The effect of artifact rejection on the performance of a convolutional neural network based algorithm for binary EEG data classification

Evaluate the effect of artifact rejection on the performance of a Convolutional Neural Network (CNN) based algorithm for classification of abnormal and normal electroencephalography (EEG) data. We developed an automated CNN-based clean versus artifact classification algorithm and applied this to abnormal and normal EEG data. Additionally, algorithms for abnormal versus normal classification were developed with and without artifact rejection beforehand. For each algorithm, five CNNs were trained using 5-fold cross-validation and the majority vote of these was used for classification of test data. We compared bootstrap test accuracies and the number of training epochs required between the scenario in which artifacts were rejected and the scenario in which they were not. The clean versus artifact classification algorithm had a test accuracy of 85% (recall:89%, precision:82%). Bootstrap test accuracies of the abnormal versus normal classification algorithms were similar with and without artifact rejection beforehand (mean [95% CI]:84% [77%-91%] and 84% [76%-91%], respectively). For the five cross-validation runs, respectively −8%, 22%, 31%, −94% and 26% less training epochs were required when artifacts were rejected beforehand. In our study, EEG artifact rejection did not improve CNN-based abnormal versus normal classification performance, but did slightly speed up training. In our study, artifact rejection was not required. The possibility to omit this step lowers the threshold for applying CNNs to EEG data and using these in clinical practice. Future studies should confirm whether this step could also be omitted in case other artificial intelligence techniques are applied to different datasets.