Towards Interpretable Machine Learning in EEG Analysis

In this paper a machine learning model for automatic detection of abnormalities in electroencephalography (EEG) is dissected into parts, so that the influence of each part on the classification accuracy score can be examined. The most successful setup of several shallow artificial neural networks aggregated via voting results in accuracy of 81%. Stepwise simplification of the model shows the expected decrease in accuracy, but a naive model with thresholding of a single extracted feature (relative wavelet energy) is still able to achieve 75%, which remains strongly above the random guess baseline of 54%. These results suggest the feasibility of building a simple classification model ensuring accuracy scores close to the state-of-the-art research but remaining fully interpretable.