Smart K Nearest Neighbor Outlier Detection for Electroencephalogram signal

Abstract Electroencephalogram (EEG) data suffer from artifacts such as poor imagination or loss of concen-tration during the recognition process. The negative impact of artifacts on the quality of informationand EEG signal analysis reduces the Quality of Service "QoS" and Quality of Information "QoI" ine-health applications. This negative impact can be avoided by identifying outliers with anomaly detec-tion algorithms. Aberrant values’ identification using Euclidean distance for the K Nearest Neighbor(KNN) process between the recorded EEG values separates the time series data recorded by a personinto neural activity and artifact. The algorithms using KNN often require knowledge of the data char-acteristics to configure one or more parameters (such as the number of neighbors K and distance D). However, our proposed solution does not require the initial setting of KNN process or any additional parameters. We propose a Smart KNN Outlier Detector (SKOD) which is an unsupervised non-parametric algo-rithm. We evaluated SKOD using various combinations of real EEG data of 140 trials with three channels of the benchmark Brain-Computer Interface "BCI competition II". We tested the perfor-mance of our solution on EEG data as provided by the studied patient (the subject) with the inclusionof different numbers of outliers. Our proposed detector achieves more than 60% of sensitivity andspecificity for detecting abnormal values with outlier detection close to 100%.