Building Your Own Custom Pipeline Extensions

In eyeris: Flexible, Extensible, & Reproducible Processing of Pupil Data

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

One key strength of the preprocessing framework within eyeris is its modularity.

While we encourage most users to use the glassbox() function for simplicity and reproducibility, advanced users can create custom preprocessing steps that seamlessly integrate into the pipeline.

This vignette walks you through the structure required to write your own eyeris-compatible preprocessing functions.

🧩 How the Pipeline Works

Under the hood, each preprocessing function in eyeris is a wrapper around a core operation that gets tracked, versioned, and stored using the pipeline_handler().

Custom pipeline steps must conform to the eyeris protocol for maximum compatibility with the downstream functions we provide.

Following the eyeris protocol also ensures: - all operations follow a predictable structure, and - that new pupil data columns based on previous operations in the chain are able to be dynamically constructed within the core timeseries data frame.

For instance: pupil_raw -> pupil_raw_deblink -> pupil_raw_deblink_detransient -> ...

If you're unfamiliar with how these columns are structured and tracked, first check out the companion vignette: 📦 Anatomy of an eyeris Object.

🛠️Creating a Custom Extension for eyeris

Let’s say you want to write a new eyeris extension function called winsorize() to apply winsorization to extreme pupil values.

1) Write the core operation function

This function should accept a data frame x, a string prev_op (i.e., the name of the previous pupil column), and any custom parameters.

To illustrate:

winsorize_pupil <- function(x, prev_op, lower = 0.01, upper = 0.99) {
  vec <- x[[prev_op]]
  q <- quantile(vec, probs = c(lower, upper), na.rm = TRUE)
  vec[vec < q[1]] <- q[1]
  vec[vec > q[2]] <- q[2]
  vec
}

2) Create the wrapper using the eyeris::pipeline_handler()

The pipeline_handler() enables your function to automatically:

• track your function within the eyeris list object's params field,
• append a new column to each block within the timeseries list of data frames, and
• update the object's latest pointer.

To illustrate:

#' Winsorize pupil values
#'
#' Applies winsorization to extreme pupil values within each block.
#'
#' @param eyeris An `eyeris` object created by [load_asc()].
#' @param lower Lower quantile threshold. Default is 0.01.
#' @param upper Upper quantile threshold. Default is 0.99.
#'
#' @return Updated `eyeris` object with new winsorized pupil column.
winsorize <- function(eyeris, lower = 0.01, upper = 0.99) {
  pipeline_handler(
    eyeris,
    winsorize_pupil,
    "winsorize",
    lower = lower,
    upper = upper
  )
}

1. Here, the first argument is always the eyeris class object.
2. Then, is the name of the function (the one you created in step 1 above).
3. Third, is the internal label you want eyeris to refer to (i.e., the one for the column name, plots, etc.).
4. Fourth position++ are the parameters you created after prev_op when creating the function winsorize_pupil in step 1.

🎉 And that's it!

You should now be able to use your new function extension as a component within a new custom eyeris pipeline declaration. To illustrate:

system.file("extdata", "memory.asc", package = "eyeris") |>
  eyeris::load_asc(block = "auto") |>
  eyeris::deblink(extend = 50) |>
  winsorize()

💪 Best Practices

• Use consistent naming: i.e., match your suffix (e.g. "winsorize") to the column and log structure:
• If you call pipeline_handler(..., "winsorize")
• Then, your function names should reflect that:
  • the public facing winsorize() wrapper function, and
  • the private winsorize_pupil() logic implementation function.
• You will then see pupil_raw_*_winsorize as the new output column!
• Respect previous steps: Your custom function should rely only on prev_op, not on any hardcoded column names!
• Return the expected data type: Be sure that you private function always returns a modified vector type, as the underlying pipeline_handler() is looking out for a vector it can transpose into the new column that will be added to the timeseries data frame within the resulting eyeris object.
• Test on individual blocks: Always try your private function logic on a single data block (e.g., eyeris$timeseries[[1]]) to debug before integrating it with the eyeris pipeline protocol.

✨ Summary

We hope you are now convinced at the power and extensibility the eyeris protocol enables! As we demonstrated here, with just a little bit of structure, you can create custom extension steps tailored to your specific analysis needs -- all while preserving the reproducibility and organizational core principles eyeris was designed and built around.

If you'd like to contribute new steps to eyeris, feel free to open a pull request or discussion on GitHub!

---

📚 Citing eyeris

If you use the `eyeris` package in your research, please cite it! Run the following in R to get the citation:

citation("eyeris")

eyeris documentation built on April 12, 2025, 2:05 a.m.