Automatic EEG Blink Detection Using Dynamic Time Warping Score Clustering

The electroencephalograph (EEG) is a powerful tool, involving multiple electrodes placed on the scalp, with the intention of measuring brain activity through the scalp. One significant application for EEG is to analyze the mental state of a subject. One of the challenges involved in using the EEG for identifying mental state in practical settings is ocular artifacts e.g. eye blinking. Eye blinks cause high amplitude noise in electroencephalograms (EEGs), the noise from these blinks can cause interference in several very important frequency bands and confuse predictive modeling e.g. introduce false positives. Prior works have employed independent component analysis (ICA) to decompose the noisy EEG signals into constituting sources and identify the eye blink sources. However, ICA requires off-line signal processing and is not suitable for online applications. More recently, time domain autoregressive features were used to model eye blink related segments in the recorded EEG data. While the autoregressive method showed high identification accuracy in isolated short trials, the goal of this work is to create a more advanced system capable of identifying and filtering blink noise in continuous trials during long and complex tasks. The proposed method detailed in this paper conducts automatic detection of eye blink noise using dynamic time warping (DTW) score clustering during wearable EEG-based cognitive workload assessment tests. The proposed eye blink detection system only uses EEG data for training and identification and does not require electrooculogram (EOG) data, which is particularly important for wearable systems. Our experimental results demonstrated the effectiveness of the proposed blink detection methodology by achieving 96.42% average accuracy of blink detection in the recorded EEG dynamics during a continuous workload assessment task.