[Eeglablist] CEEG Processing and Quantitative EEG Parameter Calculation in ICU Patients With Noisy Data

[Scott Makeig]

Thomas -

One variable you may not have thought of is nonstationarity.  ICA
decomposition finds a fixed set of  maximally (temporally) independent
information sources that sum to the scalp data. If there are spatially
fixed brain equivalent sources (e.g., areas of cortex within which LFP is
partially coherent) then a single ICA model can identify their (maximally)
distinct signals and scalp projection patterns. However, if there are any
changes in these (e.g., as in normal sleep) then the ICA source solution
must change.  Thankfully, nearly 15 years ago Jason Palmer, then a graduate
student working with SCCN, developed AMICA including a multi-model option.
The value of this option has been further revealed by the more recent work
of Sean Hsu, who showed that using as many as 20 models may act as a quite
sensitive cognitive state change indicator, and further, characterize the
EEG in each identified state model... Multi-model AMICA may also be used to
automatically perform sleep staging,
<https://sccn.ucsd.edu/~scott/pdf/Hsu18_AMICA.pdf> his results
indicated, indicate
changes in emotion during an emotion imagination exercise,
<https://urldefense.com/v3/__https://www.sciencedirect.com/science/article/pii/S1053811922000039__;!!Mih3wA!BffDCi9Zo1iNHm-3-fMiVFajOw8MJA1H148CuYQi15pC7ZtJMGUQS4BJrm9FDt4rTFBawJStf93-oJrKdEVE$ > etc.

So far as I know, AMICA has not so far been used to explore state changes
in ICU patients, or using ~21-chan 10-20 montages.  Here, the problem of
multiple sources (>21) including noise sources challenges ICA decomposition
-- and performing many-model decomposition of long records may thus make
more sense.  This could prove to be a rich avenue to explore...

Scott Makeig

On Mon, Dec 16, 2024 at 6:16 PM Johnson, Thomas via eeglablist <
eeglablist at sccn.ucsd.edu> wrote:

> Hi there,
>
> First of all, I want to thank the EEGLab community in advance for being so
> generous with their limited time in helping beginner users such as myself.
> EEG analysis is a rich and complex world, and having a community to help
> get one started makes it much less intimidating.
>
> Now to my problems:
>
> Setup:
> We are a lab studying various neurophysiologic measures in ICU patients
> that require mechanical circulatory support (e.g. ECMO). As part of this,
> we have continuous EEG data for extended periods (i.e. 24 hours per day,
> 3-7 days per study). This data is noisy due to the poor electrical
> isolation of the ICU environment and the extra machinery an ICU patient
> requires (e.g. automatically inflating mattress to reduce pressure sores,
> intermittent pneumatic compression devices to prevent DVT, infusion pumps,
> ECMO machine, etc.). We have a limited number (typically 21 or 22) of leads
> in a 10-20 system and collect data at 512 Hz via XLTek before being
> down-sampled to 256 Hz and saved on disk. We also concurrently measure
> relative cerebral blood flow (rCBF) for approximately two hours per day
> with a diffuse correlation spectroscopy (DCS) system that uses
> near-infrared spectroscopy (NIRS).
>
> For my current project, we are interested in qEEG measures during these 2
> hours of EEG and DCS co-measurement. The patients are typically comatose,
> so there are no experimental ‘events’. The data is purely observational. We
> are scaling up our project after recently receiving grant funding, and I am
> trying to make a semi or fully automated data processing pipeline for our
> group. As the first step in this process, I am comparing the results of a
> semi-automated pipeline to manually selected segments of the time series
> that an expert reviewer (i.e. an epileptologist) has deemed to have minimal
> artifact. My reasoning is as follows: if the qEEG measures are similar
> between the two methods, the semi-automated pipeline can be used, thus
> saving valuable time for the epileptologist.
>
> To begin this analysis I have established the following pre-processing
> workflow after reading the tutorials on the EEGLab website, as well as
> Makato’s pipeline:
>
>
>   1.  Import Data with BIOSIG
>   2.  Narrow Into DCS Time Range
>   3.  Populate Channel Locations Using Known Labels
>   4.  Manually Exclude Obviously Bad Channels
>   5.  clean_rawdata() to reject bad channels
>   6.  Duplicate Dataset Narrowed To Expert Reviewer Time Segments
>   7.  1 Hz High-Pass Filter
>   8.  Zapline 60 Hz
>   9.  Interpolate Removed Channels
>   10. Re-reference To Average
>   11. Calculate ICA Components
>   12. ICLabel
>   13. Review & Reject ICA Components
>
> I have a few questions that have arisen after working through this
> pipeline for the first time.
>
>
>   1.  I initially tried to exclude channels by only using clean_rawdata,
> but found that it was missing channels with very high amplitude sinusoidal
> artifacts (not 60 Hz, but certainly from a machine). I tried adjusting the
> 'ChannelCriterion' parameter, but this did not work well either at 0.70 or
> 0.90. I ended up manually excluding these channels so I could move forward.
> Would this type of artifact be better removed using ICA?
>   2.  In a similar vein, is there any benefit to removing line noise prior
> to doing ICA?
>      *   My initial thought is that if we have limited leads and thus
> limited components to use for analysis, we might as well remove one source
> of noise of which we have a priori knowledge, and ‘save’ the components for
> other noise.
>      *   However, the idea of removing all noise sources with a single
> non-destructive technique is appealing.
>   3.  Is there a benefit to using epochs in this type of continuous EEG
> data, i.e. where there are no ‘events’?
>      *   My initial thought is that this might be helpful in calculating
> alpha/delta ratio and other qEEG parameters, as I can leverage the epoch
> infrastructure to calculate parameters more easily per each epoch.
>   4.  Some of the ‘expert-reviewed’ time segments are brief (e.g. 3-5
> seconds). If I chose a 5-second epoch length, how does it handle data that
> contains boundaries within the epoch?
>   5.  Is anyone aware of a plugin that will calculate some of these qEEG
> parameters? It seems common enough in the literature that someone has
> created the scripts at some point, but I didn’t see any plugin that matched
> this description on the website.
>      *   We are interested in parameters similar to Tjepkema-Cloostermans
> et al. (2017),
> https://urldefense.com/v3/__https://pubmed.ncbi.nlm.nih.gov/28430695/__;!!Mih3wA!CO_ifNmKaFvV5SxJrxhMu3ASeZffudYh2CK5B3JUL18YRuXSl4_jREAO7b88pEFwzViEIA1GOqP04VZcG-9lDW3jIbtUFvVZCaM5m60$
> . This includes ADR, power, Shannon entropy, delta coherence, regularity.
>
> Thank you in advance for any help you can provide.
>
>
> Thomas W. Johnson, MD, PhD
> Resident Physician, PGY-4
> Department of Neurology
> University of Rochester Medical Center
>
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>


-- 
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation, University of
California San Diego, La Jolla CA 92093-0559, http://sccn.ucsd.edu/~scott