ICA and Mobile EEG: Separating Brain from Body in Moving Subjects
This post is still being written and revised, so please take any information you can use, but beware of its alpha status.
As mentioned, over the past months I have had a long look at EEG data recorded from subjects in various states of locomotion and body posture, and have since tried to make out how to get the most from it – both results and insights. In order to deal with the noise generated by bodily movement and related muscular activity, one promising approach has been to apply independent component analysis to EEG data. However, despite its successes with data recorded in ‘normal’ lab conditions, its application to mobile EEG recordings is still far from trivial, the reasons of which I intend to elaborate on here.
Anyone who reads this section, I assume already to be familiar with basic pre processing and analysis of EEG data, the EEGLab toolbox, the concept of ICA and its data driven, unsupervised ability to uncover statistically independent sources (components) of signals in neurophysiological recordings. Have a look here if you need some more information on ICA and its uses in neuroscience and neuropsychology.
ICA and Movement Artifacts
While conducting ICA on one particular data set (see the example case here), I noticed much of the decomposition was dominated by a beautiful noise free (meaning in this case: devoid of any believable neural activity) continuous movement artifact at one central electrode site (-a pedometer would not have provided a better signal), but also in lesser form at surrounding electrodes, and at other sites. The presence of this pretty waveform was obviously an error on behalf of the experiment leader (..ahum..) who might have spotted it in the continuous data on screen and subsequently fiddled with the cables until it no longer showed, but unfortunately his inexperience with mobile EEG at that time did not make him look twice.
In response to this I could have developed a new found appreciation for tying down participants to chairs as is still a good tradition in (neuro)psychology, or perhaps just fixating only their heads with some apparatus similar as used with rodents and fruit flies. Fortunately, this conveniently extreme example of an EEG recording – and in particular the ICA decomposition – gone bad due to locomotive movement provided a unique opportunity to do a closer examination of what the undesired decomposition had uncovered.
Being at the frontier of getting fundamental neuropsychological research to stand an walk on its own two feet, the topic of how to deal with motion artifacts in EEG has not seen its last publication yet (Nathan & Contreras-Vidal, 2016; Kline, Huang, Snyder & Ferris, 2015; Castermans, Duvinage, Cheron & Dutoit, 2014).
The essence: Many body movements, brain and muscular activity we assume to be independent with ICA, and would much prefer to study and analyze in isolation, are arguably quite the opposite of statistically independent – and this shows!
What already can easily be observed without too much in depth analysis is that many movements actually aren’t that statistically independent from each other, or even from brain signals. Some subjects have a tendency to blink in perfect harmony with their button responses, others appear to time their button presses using their stepping frequencies.
Adding to this; by introducing full-body movement we are introducing vestibular input into a deeply ingrained, highly automated system that luckily is bound to introduce extremely predictable and strongly correlated signals across the skull, so should not pose too great a problem for ICA.
In case of the former: in particular ocular and body movements appear to go hand in hand. Intuitively this makes sense; when a person walks while keeping their eyes fixed on any object, the position of the eyes in the skull will consistently respond to body sway movements (not merely direction of one’s gaze changes, which most will gather consists of two-dimensional pitch and yaw movements, but even tiny adjustments along the third axis occur, namely rolling of the eye; see also). As is well known in EEG research, any movement of the eyes creates vast changes in potentials recorded at the majority of scalp sites.
To Do;
-Independence of EMG?
-Data stationarity across conditions? (possibilities for AMICA)
–