Artefact Detection in Chronically Recorded Local Field Potentials: an Explainable Machine Learning-based Approach

The role of machine learning in neuroscience has been increasing through the years, in aiding diagnosis, biomarker discovery, signal analysis, and other applications. However, the lack of information of the decision-making of the models restricts their use and adoption by the community. In the process of neuronal signal acquisition, other electrical signals can distort the recording, for which a review process is necessary. Machine learning can aid by automatically detecting affected segments, speeding up the review process. However, as the ground-truth labelling is done manually or via a threshold, researchers must be able to identify the causes of false negatives and positives. This paper looks into explainable machine learning for artefact detection in invasively recorded neural signals through the use of different classifiers, trained with a feature subset produced by the combination of feature selection algorithms to reduce the dimensionality by two orders of magnitude. Our results show that the bagging decision tree model is best suited for creating a generalised model that is capable of classifying artefactual patterns in a multi-state dataset, which achieves an accuracy of 96.1%. Lastly, the predictor importance, Shapely values, and reduced feature space visualisation are used to gain insight into the model.