glassbox: The opinionated "glass box" 'eyeris' pipeline

In eyeris: Flexible, Extensible, & Reproducible Pupillometry Preprocessing

The opinionated "glass box" eyeris pipeline

Description

This glassbox function (in contrast to a "black box" function where you run it and get a result but have no (or little) idea as to how you got from input to output) has a few primary benefits over calling each exported function from eyeris separately.

Usage

glassbox(
  file,
  interactive_preview = FALSE,
  preview_n = 3,
  preview_duration = 5,
  preview_window = NULL,
  verbose = TRUE,
  ...,
  confirm = deprecated(),
  num_previews = deprecated(),
  detrend_data = deprecated(),
  skip_detransient = deprecated()
)

Arguments

file	Either an SR Research EyeLink .asc file generated by the official EyeLink edf2asc command, or a pre-constructed eyeris object. When an .asc path is supplied, the load_asc step reads and parses it. When a pre-loaded eyeris object is supplied instead (for example, the output of load_generic() for non-EyeLink trackers, or an existing load_asc() object), the load step is skipped and the remaining pipeline runs on the object as-is – enabling load_generic(...) |> glassbox()
interactive_preview	A flag to indicate whether to run the glassbox pipeline autonomously all the way through (set to FALSE by default), or to interactively provide a visualization after each pipeline step, where you must also indicate "(y)es" or "(n)o" to either proceed or cancel the current glassbox pipeline operation (set to TRUE)
preview_n	Number of random example "epochs" to generate for previewing the effect of each preprocessing step on the pupil time series
preview_duration	Time in seconds of each randomly selected preview
preview_window	The start and stop raw timestamps used to subset the preprocessed data from each step of the eyeris workflow for visualization. Defaults to NULL, meaning random epochs as defined by preview_n and preview_duration will be plotted. To override the random epochs, set preview_window here to a vector with relative start and stop times (in seconds), for example – c(5,6) – to indicate the raw data from 5-6 secs on data that were recorded at 1000 Hz). Note, the start/stop time values indicated here are in seconds because eyeris automatically computes the indices for the supplied range of seconds using the ⁠$info$sample.rate⁠ metadata in the eyeris S3 class object
verbose	A logical flag to indicate whether to print status messages to the console. Defaults to TRUE. Set to FALSE to suppress messages about the current processing step and run silently
...	Additional arguments to override the default, prescribed settings
confirm	(Deprecated) Use interactive_preview instead
num_previews	(Deprecated) Use preview_n instead
detrend_data	(Deprecated) A flag to indicate whether to run the detrend step (set to FALSE by default). Detrending your pupil time series can have unintended consequences; we thus recommend that users understand the implications of detrending – in addition to whether detrending is appropriate for the research design and question(s) – before using this function
skip_detransient	(Deprecated) A flag to indicate whether to skip the detransient step (set to FALSE by default). In most cases, this should remain FALSE. For a more detailed description about likely edge cases that would prompt you to set this to TRUE, see the docs for detransient()

Details

First, this glassbox function provides a highly opinionated prescription of steps and starting parameters we believe any pupillometry researcher should use as their defaults when preprocessing pupillometry data.

Second, and not mutually exclusive from the first point, using this function should ideally reduce the probability of accidental mishaps when "reimplementing" the steps from the preprocessing pipeline both within and across projects. We hope to streamline the process in such a way that you could collect a pupillometry dataset and within a few minutes assess the quality of those data while simultaneously running a full preprocessing pipeline in 1-ish line of code!

Third, glassbox provides an "interactive" framework where you can evaluate the consequences of the parameters within each step on your data in real time, facilitating a fairly easy-to-use workflow for parameter optimization on your particular dataset. This process essentially takes each of the opinionated steps and provides a pre-/post-plot of the time series data for each step so you can adjust parameters and re-run the pipeline until you are satisfied with the choices of your parameters and their consequences on your pupil time series data.

Value

Preprocessed pupil data contained within an object of class eyeris

What glassbox() does, step by step

In plain language, glassbox() runs the following eyeris steps in order on your pupil time series. Steps marked (default: on) run automatically; steps marked (default: off) are skipped unless you explicitly enable them. Most preprocessing steps can be turned off by passing ⁠<step> = FALSE⁠ (except load_asc, which always runs when file is an .asc path). Steps that accept parameters can be customized by passing ⁠<step> = list(...)⁠ with values you want to override (for example, deblink = list(extend = 40)).

1. Load the data (load_asc, default: on) – Reads and parses the EyeLink .asc file into an eyeris object, automatically splitting the recording into blocks and, for binocular recordings, handling each eye separately. See load_asc(). This step is skipped when a pre-loaded eyeris object is passed as file (see the file parameter).

2. Resample onto a uniform grid (resample, default: on) – Places each block on the expected uniform sampling grid. For hardware that drops samples (instead of zero-filling) when pupil data is missing, this interpolates local sub-period timing jitter and inserts NA rows at the dropped timestamps, so the rate-dependent steps that follow stay valid. A guaranteed no-op for already-uniform data (e.g., EyeLink). See resample().

3. Remove blinks (deblink, default: on) – Replaces the missing data around blinks with NAs, extending each gap by 50 ms on either side so that the rapid dips and spikes that surround a blink are removed too. See deblink().

4. Remove transient artifacts (detransient, default: on) – Rejects pupil samples that change faster than is physiologically plausible, using a speed-based median absolute deviation (MAD) threshold. See detransient().

5. Interpolate missing samples (interpolate, default: on) – Fills the NA gaps left by the resample, deblink, and detransient steps using linear interpolation, producing a continuous, gap-free time series. See interpolate().

6. Smooth the signal (lpfilt, default: on) – Applies a low-pass filter (default 4 Hz passband) to remove high-frequency noise while preserving the slower pupil dynamics of interest. See lpfilt().

7. Downsample (downsample, default: off) – Optionally lowers the sampling rate using an anti-aliasing filter, which preserves the temporal dynamics of the signal. Cannot be combined with bin. See downsample().

8. Bin (bin, default: off) – Optionally lowers the sampling rate by averaging samples within equal-width time bins. Cannot be combined with downsample. See bin().

9. Detrend (detrend, default: off) – Optionally fits a model of pupil ~ time and returns the residuals (along with the fitted trend) to remove slow drift. By default (detrend = TRUE) a straight-line (method = "linear") trend is removed; pass detrend = list(method = "spline", spline_df = 5) to instead remove a smooth, potentially nonlinear trend via a natural cubic spline of time. Use with care – see detrend() for when this is appropriate.

10. Z-score (zscore, default: on) – Rescales the pupil time series to a mean of 0 and a standard deviation of 1, making values comparable across participants and recordings. See zscore().

After preprocessing, glassbox() calls summarize_confounds() to compute per-step confound metrics (e.g., missingness and gaze statistics) and store them in ⁠$confounds⁠.

Crucially, each step adds a new column to the time series rather than overwriting the previous one, so every intermediate stage is preserved inside the returned eyeris object. This is what makes the pipeline a "glass box": you can inspect, plot, and compare the data before and after each transformation (for example, plot(output, steps = c(1, 5))).

See Also

lifecycle::deprecate_warn()

Examples

demo_data <- eyelink_asc_demo_dataset()

# (1) examples using the default prescribed parameters and pipeline recipe

## (a) run an automated pipeline with no real-time inspection of parameters
output <- eyeris::glassbox(demo_data)

start_time <- min(output$timeseries$block_1$time_secs)
end_time <- max(output$timeseries$block_1$time_secs)

# by default, verbose = TRUE. To suppress messages, set verbose = FALSE.
plot(
  output,
  steps = c(1, 5),
  preview_window = c(start_time, end_time),
  seed = 0
)

## (b) run a interactive workflow (with confirmation prompts after each step)

output <- eyeris::glassbox(demo_data, interactive_preview = TRUE, seed = 0)


# (2) examples of overriding the default parameters
output <- eyeris::glassbox(
  demo_data,
  interactive_preview = FALSE, # TRUE to visualize each step in real-time
  deblink = list(extend = 40),
  # only interpolate gaps up to 100 ms; longer gaps are left as NA
  interpolate = list(max_gap_ms = 100),
  lpfilt = list(plot_freqz = TRUE) # overrides verbose parameter
)

# to suppress messages, set verbose = FALSE in plot():
plot(output, seed = 0, verbose = FALSE, preview_window = c(10, 12))

# (3) examples of disabling certain steps
output <- eyeris::glassbox(
  demo_data,
  detransient = FALSE,
  detrend = FALSE,
  zscore = FALSE
)

plot(output, seed = 0, preview_window = c(10, 12))

eyeris documentation built on Aug. 1, 2026, 1:07 a.m.