Detection of Error-related Brain Oscillatory Activity using Machine Learning

Resting-state EEG (rsEEG) can reflect error-related activity formed during and after solving a cognitive task. We analyzed EEG data of 19 healthy volunteers who solved mental tasks (verbal - anagram, and spatial - assembling a figure) presented on a computer screen. The maximum allowed solving time was 40 s, but if a subject was ready with an answer sooner than this period, he/she pressed a button causing a task to disappear. In between tasks, participants observed a black screen for about 6 s. Applying two machine-learning algorithms (Perceptron without hidden layers, and k nearest neighbors) with spectral patterns of rsEEG as input vectors, we tested the accuracy of discriminating inter-stimulus rsEEG after correctly solved tasks from rsEEG after erroneously solved tasks, as well as rsEEG after completed tasks from rsEEG after uncompleted tasks. K nearest neighbors (KNN) classifier performed better than Perceptron in recognition of rsEEG after correct versus incorrect answers (75% of accuracy rate). Both algorithms showed robust accuracy rates (more than 71%) in the detection of rsEEG after completed versus uncompleted tasks. Our findings indicate that both erroneous answers and the incompleteness of the task are reflected in the rsEEG rhythmic patterns. The obtained results suggest that the neural activity elicited by goal-directed learning and error-monitoring is not only response-locked but may retain traces for several seconds after the response.