have high quality input data with a low amount of artifacts.
We show that noise (mainly electromagnetic interference and motion artifacts) may affect more than 25% of the recording length in a clinical MER database. New method: We present several methods for automatic detection of noise in MER signals, based on (i) unsupervised detection of stationary segments, (ii) large peaks in the power spectral density, and (iii) a classifier based on multiple time- and frequency-domain features.
We evaluate the proposed methods on a manually annotated database of 5735 ten-second MER signals from 58 Parkinson's disease patients. Comparison with existing methods: The existing methods for artifact detection in single-channel MER that have been rigorously tested, are based on unsupervised change-point detection.
We show on an extensive real MER database that the presented techniques are better suited for the task of artifact identification and achieve much better results. Results: The best-performing classifiers (bagging and decision tree) achieved artifact classification accuracy of up to 89% on an unseen test set and outperformed the unsupervised techniques by 5-10%.
This was close to the level of agreement among raters using manual annotation (93.5%). Conclusion: We conclude that the proposed methods are suitable for automatic MER denoising and may help in the efficient elimination of undesirable signal artifacts.