saccadr is a modular and extendable R package to extract (micro)saccades from gaze samples via an ensemble of methods approach.
Although there is an agreement about a general definition of a saccade, the more specific details are harder to agree upon. Therefore, there are numerous algorithms that extract saccades based on various heuristics, which differ in the assumptions about velocity, acceleration, etc. The package uses these methods to label individual samples and then applies a majority vote approach to identify saccades. The package includes three methods (see Implemented Methods vignette) but can be extended via custom methods (see Using custom methods vignette). It also uses a modular approach to compute velocity and acceleration from noisy samples (see Velocity computation vignette). Finally, you can obtain methods votes per gaze sample instead of saccades (see Using sample votes vignette).
extract_saccades() function uses several methods to label individual samples as belonging to a saccade, classifies a sample as a potential saccade if its proportion of votes exceeds a preset threshold, and then identifies saccades based on minimal saccade duration and minimal time between the saccades. For binocular data, 1) samples can be averaged before velocity computation, 2) votes can be merged so that function returns binocular saccades, or 3) saccades are extracted for each eye separately.
Currently, the library implements saccade detection using on the following saccade detection methods. When using this package, please cite both the package and individual methods.
method_ek: Engbert, R., & Kliegl, R. (2003). Microsaccades uncover the orientation of covert attention. Vision Research, 43(9), 1035–1045. https://doi.org/10.1016/S0042-6989(03)00084-1
method_om: Otero-Millan, J., Castro, J. L. A., Macknik, S. L., & Martinez-Conde, S. (2014). Unsupervised clustering method to detect microsaccades. Journal of Vision, 14(2), 18–18. https://doi.org/10.1167/14.2.18
method_nh: Nyström, M., & Holmqvist, K. (2010). An adaptive algorithm for fixation, saccade, and glissade detection in eye tracking data. Behavior Research Methods, 42(1), 188–204. https://doi.org/10.3758/BRM.42.1.188
For current stable version use
To install the development version from github
The main function is
extract_saccades(). Minimally, it takes x and y gaze samples, and sampling rate returning a table with extracted saccades. Note that the function expects that units of the gaze samples are degrees of visual angle, as some methods use physiologically plausible velocity and acceleration thresholds.
data("single_trial") saccades <- extract_saccades(single_trial$x, single_trial$y, sample_rate = 500)
When the recording spans multiple trials, you need to specify this via
trial parameter. This way velocity computation and saccade detection methods respect trial boundaries.
There are three ways in which binocular data can be treated based on the value of the
binocular = "merge"(default): sample votes are obtained from both eyes and for all methods and then averaged. This way only binocular saccades (i.e., eye movements with sufficient temporal overlap between eyes) are detected.
Eye = "Binocular"in saccade description.
binocular = "cyclopean": binocular data is converted to an average cyclopean image before voting and saccades detection.
Eye = "Cyclopean"in saccade description.
binocular = "monocular": saccades are extracted independently for each eye.
Eye = "Left"or
Eye = "Right"in saccade description.
data("single_trial_binocular") # binocular saccades only saccades_b <- saccadr::extract_saccades(single_trial_binocular[, c('xL', 'xR')], single_trial_binocular[, c('yL', 'yR')], sample_rate = 1000) # cyclopean saccades from binocular data saccades_c <- saccadr::extract_saccades(single_trial_binocular[, c('xL', 'xR')], single_trial_binocular[, c('yL', 'yR')], sample_rate = 1000, binocular = "cyclopean") # monocular saccades from binocular data saccades_m <- saccadr::extract_saccades(single_trial_binocular[, c('xL', 'xR')], single_trial_binocular[, c('yL', 'yR')], sample_rate = 1000, binocular = "monocular")
By default, all implemented methods are used for saccade detection but, if necessary, you can use their subset or even a single method. Note that you can also supply your own saccade detection function, please see Using custom methods vignette.
# Using a single method saccades <- extract_saccades(single_trial$x, single_trial$y, 500, methods = method_om) # Using two methods saccades <- extract_saccades(single_trial$x, single_trial$y, 500, methods = list(method_ek, method_om))
Parameters for individual methods are passed via the
options argument, which is a named list with
<parameter-name> = <value> pairs. You can find information on specific parameters and their default values in Implemented Methods vignette. Here is an example of modifying a velocity threshold, measured in units of standard deviation, for Engbert & Kliegl (2003) method. The default value is 6 but we can make it stricter
The voting threshold is the number of methods that must label a sample as a potential saccade. By default, all but one method must agree for a sample to be considered for a saccade (
vote_threshold = length(methods) - 1) but is 1, if only a single method was passed to the function. You can make voting more or less restrictive via
# A strict unanimous decision threshold saccades <- extract_saccades(single_trial$x, single_trial$y, 500, vote_threshold = 3) # A slacker criterion that at least one of the three methods must label sample as a saccade saccades <- extract_saccades(single_trial$x, single_trial$y, 500, vote_threshold = 1)
Because the gaze samples tend to be noisy, different methods use various approaches for computing velocity from noisy samples. Methods by Engbert & Kliegl (2003) and Otero-Millan et al. (2014) used the same approach based on averaging over multiple samples to compute velocity, whereas Nyström & Holmqvist (2010) compute a simple derivative and then filter it. By default, package uses the former approach (
velocity_function = diff_ek) but you can also use the latter (
velocity_function = diff_nh) or implement a custom method (see Velocity computation vignette). Acceleration is computed the same way but from velocity samples. Here is an example of using Nyström & Holmqvist (2010) velocity computation
saccades <- extract_saccades(single_trial$x, single_trial$y, 500, velocity_function = diff_nh)
Once the votes are in, saccades detection is based on their minimal duration (
minimal_duration_ms parameter, defaults to 12 ms) and minimal time between the saccades (
minimal_separation_ms, defaults to 12 ms).
# Only longish saccades are extracted saccades <- extract_saccades(single_trial$x, single_trial$y, 500, minimal_duration_ms = 20)
extract_saccades() function returns a table with following columns:
Eye“Monocular” for monocular inputs. “Cyclopean” for binocular data that was averaged before applying algorithms. “Binocular” for binocular data with votes averaged after applying algorithms. “Left” or “Right” for binocular data when eyes are processed independently.
OnsetSampleIndex of the first sample.
OffsetSampleIndex of the last sample.
OnsetOnset time relative to the trial start in milliseconds.
OffsetOffset time relative to the trial start in milliseconds.
DurationDuration in milliseconds.
DisplacementXHorizontal displacement measured from the first to the last sample.
DisplacementYVertical displacement measured from the first to the last sample.
DisplacementDisplacement magnitude measured from the first to the last sample.
DisplacementPhiDisplacement direction measured from the first to the last sample.
AmplitudeXHorizontal displacement measured from the leftmost to the rightmost sample.
AmplitudeYVertical displacement measured from the lowest to the uppermost sample.
AmplitudeDisplacement magnitude measured from the most extreme samples.
AmplitudePhiDisplacement direction measured from the most extreme samples.
AccelerationStartPeak acceleration before peak velocity was reached.
AccelerationStopPeak acceleration after peak velocity was reached.