A fast algorithm has been developed, to classify eye movements and blinks, from the signals provided by eyewear-mounted sensors. The sensors (JINS Inc., Tokyo) measured electrooculography (EOG) by means of two electrodes mounted on the bridge of the user glasses, and acceleration of the head, at a constant rate of 100 Hz. An adaptive filter combining a low-pass filter, linear regression and Wiener filtering was used to remove motion artefacts from the EOG data streams. Batches of 20 seconds of data were decomposed by DB4 wavelets into 8 levels of detail. The root mean square of the signal obtained by reconstructing the 3rd-to-5th levels was used to identify and distinguish horizontal and vertical movements, and the value of the 8th level was used to classify the orientation of the movement (left vs. right, or up vs. down). The sensitivity of the algorithm to detect movements was greater than 90% in all directions, and a specificity between 70% and 90%, depending on the direction of the movement. The artefacts of head motion could be successfully removed at slow rotational velocities, compatible with normal visual tasks at rest. Quick movements without clear directional features were classified as blinks, without false detections (100% specificity) but low sensitivity (35%). This algorithm runs at high speed, with the possibility of providing real-time feedback of eye movements from normal eyewear.

{"references": ["Henderson et al. (2013). \"Predicting Cognitive State from Eye Movements\", PLOS ONE"]}