Magnetoencephalography II Flashcards
What is brain oscillation?
- Rhythmic or repetitive neural activity in the brain
Not just combination of sinus over time
Can get distinct frequencies out of it - Five major bands of oscillatory frequency:
Delta waves: 0-4 Hz
Theta waves: 4-8 Hz
Alpha waves: 8-12 Hz
Beta waves: 12-40 Hz
Gamma waves: 40 Hz and above
Mu = high alpha band = motor activity
Beta and gamma can be split into low and high - Many task-related dynamics in MEG data that are retrievable using only spectral analysis approach because it captures more multidimensional features of brain processes. Changes of neurona bvr
A neuron being bombarded with inputs will give an output but will also inhibit itself a but and this creates the oscillations
= recussing self-inhibitory loops
Periods of activity followed by periods of non activity creates the rythmic cycles of activity
What are spectral responses?
Spectral analysis aims to decompose MEG signals into the (time-) frequency domain.
Can get the power of the frequencies or can do frequency transforms oer small frequencies of time
You can extract these components in a few different ways
* Fourier transform – assumes MEG data are static the in frequency domain and gives you the frequency components but you lose timing information
Decompose any signal complex into a finite set of sin and cosin waves of all diff f and if build up a set of those, you can use the set to tell you what fs are in the signal
- Wavelet transform – accepts the temporal convolution of frequency responses and gives you frequency information and time information
Get wavelets (sin waves) and convolve it with a Gaussian and shorten it to make high or low fs
Gives you f info but also returns some time info (there are some spots in time where its correlated
Caveats: evoked vs. induced brain oscillations
Differences between evoked vs. induced brain oscillations
* Evoked activity is locked to the stimulus, time locked and phase locked
Tends to reflect lower level or sensory processing
When stim someone, get some activity that’s time and phase locked to the stim and others are time but not phase locked so when average, you lose the non phase locked
* Induced activity appears with a latency jitter (i.e., it doesn’t appear a fixed time after the stimulus) or is time-locked to the stimulus but not phase-locked
Tends to reflect higher-level processing (more cognitive)
Ex stim = sound, causes activity in auditory cortex that then goes to frontal and then back to auditory so it gets all mixed up
Because of the latency jitter, induced activity can be absent if you average across trials before extracting frequency information.
Instead you need to extract the frequency information first and then average across trials.
Have a magnitude of every trial and then can average it even if phases are diff
What research questions you can ask with spectral analysis? (5)
- Is a brain oscillatory activity present in one condition and absent in another?
- Is a brain oscillatory activity delayed in one condition relative to another condition?
- Does the size of a brain oscillatory activity vary with some feature of the stimulus?
- Is a brain oscillatory activity present in one population and absent/reduced/delayed in another?
- Does the brain distinguish between two conditions in terms of the changes in oscillatory responses?
What are 2 limitations of spectral analyses?
- Decrease of temporal precision
In wavelet transforms, to reliably calculate the f, have to sacrifice some time, it gets smeared - “the paralysis of analysis”
When have a big dataset and do transforms that can do lots of analyses, can do too many analyses, have to have a good idea of what you want to do with the data beforehand
How do we report results for sensor level analysis? (4)
- original waveforms/ oscillatory activity of selected channels covering a region of interest
So ppl can judge the quality of the raw data - topography/contour maps of the time window of interest
See structure in the topography of the signal to see where the source analysis came from - the RMS waveforms over all sensors or the regions of interest
Increases signal to noise ratio - describe the full statistical model used
What are 8 other parameters that need attention when reading MEG papers?
- N of trials: early ERFs
The later in time you want to look for smtg, the more trials
you need bc there’s smearing so signal to noise ratio dims - ISI: interstim interval, needs to be big enough so the stims dont sum
- Baseline selection criterion
Period of time previous to the stim to normalise the amplitude or normalise what happens after the stim - Data transformation such as log-transform on the unit values of figures
- Group analysis: individual-variability of the type of MEG responses analysed, movement compensation techniques applied before grand-averaging or inter-run averaging
- Details about the head model used in source reconstruction
- Multiple comparison correction applied to statically analysis when contrasting across space/time/frequency
- effect size as well as their stats
What is source reconstruction?
One of the major advantages of MEG over EEG is that it’s easier to figure out where in the brain stuff is happening
* With EEG you need information about the conductivity of the skull and scalp because these conductivities distort the signal
By the time they reach the sensor, they’re mixed together
* MEG signals pass the skull and scalp unperturbed making the problem of source reconstruction easier
* Note that the huge dimensionality of the data allows you to infer a lot more than source location. (functional/effective connectivity analysis)
What are 3 approaches to source analysis?
Dipole fitting
Distributed source models
Beamformers
What is dipole fitting?
- Attempts to find dipoles that produce the observed magnetic field
- You need a high SNR so it is generally used for sensory-evoked activity
- It can be tricky knowing how many sources to include (made simpler by assumptions)
- Better for short time windows because then the assumptions are more likely to hold
What is Distributed source models?
- Instead of estimating the location of one or two dipoles, a large number of dipoles are distributed through the cortex and their current density is estimated
- Approaches such as MNE attempt to find the solution with the minimum power
- Some of these approaches have a depth bias and favour surfaces sources but there are techniques that can be used to minimise this bias
If the sources are further from the sensors, the signal is weaker
What is beamformers?
- Uses a spatial filtering technique. Weighting sensors in a particular way emphases signals from a particular location and attenuates signals from other locations
- Generally blind to correlated sources (but there are ways around this by, for example, splitting the sensor array in two)
If 2 parts of the brain are active at the same time with low latency, you’ll only get info from 1 of the sources - Allows you to look at induced and evoked activity so it’s good for studying higher level cognitive functions. It can be tricky knowing how many sources to include
Dont need to average the data, can look at time series
What is the inverse problem in source analysis?
In general, source modelling is an ill-posted issue for MEG and EEG because a large number of source configurations can produce the same pattern of magnetic fields (the inverse problem).
And if people do care about where do the processes take places in the brain, go for fMRI or PET
What are connectivity analyses and 4 things that we should consider?
Presume that the brain has fctional specialised nodes
Bi- vs multivariate interactions
Volume conduction
Connectivity over time vs. over trials
Directionality
Bi- vs multivariate interactions
If someone taps there finger when there’s a beep, there
should be an interaction between auditory and motor cortexes
If the person is well trained in that task, the relationship should
be stronger
