updated for 2024: list of summer schools & short courses in the realm of (computational) neuroscience or data analysis of EEG / MEG / LFP and the like:
invasive recording over the human visual cortex: a beautiful alpha-rhythm. a visual stimulus is shown during the shaded intervals, the stimulus attenuates the rhythm. the rhythm returns when the stimulus is turned off. (neat to see this classic effect so clearly in ECoG ♥️)
"but the correlation is so large, it must be a robust effect" 🤔 for small samples, large values for the computed correlation coefficient are more likely to appear, even in the absence of any relationship. see here for 2 independently generated variables & different sample sizes:
in the field of brain rhythms, one of the rock-solid empirical findings is that frequency increases across development, for alpha and mu rhythms.
in this new 〰️preprint〰️, we also look at waveform shape across development & neurodevelopmental disorders:
the EEG signal is a mixture of many different contributions. for example: midfrontal theta-, sensorimotor mu- and posterior alpha-rhythms, as well as muscle noise & eye blink artifacts. below, these components are added up successively onto electrode signals for illustration.
& I have been awarded a Marie Skłodowska-Curie postdoc fellowship for my project to study traveling waves in the brain, here at
@ESI_Frankfurt
. looking forward to tell you all more about it in the next 2 years! 🌊🙂
#msca
#wavescope
I like this tool for graphical exploration of scientific literature: you put in some seed papers & it returns a network, with papers that are commonly citing or cited by seed papers. very helpful, often papers already exist, one just needs to find them. 🙂
intracranial EEG data from the human brain features extremely rich dynamics. this is a scan through frequency space to find pronounced oscillations in different regions.
〰️🙂〰️
some of these rhythms are quite hard to see with non-invasive recording modalities!
I saw lots of beautiful beta-rhythm work lately! there is one specific aspect which I find non-optimal: in the human brain, there are also lots of alpha-rhythms, which may produce harmonics in the beta-band. I wrote about it here in this short article:
waveform shape of oscillations can reflect properties of underlying spiking dynamics. if neuronal units fire in synchrony, the local field potential can be more asymmetric (here: sharp troughs).
read about this & cool bat rhythms 🦇 in our new paper:
A large portion of measured EEG signal is of non-neuronal origin. Signal separation techniques like independent component analysis can help to identify these. Here, noise components are slowly projected out, revealing the interesting bits, the alpha oscillations.
new preprint!🌟we show how data-driven referencing can be useful for analyzing oscillations in intracranial electrophysiological recordings and explore waveform shape & spatial spread & variability across participants: (1/n)
A classic effect in the sensorimotor cortex: event-related desynchronization. Strong mu-rhythm oscillations attenuate as soon as the participant starts to contract a muscle. Where do they go? 🧐
many EEG analyses are done in sensor space. in this preprint with Vadim Nikulin
@MPI_CBS
, we investigate how different types of alpha rhythms contribute to activity of individual EEG electrodes. 1/8
there are different models of how the brain responds to rhythmic input, 1) by adjusting an intrinsic oscillator or 2) by a series of evoked responses. adjusting the input rate, the 2 models have different properties, exhibiting a constant or variable phase shift to the input.
1/f-exponent & E/I-balance: while changes in synaptic time constants certainly influence spectral measures, not all changes in the 1/f-exponent can be interpreted as changes in E/I-balance: starting with the most basic things like artifacts, here shown for EEG & eye blinks.
oscillations recorded with EEG electrodes are a mixture of different sources. therefore, phase of sensor space signals does not necessarily reflect the phase of a singular oscillation; swinging back & forth in phase space, depending on synchronization of underlying rhythms.
two oscillatory sources, producing rhythms of same frequency. both contribute to activity of nearby electrodes with a distance-dependent weighting. changing the phase relationship of the sources changes the traveling wave direction that can be measured on the electrode signals.
you want to extract instantaneous phase of a signal, the textbooks say to narrowband filter before computing the analytic signal. but why?
using a broadband signal, the resulting phase corresponds to the circular mean of the analytic signals of the constituents (here: 🔴=🟢+🔵).
spike-field coupling: when spikes occur at a preferred phase of an ongoing oscillation. in below example, a homogeneous Poisson spike train (spiking is distributed uniformly across time) is morphed into an inhomogeneous one, with more frequent spikes at the trough of the LFP.
the two hemispheres have their own neuronal rhythms. sometimes they swing together for a bit, then drift apart again. (EEG, resting state sensorimotor mu-rhythms)
when recording EEG, the orientation of underlying dipoles will heavily influence what kind of field will be picked up on the sensor level. (single dipole placed in hand knob area of primary motor cortex)
the institute storage rooms just keep on giving, found some beautiful old scientific posters there:
"electrical stimulation points in the human cortex" (1931)
I saw lots of beautiful beta-rhythm work lately! there is one specific aspect which I find non-optimal: in the human brain, there are also lots of alpha-rhythms, which may produce harmonics in the beta-band. I wrote about it here in this short article:
during a long experiment, there can be different sources of variability. for example, the participant can get sleepy, which influences measures like reaction times on a slower timescale. these slow drifts can then mask differences between experimental conditions.
when analyzing LFP & spikes, only few units exhibit spike-field coupling. but there are lots of rhythms in the brain, how do they arise? this toy model here shows that only a small fraction of oscillatory units need to be active in synchrony to result in a discernible rhythm.
a glimpse on the electric orchestra that is playing all the time in the brain, with EEG rhythms appearing and disappearing concurrently. for all of them: distinct peak frequencies and waveforms, distinct spatial topography, distinct modulation by behavior. 🌊
when recording from the brain using MEG, distance to the sensors is crucial in determining the signal. here is a fun test we ran when presenting auditory stimuli: squeezing the participant close to the sensors results in quite a boost in evoked response amplitude.
how could one arrive at a traveling wave signal with EEG? take 1 occipital & 1 sensorimotor alpha rhythm. the resulting EEG signal will have varying peak times along the posterior-anterior direction. apparent wave direction will depend on phase shift between rhythms.
does phase of a rhythm matter? had a chance to talk about our closed-loop TMS experiments recently and decided to plot everything in a more single-trial manner. I would say it looks quite nice. 🙂 a macroscale rhythm, detected in a non-invasive way, with effects on excitability.
in auditory paradigms using MEG, peak activation is over temporal sensors. with EEG, curiously often peak activation is on vertex channels. 🤔
it feels counterintuitive, given the distance from source, right? but it is expected, as this simulation with 2 active dipoles shows.
♥️ anatomical line drawings in old papers, here: distribution of alpha and beta rhythms across the cortex in electrocorticography recordings. Jasper & Penfield (1949)
we often use Fourier analysis (=sinusoids) to represent brain signals. does this mean that there are many sine oscillators in the brain (1 per frequency)? we can also use triangular functions to reconstruct the signal. so are there 1000 tiny triangular oscillators in the brain?
sometimes for electrophysiological signals, mean spectral power in specific frequency bands is computed, often with a kind of normalization.
➡️ demo how to get different ratios of low & high freq power without any oscillatory contribution (by changing 1/f-exp. & spectral knee).
Fourier Transform: transform time-domain signals into the spectral domain, easy. but why use only these two domains?
➡️ fractional Fourier Transform, for transform into the space in between. time domain – alpha=0; spectral domain – alpha=1, other alphas: a strange world. 🙃
one reason why I think it's always good to see original spectra before running any kind of fancy analysis: it's quite easy to create oscillations magically out of noise. for instance, in the below demo by grouping trials (which contained only 1/f-activity) by frequency content.
recorded EEG recently for the first time in 3 years! still feels great watching the squiggly lines on the screen. 🙂 & my colleague
@pwdonh
has some amazing mu-rhythms.
besides occipital alpha- & sensorimotor mu-rhythms, there are also other ~10 Hz rhythms in the human brain. for instance, tau-rhythms originating in the temporal lobe, which have been linked to auditory function. have not seen a temporal rhythm so clearly before. ECoG ❤️!
influence of muscle noise on spectrum & 1/f-exponent estimation, example from EEG. generally: the more muscle noise, the flatter the spectrum at higher frequencies.
for me, the hardest step is always averaging over subjects, since they display large heterogeneity. here an example of how alpha-power is related to visual target detection (the classical result: lower occipital power for hit trials), in the average & for some selected subjects.
Sensorimotor oscillations are not passively idling, they modulate cortical excitability with behavioral consequences. Magnetic stimulation given at the trough of the EEG mu-rhythm results in much larger muscle responses compared to peak stimulation.
I am moving to Germany/Rhine-Main area in the fall & looking for a new postdoc-type job then! if you know something cool, send it my way. 🙂 love data and all the beautiful patterns in EEG/ECoG/LFP-recordings of brains.
MEG/EEG: "we use source reconstruction" – all problems solved?
plotted here: crosstalk – possible contribution from other locations to reconstructed activity @ red dot. no strong activity at these locations: 👍 otherwise: 🫠
➡️crosstalk can be from distant locations
calculating power spectra with Welch's method entails splitting the signal into segments of a certain length and averaging across segments. too short segments: low frequency resolution. low number of segments: noisy estimates. the sweet spot: somewhere in between. 🤔
2 different signals, for which the phase is extracted via the hilbert transform. one will get some numerical values for phase in any case, but these values are not always meaningful in a physiological sense, without an oscillation present.
sound direction can be inferred from temporal differences of sound arrival at each ear. Jeffress model accomplishes this by mapping time into space with delay lines: spatially arranged units are activated depending on when inputs from both ears arrive simultaneously at the unit.
EEG analysis in sensor space may yield distorted measures. 2 rhythms (at red & blue locations) with slightly different peak frequency mix across space. the power spectrum of the signal recorded at the green electrode now has 2 peaks in the alpha band. 🤔
apparently, many students at
@neuromatch
academy have interest in working with EEG/MEG/ECoG (good choice 🙂). the organizers are looking for mentors (postdocs & PIs) who can help guiding data analysis projects! small time commitment in July, easy sign-up:
if you combine two rhythms, what determines the resulting modulated signal? one aspect is the difference between the two frequencies, resulting in two spectral peaks left and right from the center frequency.
"beta-activity as bursts", I am def 💯 OK with that.
but imho the pendulum has swung too far in the direction that all beta activity equals close to 1-2 cycle bursts. so just to illustrate diversity, here is some more sustained beta-activity, resting-state EEG.
〰️🙂〰️
Inter-trial coherence measures phase-synchronization across trials. It is the circular sum of phases at a certain point in time (length of red arrow). It reaches its maximum value of 1 for perfectly phase-aligned signals and becomes 0 as the phase distribution becomes uniform.
tracking phase with a phase-locked loop: phase of a signal is compared to a reference oscillation. depending on the low-pass filtered phase difference, the reference frequency is adjusted. filter properties determine whether phase can be tracked after sudden frequency jumps.
intracranial activity from the awake but resting human brain in the delta-frequency band (1–4 Hz), a couple of bursting cycles, not a sustained rhythm.
what happens if you apply PCA on traveling wave type activity? this will result in two phase-shifted principal components, which can be combined to reconstruct the original wave. (example for laminar LFP recordings)
sadly, this interesting representation of EEG power spectra in the form of chernoff faces (ten spectral parameters mapped onto features of faces) from 1987 did not reach mainstream. 😀
traveling waves & sources, the empirical edition: for hippocampal theta, 2 different sources can reliably be found, which show specific phase shift. the phase extracted from raw LFP points to a traveling wave direction that is in agreement with that (here: ~top to bottom row).
oscillations and ERPs often feel like separate worlds, but in some ways, they are different sides of 1 coin. according to the baseline shift account, late ERPs arise because of the non-zero mean property of oscillations, which change their power in the presence of stimuli.
the alpha rhythm is the most prominent rhythm in the human brain. some animals also have prominent resting alpha-rhythms, including cats & dogs. while brain size changes a lot when comparing across species, alpha peak frequency curiously seems to be quite consistent.
imagine: you train a classifier to predict something from a bunch of recording electrodes. you obtain weights, 1 value for each electrode. if only 1 weight is unequal 0, does it mean the informative signal is only present on that single electrode? 🧵⬇️
the power spectrum & the autocorrelation function (ACF) of a signal are heavily related. so, spectral & ACF measures will capture similar aspects of the data. shown here: how changes in spectral exponent lead to changes in ACF full width at half-maximum amplitude.
our article illustrating sensor space spatial mixing of different alpha-rhythms is now out in
@NeuroImage_EiC
: you can check out the thread describing the preprint here:
many EEG analyses are done in sensor space. in this preprint with Vadim Nikulin
@MPI_CBS
, we investigate how different types of alpha rhythms contribute to activity of individual EEG electrodes. 1/8
EEG alpha peak frequency varies as a function of alpha power for some subjects, as shown by below spectra computed for segments sorted by power in alpha band. I wonder how this could influence gradients in peak frequency along posterior-anterior axis. 🤔
new preprint!🌟we show how data-driven referencing can be useful for analyzing oscillations in intracranial electrophysiological recordings and explore waveform shape & spatial spread & variability across participants: (1/n)
good looking N100 🙃
(the university hospital where we record MEG had a cyber attack in October and it took until last week to get MEG data access reinstated... but now the data is back! ♥️🧲🙂)
a report from Adrian & Yamagiwa (1935) on the EEG alpha-rhythm, recorded with electrodes placed on the midline. it is mostly about rhythm-localization, but one interesting figure shows alpha as a traveling wave, with propagating peaks and troughs. 🌊
behavioral and neuronal measures acquired over the course of an experiment display fluctuations on fast as well as slow scales. if one computes a correlation between two _independent_ sets of such measurements, very likely strong spurious correlations will be seen.
if you are interested in time series analysis of oscillations, check out this! my most common burst recommendations are implemented:
✅ determine frequency band of interest via spectrum
✅ sensitivity analysis of main output measures to show independence from chosen parameters
Bats have beautiful rhythms! In a new preprint,
@nschawor
,
@Talking_Bat
and myself characterize and compare the waveform shape of delta and gamma oscillations in the bat auditory and frontal cortex. We found some cool things! 🧵
A single ion channel is stochastic; but looking at many of them at once will reveal regularities, slowly approaching a time course as described by the Hodgkin-Huxley model. (here: delayed rectifier potassium channels in a voltage clamp setup; inspired by rereading Dayan & Abbott)
"7 things to consider when analyzing oscillations in the brain"
are we really measuring what we intend to measure? for me, getting a definite "yes!" to this question is the foundation of solid science. we try to provide a field guide to methods aiming at that for oscillations.👇
New Preprint 🎉
"Methodological considerations for studying neural oscillations"
With Natalie Schaworonkow (
@nschawor
) and Bradley Voytek (
@bradleyvoytek
), we review key methodological issues and concerns for analyzing oscillatory neural activity.
🧵:
traveling waves, 1973 edition: a video showing the work of Walter Freeman III in the olfactory bulb & cortex, looking at evoked responses and spontaneous waves with an electrode grid in the cat.
(thanks to
@researchEnginee
for this obscure find!♥️)
Across participants, there is a large range of measurable EEG oscillation strength. For some participants, oscillations are very pronounced, for others miniscule. One can exclude participants for the sake of good measurement, but then the generalizability suffers. Tough choice!
Reproducing really old results (~1930s): EEG visual alpha frequency increases over development.
So many subjects in one scatter plot 😮. All of the extensive data wrangling by Voytek lab graduate rotation student Andrew Bender.
a small side project that I've been working on: an analysis of aperiodic activity and alpha oscillations in an open-data longitudinal infant EEG dataset: 1/n
I really like papers with many beautiful raw traces. These are from Contreras & Steriade (1995, ) and shows simultaneous EEG&intracellular recordings in motor cortex. Cool looking bursts! (sadly, this presentation form seems to have gone a bit out of style)
built a spikeling: a model neuron that runs on a small arduino nano chip & can fire action potentials. for instance, by shining light on a photodiode the neuron can be excited (1st half of video) or inhibited (2nd half). love the sound output!
updated for 2024: list of summer schools & short courses in the realm of (computational) neuroscience or data analysis of EEG / MEG / LFP and the like:
the whole lab has pretty diverse backgrounds. so in order to establish some common ground, we decided to read a book in journal club earlier this year. we picked this one, "slim" in the title was also decision factor 🙂 & I liked that it featured a bunch of very recent studies.
I couldn't find an autocorrelation GIF I liked, so made my own!
#Autocorrelation
function (ACF) measures self-similarity of a signal by correlating it to itself + a time lag. As the lag increases, the correlation gets weaker.
ACF can be used to approximate neural timescales.
reading some older papers without any kind of uncertainty estimates reminded me of this nice figure. inspiration of what to show in plots comparing traces between two conditions. (figure from: )