Rejecting bad data spans

This tutorial covers manual marking of bad spans of data, and automated rejection of data spans based on signal amplitude. :depth: 2

We begin as always by importing the necessary Python modules and loading some example data <sample-dataset>; to save memory we'll use a pre-filtered and downsampled version of the example data, and we'll also load an events array to use when converting the continuous data to epochs:

Annotating bad spans of data ^^^^^^^^^^^^^^^^^^^^^^^^^^^^

The tutorial tut-events-vs-annotations describes how :class:~mne.Annotations can be read from embedded events in the raw recording file, and tut-annotate-raw describes in detail how to interactively annotate a :class:~mne.io.Raw data object. Here, we focus on best practices for annotating bad data spans so that they will be excluded from your analysis pipeline.

The reject_by_annotation parameter ~~~~~~~~~~

In the interactive raw.plot() window, the annotation controls can be opened by pressing :kbd:a. Here, new annotation labels can be created or existing annotation labels can be selected for use.

.. sidebar:: Annotating good spans

The default "BAD\_" prefix for new labels can be removed simply by
pressing the backspace key four times before typing your custom
annotation label.

You can see that default annotation label is "BAD_"; this can be edited prior to pressing the "Add label" button to customize the label. The intent is that users can annotate with as many or as few labels as makes sense for their research needs, but that annotations marking spans that should be excluded from the analysis pipeline should all begin with "BAD" or "bad" (e.g., "bad_cough", "bad-eyes-closed", "bad door slamming", etc). When this practice is followed, many processing steps in MNE-Python will automatically exclude the "bad"-labelled spans of data; this behavior is controlled by a parameter reject_by_annotation that can be found in many MNE-Python functions or class constructors, including:

• creation of epoched data from continuous data (:class:mne.Epochs)
• independent components analysis (:class:mne.preprocessing.ICA)
• functions for finding heartbeat and blink artifacts (:func:~mne.preprocessing.find_ecg_events, :func:~mne.preprocessing.find_eog_events)
• covariance computations (:func:mne.compute_raw_covariance)
• power spectral density computation (:meth:mne.io.Raw.plot_psd, :func:mne.time_frequency.psd_welch)

For example, when creating epochs from continuous data, if reject_by_annotation=True the :class:~mne.Epochs constructor will drop any epoch that partially or fully overlaps with an annotated span that begins with "bad".

Generating annotations programmatically ~~~~~~~~~~~

The tut-artifact-overview tutorial introduced the artifact detection functions :func:~mne.preprocessing.find_eog_events and :func:~mne.preprocessing.find_ecg_events (although that tutorial mostly relied on their higher-level wrappers :func:~mne.preprocessing.create_eog_epochs and :func:~mne.preprocessing.create_ecg_epochs). Here, for demonstration purposes, we make use of the lower-level artifact detection function to get an events array telling us where the blinks are, then automatically add "bad_blink" annotations around them (this is not necessary when using :func:~mne.preprocessing.create_eog_epochs, it is done here just to show how annotations are added non-interactively). We'll start the annotations 250 ms before the blink and end them 250 ms after it:

Now we can confirm that the annotations are centered on the EOG events. Since blinks are usually easiest to see in the EEG channels, we'll only plot EEG here:

See the section tut-section-programmatic-annotations for more details on creating annotations programmatically.

Rejecting Epochs based on channel amplitude ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Besides "bad" annotations, the :class:mne.Epochs class constructor has another means of rejecting epochs, based on signal amplitude thresholds for each channel type. In the `overview tutorial

` we saw an example of this: setting maximum acceptable peak-to-peak amplitudes for each channel type in an epoch, using the reject parameter. There is also a related parameter, flat, that can be used to set minimum acceptable peak-to-peak amplitudes for each channel type in an epoch:

The values that are appropriate are dataset- and hardware-dependent, so some trial-and-error may be necessary to find the correct balance between data quality and loss of power due to too many dropped epochs. Here, we've set the rejection criteria to be fairly stringent, for illustration purposes.

Two additional parameters, reject_tmin and reject_tmax, are used to set the temporal window in which to calculate peak-to-peak amplitude for the purposes of epoch rejection. These default to the same tmin and tmax of the entire epoch. As one example, if you wanted to only apply the rejection thresholds to the portion of the epoch that occurs before the event marker around which the epoch is created, you could set reject_tmax=0. A summary of the causes of rejected epochs can be generated with the :meth:~mne.Epochs.plot_drop_log method:

Notice that we've passed reject_by_annotation=False above, in order to isolate the effects of the rejection thresholds. If we re-run the epoching with reject_by_annotation=True (the default) we see that the rejections due to EEG and EOG channels have disappeared (suggesting that those channel fluctuations were probably blink-related, and were subsumed by rejections based on the "bad blink" label).

More importantly, note that many more epochs are rejected (~20% instead of ~2.5%) when rejecting based on the blink labels, underscoring why it is usually desirable to repair artifacts rather than exclude them.

The :meth:~mne.Epochs.plot_drop_log method is a visualization of an :class:~mne.Epochs attribute, namely epochs.drop_log, which stores empty lists for retained epochs and lists of strings for dropped epochs, with the strings indicating the reason(s) why the epoch was dropped. For example:

Finally, it should be noted that "dropped" epochs are not necessarily deleted from the :class:~mne.Epochs object right away. Above, we forced the dropping to happen when we created the :class:~mne.Epochs object by using the preload=True parameter. If we had not done that, the :class:~mne.Epochs object would have been memory-mapped_ (not loaded into RAM), in which case the criteria for dropping epochs are stored, and the actual dropping happens when the :class:~mne.Epochs data are finally loaded and used. There are several ways this can get triggered, such as:

• explicitly loading the data into RAM with the :meth:~mne.Epochs.load_data method
• plotting the data (:meth:~mne.Epochs.plot, :meth:~mne.Epochs.plot_image, etc)
• using the :meth:~mne.Epochs.average method to create an :class:~mne.Evoked object

You can also trigger dropping with the :meth:~mne.Epochs.drop_bad method; if reject and/or flat criteria have already been provided to the epochs constructor, :meth:~mne.Epochs.drop_bad can be used without arguments to simply delete the epochs already marked for removal (if the epochs have already been dropped, nothing further will happen):

Alternatively, if rejection thresholds were not originally given to the :class:~mne.Epochs constructor, they can be passed to :meth:~mne.Epochs.drop_bad later instead; this can also be a way of imposing progressively more stringent rejection criteria:

Note that a complementary Python module, the autoreject package_, uses machine learning to find optimal rejection criteria, and is designed to integrate smoothly with MNE-Python workflows. This can be a considerable time-saver when working with heterogeneous datasets.

.. LINKS