Fig 1.
(A,C) Two samples of 22-minute recordings of raw in-ear EOG. (B,D) Detailed view of the EOG waveform on a scale of 20 seconds. Straight solid lines denote the position of the visual targets. Small rectangles in the panels A and C indicate where the detailed views were taken from. The scale of y-axis in the zoomed in-panels B and D was kept fixed to 2 mV, x-axis was fixed to 20 s, and it shows actual time during the experiment.
Fig 2.
(A) Schematics of the experimental procedure. (B) The participant was equipped with the mobile bio-amplifier (blue box) attached to the headband and the video-based eye tracker. (C) Detail of the in-ear electrodes.
Fig 3.
Model scheme.
Fig 4.
Step response of the IIR HP Butterworth filters as a function of half-power frequency.
Fig 5.
Detail of the saccade detection algorithm over 12 seconds of the experiment.
An example of EOG recording (A) and (B)-(D) estimated parameters Sx, So, and Sg. (B) Sx saccade magnitude parameter. (C) So—saccade time shift parameter. (D) Sg—saccade gain parameter. Circles with lines show detected saccades and their magnitude (right y-axis). The magnitude of a saccade was obtained by multiplying measured voltage Sx with Ci parameter obtained in the training period.
Fig 6.
Limits of the saccade integration scheme.
The figure shows the performance of the saccade integration scheme for the current experiment assuming the perfect saccade detection. The r2 values of the estimated and actual eye gaze position (y-axis) were computed for different values of Nm and Nx,min. The Nm is on the x-axis, Nx,min is represented by different types of line. The lines show across-subject means; error bars show standard errors of the mean (SEM).
Fig 7.
The change in the in-ear EOG as a function of the change of target angle.
The magnitude of the change in the visual target is shown on the x-axis. The change was defined as the difference between the medians of the pre-transition and post-transition periods. This was computed for each trial and each participant. Data of individual participants are shown using thin grey lines, the across-subject mean value and SEM error bars are shown with the dashed line.
Fig 8.
Performance of SACCINT during the training period.
The figure shows across-subject mean (±SEM) performance as a function of Nx,min parameter (top row) and Nm parameter (bottom row). (A, E) The r2 values of the actual and predicted eye angles, greater r2 indicates better performance. (B, F) the standard deviation of error of the linear model. (C, G) The gain of the linear model. (D) Detection performance in F values (fHP = 0.03 Hz).
Fig 9.
Data of individual participants are shown by ‘x’ symbols. The x-axis shows the performance of the EOGHP. Y-axis shows the performance of the SACCINT. Each panel shows data for single cut-off frequency of the HP filter. Solid lines show predictions of the performance of SACCINT obtained from fitting a linear function on the individual data. The dotted line shows the theoretical maximum of the saccade integration for the current set of parameters. The saccades were obtained from the video based eye tracker were delayed by the time of half of the length of the temporal window for the purpose of this comparison and they were integrated with the same parameters as the EOG based data; Nm = 9° and Nx,min = 8°.
Fig 10.
Across-subject performance of SACCINT and EOGHP.
(A) The left panel shows across-subject r2 values, separately for SACCINT (circles), EOGHP (squares), and the “ideal” performance for the current set of parameters obtained from the ground truth (triangle) using the same parameters. The second and third columns show the data without HP filtering. (B) The right panel shows across-subject RMS error using the same symbols.
Fig 11.
Saccade detection evaluation for one participant.
(A) Scatter plot of the computed versus the actual saccade magnitudes. The solid line with open triangles shows the linear regression with the constant term fixed to zero. (B) Histogram of missed saccades, the white bars show misses smaller than Nx,min (8°), the black bars show misses larger than Nx,min. (C) Histogram of false alarms. White bars in the histogram show the data of saccades with magnitude smaller than Nx,min (8°), black bars show the rest of the dataset. The numbers inside panels indicate a total number of points in each histogram. Subscript 1 refers to the black bars, and subscript 2 refers to the white bars.
Fig 12.
Temporal distribution of missed and falsely alarmed saccades of two example participants.