cili 0.5.3

Eyetracking data tools based on pandas

cili

Cili is meant to reduce the overhead of basic eyetracking data processing. While it sets the stage for more advanced work by providing data in the form of pandas DataFrames, manipulating the contents of those DataFrames works like manipulating any other pandas DataFrame, so that’s where cili stops - we leave it to the user to learn to work with pandas. If you’re going to be dealing with eyetracking data in python, you’ll be glad you did.

At the moment, we support EyeLink data only. We’d be happy to support other manufacturers, but don’t have the data on hand to do so. If you have the data, and would like us to support something in particular, please drop us a message at https://github.com/beOn/cili/issues, or add it yourself and submit a pull request.

Please note that this is an alpha release.

Installation

Before installing cili, you’ll need to install numpy. The rest of the dependencies should install when you install cili, but numpy is special. Then you can install cili using pip or easy_install:

pip install numpy
pip install cili

Or grab the latest development version from https://github.com/beOn/cili, and install using setup.py.

Examples

Parsing EyeLink Data

Cili can parse samples and events from EyeLink .asc files, and samples from EyeLink Experiment Builder .txt files. You can use the same method for both, but need to ignore the second returned value if you’re using a .txt.

from cili.util import *
# using .asc
samps, events = load_eyelink_dataset("/some/file.asc")
# using .txt
samps, _ = load_eyelink_dataset("/some/file.txt")

Handling Blinks and Missing Data

Try as we might to recruit cooperative subjects, chances are your data is full of blinks; try as we might to set up the eyetracker correctly, chances are there’s some signal dropout. Cili provides a couple of convenient methods for deriving values missing for either of these reasons using linear interpolation.

Handling blinks in EyeLink data doesn’t seem to tricky at first blush, but there are a couple of subtleties and related options to be aware of.

First, EyeLink embeds every blink event within a saccade, and recommends that if you plan to scrub out blinks, you should also scrub out saccades containing blinks. Cili does this automatically when you call mask_eyelink_blinks().

Second, EyeLink’s blink marking algorithm is a little too aggressive for my taste when it comes to declaring a blink’s end time. Even when we interpolate over the containing saccade, the reported pupil size for several dozen milliseconds after the interpolated range often contains absurdly low values, with an absurdly high slope. This can add a lot of noise to your data, and I can’t imagine anyone arguing that these values have any bearing on the eye’s real state. So cili will optionally creep the blink recovery time forward by looking for the first point within 1000ms where the 100-sample rolling average of the z-scored derivative of the pupil timecourse drops to within .1 of the entire timecourse’s average. This method has worked pretty well for us, but there’s still room for improvement.

All non-blink signal dropout will be recorded as 0s. We don’t make any adjustments to the dropout onset/offset times, but otherwise these values get handled in much the same way as blinks. But you should always clean dropout after cleaning blinks, otherwise the blink recovery index method described above won’t work.

So, let’s roll up our sleeves and clean some data!

from cili.util import *
from cili.cleanup import *
samps, events = load_eyelink_dataset("/some/file.asc")
samps = interp_eyelink_blinks(samps, events, interp_fields=["pup_l"])
samps = interp_zeros(samps, interp_fields=["pup_l"])

Well that was kindof anticlimactic.

Note that if you collect the right pupil, “pup_l” should be changed to “pup_r”, and if you collect both eyes, you’ll want to include both pup_l and pup_r in interp_fields. EyeLink’s Experiment Builder calls the same values “right_pupil_size” and “left_pupil_size” in the .txt files it generates, so if your samples came from a Experiment Builder .txt, use those names instead.

Check the documentation on these methods FMI.

Smoothing Data

If you look closely at EyeLink data, you’ll probably notice a little high frequency noise. This can be a little problematic in several circumstances. To deal with it, cili provides a butterworth filter function with default settings based on previously published pupillometry studies. You can modify the order and cutoff frequency of the filter if you like, but the basic usage looks like this:

samps = butterworth_series(samps, fields=["pup_l"])

FMI, check out the documentation on butterworth_series.

Events from a List of Dicts

Sometimes you are interested in events recorded using something other than EyeLink software. For those crazy times, if you can turn that data into a list of dicts, each containing a name, onset and duration, then it’s pretty easy to create a cili Events object. Assuming you already have your list of dicts, and that it’s called list_o_dicts, then all you do is:

from cili.models import Events
events = Events.from_list_of_dicts(list_o_dicts)

Extracting Event-based Ranges

To my mind, this is where things start to get interesting. In many eye tracking and pupillometry studies, the goal is to examine a collection of sample ranges surrounding certain events. So cili provides a method for extracting sample ranges based on event timing, returning a DataFrame with a MultiIndex (event #, sample #).

Suppose you were interested in the 10 seconds following every event in some Events object, called “events,” and you have a 1kHz Samples object, samps. To extract this range for every event in events, you would:

from cili.extract import extract_event_ranges
ranges = extract_event_ranges(samps, events, end_offset=10000)

Often, pupillometric sample ranges will be transformed into a % deviation from baseline measure, where the baseline is an average of some small range immediately preceding the range of interest. Continuing the example above, let’s extract baseline measures for each of the events, then divide the ranges of interest by the baselines for the field “pup_r”:

baselines = extract_event_ranges(samps, events, start_offset=-100, end_offset=-1).mean(level=0)
ranges.pup_r = (ranges.pup_r / baselines.pup_r - 1).values

Not so painful! For more info on range extraction, check out the documentation on extract_event_ranges. To work with the returned data effectively, You’ll probably also want to take a minute to learn about pandas MultiIndex objects.

Borrowing Event Attributes

If your events each have a field, let’s say “subject”, and you’d like to insert each event’s value for that field into every row of the corrisponding range under a column of the same name, you can “borrow” event attributes using borrow_attributes, like so:

ranges = extract_event_ranges(samps, events, end_offset=10000, borrow_attributes=["subject"])

Saving and Loading

If you keep reading and writing large .txt files, you’ll die young. Or at least having spent too much of your time waiting for .txt files to be read or written. So cili uses hdf5 to speed things up. To use this, you’ll need to install h5py and its dependencies, as documented at http://docs.h5py.org/en/latest/build.html.

Once that’s done, saving and loading Samples and Events objects is pretty easy. It works the same way in both cases, so we’ll just work with samples below:

from cili.models import Samples
samps.save("some_filename.hdf")
samps_2 = Samples.load_saved("some_filename.hdf")

Exporting to .txt

If you have to export samples or extracted ranges to a .txt file, fine. Ok. We understand.

Luckily, pandas datasets already include a function for writing csv files, any several other formats as well (check their documentation for the complete list: http://pandas.pydata.org/pandas-docs/stable/io.html). For example, to create a tab delimited .txt file:

samps.to_csv("some_filename.txt", sep="\t")

To create a Zamboni delimited .txt file, just set sep to “Zamboni”.

Checking For Signal Dropout

Sometimes, for one reason or another, eyetracking sessions can go pretty poorly. Usually, this means that there’s a high level of signal dropout due to blinks, or the tracker losing track of the eye. One way to check for this is to see what percentage of the timeline’s pupil value(s) were recorded as 0. Cili’s util.py offers a convenient way to check the dropout rate for all of the .asc files in a directory, like so:

util.py --dropout -d /path/to/dir/containing/ascf_iles/

Reporting Bugs, Requesting Features

Submit all bug reports and feature requests using the github ticketing system: https://github.com/beOn/cili/issues

Please make an effort to provide high quality bug reports. If we get one that just says, “sample range extraction is broken,” we’ll probably trash it without a second look, because the submitter is probably the kind of person who saps energy from everything they touch.

A good bug report should include three things:

1. Steps to reproduce the bug

2. Expected result

3. Actual result

The goal is to give the developers the ability to recreate the bug before their own eyes. If you can give us that, we’ll take a very close look.

Why Cili?

Because, like the mighty ciliary muscles, it brings your eye data into focus.

TODO: Thanks, credit to CCP Lab