Lecture 10: Frequency-Based and Advanced Analyses in MEG Flashcards
1
Q
Last week (L9) we covered
A
event-related analyses
2
Q
There are two more ways to analyse MEG activity, aside form event-related analyses, such as - (2)
A
- Frequency and time-based approaches
- Multivariate approaches
3
Q
Aside from analyse MEG activity, we can also analyse the
A
functional connectivity in MEG
4
Q
The event-related, frequency and time-frequency approaches of MEG data analysis as well as functional connectivity can - (2)
A
be done in sensor or source space
although connectivity analyses more meaningful in source space
5
Q
MEG signals is ___ and measured in __ whilst noise (artefacts) are __
A. small, milli Teslas, big
B. small, femtoTeslas, big
C. big, picoTeslas, small
D, big, Teslas, imaginary
A
B. small, femto Teslas, big
6
Q
MEG data consists of a matrix (database) of what by what
A. Time by magnitude
B. Magnitude by sensors
C. Time by sensors
D. Sensors by Versace
A
C. time by sensor
7
Q
Which of these are important sources of external noise in MEG data? - (2)
A. Earth’s magnetic field and mains electric
B. Mains electric and passing satelites
C. Earths magnetic field and other planets
D. Passing traffic and other people
A
A Earth;s magnetic field and mains electric
Passing traffic would be source of noise but other people wouldn’t be source of external noise
8
Q
Which of these are the most important internal sources of noise? - (2)
A. Breathing and heartbeat
B. Heartbeat and swallowing
C. Blinking and heartbeat
D. Blinking and screaming
A
C. Blinking and heartbeat
Breathing isn’t main ones unless moved around a lot
9
Q
Low-frequency scanner drift may be removed by a low-pass filter
True or False?
A
False –> because low pass lets low frequencies through
10
Q
High-frequency mains artefact may be removed with a low pass filter
True or false
A
True –> low pass only lets lower frequencies through
11
Q
A bandpass filter only allows a small range of frequencies through
True or False
A
False –> la band lets a chunk through and notch fit gets rid of small range
12
Q
Which of these preprocessing steps should be done first?
A. Filtering
B, Panic
C,Automatic artefact removal
D, Bad channel removal
A
D. Bad channel removal
13
Q
It is only appropriate to use event-related analyses if we have?
A. Evoked responses
B. Induced responses
C. Either evoked or induced responses
D. Positive responses
A
A. evoked responses
14
Q
MEGS are like EEG ERPs expect they have a more variable
A. Magnitude
B. Personality
C. Time
D. Direction
A
D. direction - can be positive or negative
15
Q
Electric potential/magnetic field strength plotted acorss time in MEG and EEG gives us
A
osciliations -> up and down over time
16
Q
Oscillations have
A
cycles - up and down and back up again
17
Q
Location in a cycle is called a
A
phase –> 4 different points
18
Q
Number of cycles per second is equal to
A
frequency in Hertz
19
Q
If we have a 5 Hz MEG/EEG signal then it means
If we have a 10 Hz MEG/EEG signal then it means
- (2)
A
Going at full cycle 5 times per second
Going at full cycle 10 times per second
20
Q
We can identify brain activity at different frequencies with
A
MEG or EEG
21
Q
Diagram of the different frequencies bands that different brain activity can occur in - (2)
A
Delta 2-4 Hz to low and high gamma
MEG/EEG Data can look at the top - broadband signal which we be broken down into different frequency bands
22
Q
Diagram of evoked, induced responses or neither and explain them and analysis to do them - (3)
A
evoked = same response happening on different trials at exactly same timel and peak same time on all trials = time locked (happening at same time) and phase locked (phase -top of peak in all trials) –> event-related
induced responses = time locked ish as all responses happening after lets say 100 ms after stimulus onset but at differet points of a phase (not phase-locked) – >frequency based analyses
Neither =after 100 ms saw response in one trial but another trial saw response at 150 ms and another trial response at 50 ms
23
Q
Event-related analyses average over the time course meaning they
A
analyze in the time domain meaning better signal than noise but likely to miss responses that are not evoked (i.e., induced or neither)
24
Q
Frequency-based analyses do not average the time course and able to
A
detect any of these response types - like induced , evoked or neither responses
25
Frequency-based analyses in simple versions just ignores - (2)
time entirely and use whole trials
(e.g., looking trials of seeing face and looking across all of the trials and seeing its frequencies)
26
Frequency-based analyses simplest version uses the
Fourier transform
27
The fourier transform calculates the - (2)
average power at each frequency across trials
(strength of oscillations at different frequencies range e.g., alpha, beta)
28
The outcome of fourier transform is a
power spectrum
29
Diagram of power spectrum
30
What is on y and x-axis of power spectrum? - (2)
Y axis is power (which is amplitude squared of magnetic field strength)
X axs is frequency Hz
31
The power spectrum can be for a
whole sensory array or a specific sensor
32
Frequency-based analyses 's power sectrum can be
compared across conditions (within-subject) or participant groups (between subject)
33
Fourier transform in Frequency-based analyses ask
in what frequencies are there changes in magnetic field strength
34
Fourier analysis in frequency based analyses asks:
in what frequencies are there changes in magnetic field strength
This is different question from event-related analyses as
Frequency-based analyses using Fourier transform is not saying when is there a change in amplitude of magnetic field strength but saying in what frequencies are we finding activity (e.g., finding activity for faces at 2Hz)
35
# word
Describe power spectrum graph - (2)
example of 1/f finding that brain activity occurs in lower frequencies than high so this slope down in high frequencies
Also getting big spikes in 9 Hz of alpha band activity and 15 Hz of beta band activity - which can compare across conditions (e.g., face trials or tool trials) or participants grps
36
In frequency-based analyses using Fourier transform we are analysing in .. which means ... - (2)
Analysing 'in frequency domain' (not 'time domain')
Since we are looking at every frequency
37
Analysing in 'time domain' is what type of analysis?
Event-related analyses
38
In frequency based analyses using fourier transform
effects don't have to be....
can detect what kind of responses....
can even do this with .... - (3)
1. Effects don't have to be phase-locked or time-locked to an event
2. Can detect evoked, induced or neither responses
3. Can even do with runs of resting-state data
39
Frequency-based analyses response types do not have to be
time-locked or phase-locked
40
Time-frequency analyses looking at the frequency changes across time in
each trial
41
Example of time-frequency analyses
Around 100ms is there osciliations in alpha band, around 500 ms is there osciliations in gamma band
42
The outcome of time-frequency analyses is a
time-frequency plot
43
What is on y-axis and x-axis of time-frequency plot and colour? - (3)
Y axis is freuency Hz
X axis is Time (s) - (-1 = 1s before stimulus, 1 = 1 second after stimulus)
Colour is power of osciliations at specific frequencies (e.g,, can say at 100ms how strong is 1 Hz frequency)
44
Time-frequency analyses ask the questtion of
how does activity in different frequencies change over time OR over time what frequencies do we have osciliations in
45
The time-frequency analyses, red, blue and black in time-frequency plot represents... - (3)
* Red = positive freq change
* Blue = negative freq change
* Black = no freq change
46
Time-frequency analyses uses a similar method to fourier analysis (breaking things into different frequencies) which is...
more suited to shorter time windows
47
Time-frequency analyses uses a similar method to Fourier analysis, which is more suited to shorter time windows, called
wavelet decomposition
48
Time-based freqency analysis does not include
whole trials
49
Time-frequency analysis using wavelet decomposition analyses both in
time and frequency domains
50
The wavelet decomposition is performed using
each consecutive time-point
51
The wavelet decomposition in time-frequency analysis is different to frequency-based analysis Fourier analysis as
responses do need to be time-locked to an event
52
wavelet decomposition does not need to be ... but needs to be
phase-locked but needs to be time-locked
53
In wavelet decomposition in time-frequency analysis still discard phase term so when we
average the power in each frequency across trials so responses do not have to be phase locked
54
In time-frequency based analysis wavelet decomposition we average the power in each frequency across trials but do so
per time so respones do need to be time-locked to an event e.g., looking at responses after 500 ms of stimulus
55
Time-frequency analysis wavelet decomposition can detect what responses
evoked or induced responses
56
What does this time-frequency plot show? - (2)
In this example we can see an initial low frequency response (the MEP ), and subsequent bursts of gamma activity at higher frequencies (50 – 100Hz) when faces shown - induced response of gamma band bursting
Event-related analysis - evoked at bottom
57
Time frequency analysis wavelet decomposition can be used to be
compared across conditions (within) and participants groups (between subject)
58
In both time-frequency and frequency -based analyses we
we take frequency of each osciliations and its power then averaging that across trials
59
Why might we use frequency-based analyses? - (2)
When we want to know the frequencies involved
e.g., because we think osciliations at this frequency are meaningful
60
Different frequenci band can be linked to - (2)
different cognitive processes
e.g., gamma bust leading to -coherent visual perception seeing dog or not
61
Different frequency bands are typically localised to
different brain regions - source localisation
62
Example of different frequency bands associated to different regions of the brain - (3)
delta usually in anterior temporal lobe
alpha associated with alertness and visual comes from back of brain
gamma comes from medial prefrontal cortex but gamma is variable and come from lots of places
63
Different frequency bands can be linked to different functions and brain regions
Have to think about
Are these the cause of behaviour (do they tell us something about mechanisms of how the process works) or are they an epiphenomena of brain activity (like BOLD)?
64
Other reason for using frequency-based analyses other than When we want to know the frequencies involved e.g., because we think oscillations at this frequency are meaningful
is that we want to know
when we thinkwe have responses that are not phase-locked (not evoked - induced or neither responses [+not time-locked] ) that would be miseed with event-related analyses
65
We do frequency0based analyses when When we (think we) have responses that are not phase-locked (not evoked), that would be missed with event-related analyses
If responses are time-locked(induced) we can look at them over time with
time-frequency analysis
66
We do frequency-based analyses When we (think we) have responses that are not phase-locked (not evoked), that would be missed with event-related analyses
If responses are not time-locked( induced) or phase-locked - neither responses then do
standard frequency analysis
67
Fourier analysis/transform is a mathematical therom that
specifices any waveform can be described as a sum of basic functions like lego bricks
68
The basis functions are the
sine waves of different frequencies
69
The fourier analysis lets us break down a single to
its component sine waves at different frequencies (e.g., red sine wave at 1 Hz and orange sine wave at 2 Hz)
70
Diagram of Fourier analysis/transform - adding sine waves together - (2)
Add up a sine wave of low-frequency and add up a sign wave of higher frequency which approximate data as well as possible
We can do a lot more than adding 2 sine waves, at different frequencies we can decide how much of each is included = power of osciliations at each frequency
71
Diagram of Fourier-Analysis transform - how to find recipe of smoothie
72
Fourier analysis can make very different and highly complex functions by combing many sine functions such as
can approximate a step function - on and off in square boxes by adding more and more of sine waves of different frequencies
73
Fourier analysis works best on longer segments of data (e.g., full trial) as
expects sine waves to continue throughout
74
To look across (tiny) time points within time window of trials, we often use a .. instead of fourier analysis
wavelet decomposition
75
The wavelet decomposition uses - (2)
moretemporally limited basis functions called wavelets (short waves) that taper off smoothly to either side
At top high frequencyes and at bottom low frequencies
76
The wavelet decomposition slide across
time window to see how much power in that freuency around that time
77
Diagram of summary of frequency-bases analyases
Red is standard frequency analysis and green is time-frequency analysis
78
Frequency and event-related analyses are all
univarate approaches
79
Univarate analyses look for
difference in mean activity betwen condition A (e.g., images of bird) and B i(e.g., images of furniture) ndependently for each sensor or vertex in source space
80
The problem with univarate analysis might miss a region that processes and represents whether input is
A or B stimuli if it does not simply have more activity throughout (or all sensors/vertices might be important - it's pattern across all sensors/vertices )
81
Multivarate analyses ask what info is represented in signals we measure in other words
is pattern different for conditions A or B or are there meaningful differences between each stimuli
82
In multivariate approaches we use the pattern across ... instead of the amount of change in each sensor or vertex.
multiple sensors (or vertices)
83
In MEG we can look at the information represented
across time (and space)
84
Diagram of univarate, multi-variate approaches(MVPA - classifiers and RSA) in MEG
85
Pattern classification in MVPA involves training a computer algorithm called a
classifier
86
Pattern classification trains a classifer to distinguish between
conditions based on activiy in many sensors/vertices
87
After training a classifer, we test its accuracy on
eparate test data (e.g., correctly labels red squares and blue circles)
88
If the classifier can predict the different conditions - red and blue better than chance (50% = since 100/2) in test data,
then there must be..
suggests that.. - (2)
there must be differences in brain activity for the two conditions there must be differences in brain activity for the two conditions
Suggesting the brain or brain region processes them differently, perhaps identifying them as different categories
89
The classifier accuracy on test trials give some indication of
similarity between the brain’s response in two conditions (although highly dependent on algorithm details)
90
In pattern classification it tests the classifer accuracy on
separate data (data not used to train classifer)
91
Pattern classification in MEG (since it has great temporal resolution) will give
timecourse of classifer accuracy -- > seeing how info changes over time
92
What does this diagram of pattern classification across time graph show? - (3)
Classification was done on the pattern of activity across sensors
Gives us a timecourse of classifier accuracy of classifer whether stimuli was object or living thing
So we can see info changes over time
93
In representational similarity analysis, it involves
lots of different stimuli and possible to calculate how similar the neural response is for each pair of stimuli
94
Representational similarity analysis produces a
data representational dsimilarity matrix
95
In RSA, we can compare the data representational dssimilarity matrix with
model representational disimilarity matrix
96
In RSA, tests the correlation between data representational disimilarity matrix to model representational disimilairty matrix and if its high then
brain contains similar info to that in your tehoretical/behavioural model
97
In RSA, can do it
over time for MEG for different ROIs
98
RSA over time is a modality-invariant measure, so we can also compare it across different imaging techniques and with computational models
99
Representatational similarity analysis (RSA) can combine
MEG and fMRI data
100
Functional connectivity looks at how different brain regions in MEG
communicate and transfer information
101
Diagram of looking at functional connectivity in fMRI
in fMRI, take time series in one ROIs and time series of another ROI and correlate them in simple version (if this ROI goes up then another ROI goes up if so maybe talking with each other)
102
Functional connectivity analysis in MEG is closely related to
fMRI measures of connectivity , with greater temporal resolution
103
Functional connectivity analysis in MEG closely related with fMRI measures of connectivity, with greater temporal resolution - (2)
Potentially improves measurement and gives extra information (time lag, better directionality estimates, frequency…) - does both ROIs turn on same time - are they talking to each other?
But adds a level of additional complexity - not simple correlation in fMRI connectivity
104
In MEG we also have freuency-based measures which can also be used in
functional connectivity analyses in MEG as well
105
In MEG and fMRI functional connectivity analyses remember that
Remember isn’t structural (anatomical) connectivity and needs to be understood in the context of the structural connectivity
106
In correlation-based approaches of functional connectivity
suppose we measure activity at locations A and B and correlate two signals - (2)
If they are highly correlated, they might be representing similar information
But generally signals take time to get around
107
Correlation-based approaches for functional also gives insight as to what the direction
of two signals are --> interaction (not causality)
108
As compared to correlation-based approaches of functional connectivity a more informative approach in MEG is - (2)
cross-correlate the signals with a range of ‘lags’ (delays) – get timing information as well and choose one with strongest correlation
Also start to get clues as to the direction of the interaction (but not causality)… --> shift to Granger causality
109
Granger causality is no longer a ... approach/method in functional connectivity in MEG
correlational
110
The granger cusality uses directed cuasality which let us
infer the direction of information flow and speak about one region caually influencing another
111
Grangier causality could be thought of in two ROIs( A and B) - (2)
past activity in A predicts the future activity of B (more than B predicts futre of A)
Think of asking does the past activity in A add something more to the prediction of future B than the past of B alone?
If so we can say locaiton A is driving the activity in location B
Granger causality lets us quantify this
Can call ‘effective connectivity’ as causal
112
Conditional granger causality factors out the - (2)
effects of other areas to ask if this is a direct influence (e.g., A influences B, B influences C so will A influence C?)
But only compares to other specificed ROIs (not everywhere)
113
(Conditional) Granger causality has clinical utility for example..
tracking how seizures propgate across the brain
114
In within-frequency connectivity we are interesed in.. or ( 2)
Sometimes connectivity is mostly within a specific range of frequencies
Or we are interested in what frequency regions are interacting at
115
In wtihin-frequency connectivity can be conducted as
we can filter the data first, and just look at the connectivity in a single frequency band or set of frequencies (can do multiple and compare
116
In within-freuency connectivity
networks often correspond to activity in a specific frequency band e.g.,
the default mode network typically involves the alpha band
117
Frequency-based connectivity analyses coherence is
ommonly used and like correlation but in frequency domain
118
Correlation only considers the
relationship between changes in amplitude between two regions
119
Coherence also considers
the extent to which their phase relationship is consistent
120
Coherence means in an example
the phase (or peak) in one signal predicts the phase (or peak) in the other with some consistent lag time
121
Frequency-based connectivity analyses coherence can be computed at a
given frequency (can summarise across bands/all frequencies)
122
Cross-frequency coupling in MEG proposes that - (3)
there could be a relationship between regions that don’t have the same frequency
Could be between any measure in region A and any measure in region B
Trickier to think what this may mean
123
One popular method is
phase amplitude coupling
124
Sometimes the amplitude of high frequency oscilaations changes over time according to
a lower frequency
125
Phase-amplitude coupling proposes that
amplitude of the higher frequency depends on the phase of the lower frequency
126
Phase-amplitude coupling happens
both within and between brain regions, and may be a key mechanism for transmitting or ‘gating’ information
127
Many different ways to assess activity and connectivity in MEG but..
don't get carried away looking for a real effect and want a specific data plan including hypotheses to test and consider prereg these in advance and replicating results after
128
What are the two frequency-based analyses? in MEG
1. Frequency-based analyses
2. Time-frequency Analyses
129
130
What are the two multivaraiet approaches in MEG?
1. MVPA
2. RSA
131
What are the functional connectivity methods in MEG?
1. Granger causality
2. Within-frequency connectivity
3. Frequency-based connectivity coherence
4. Cross-frequency
5. Phase amplitude coupling
132
Is cross-frequency coupling and phase amplitude connectivity analyse same in MEG?
In summary, while both CFC and PAC involve the interaction between different frequencies of neural oscillations, PAC specifically addresses the coupling between the phase of low-frequency oscillations and the amplitude of high-frequency oscillations, whereas CFC is a broader concept that encompasses interactions between the phases or amplitudes of oscillations across different frequency bands
133
Difference between RSA and MVPA
RSA does not directly use pattern classification in traditional sense as focuses on similarity or disimilarity structure of neural representations
MVPA uses pattern classification and uses machine learning algortihms to decode patterns of neural activity associated with exp conditions
While both RSA and MVPA are used to analyze patterns of neural activity, MVPA is explicitly focused on classification or decoding, aiming to predict experimental conditions, whereas RSA is more concerned with understanding the structure and relationships within neural representations without a primary focus on classification or prediction.
134
With regard to phase-locked and non-phase-locked responses, which of the following statements is true?
A. To look at phase-locked responses, we can average the spatial maps, whilst when analysing non-phase-locked responses, we should instead average the timecourses.
B. To look at phase-locked responses, we can average the timecourses, whilst when analysing non-phase-locked responses, we should instead average the spatial maps.
C. To look at phase-locked responses, we can average the activity in each frequency band, whilst when analysing non-phase-locked responses, we should instead average the timecourses.
D. To look at phase-locked responses, we can average the timecourses, whilst when analysing non-phase-locked responses, we should instead average the activity in each frequency band.
D.
135
What information about the data do coherence analyses consider that correlation-based functional connectivity methods do not?
A. Power
B. Phase
C. Amplitude
D. Tmiing
B. Phase
136
If you were performing an MVPA classification analysis in which you had the categories faces, scrambled faces, objects, houses and places, what would we expect chance level to be?
A. 25%
B20%
C. 15%
D. 10%
B. 20% since 100/5
137
In which of these MEG analysis situations do we NOT typically use ‘power’?
A. plotting the strength of osciliations at different frequencies
B. summing the total amount of change in the magnetic field across the brain
C. plotting the timecourse of a trial
D. estimating the amount of artefact at a specific frequency
C. plotting the timecourse of a trial
