Speech Transcript - A FAST SOLUTION TO ROBUST MINIMUM VARIANCE BEAMFORMER AND APPLICATION TO SIMULTANEOUS MEG AND LOCAL FIELD POTENTIAL

0:00:15	okay so um the next speaker is
0:00:18	and robert rubber smith
0:00:19	oh she will be presenting
0:00:21	a a a a method
0:00:23	for robust minimum variance beamformer and its application
0:00:28	to um and E G and
0:00:30	a little potential
0:00:54	i can start
0:00:55	a
0:00:56	the um that top like a to see it
0:00:59	um
0:01:00	okay a i it
0:01:01	that's that some work which we do doing in oxford
0:01:05	between a of us
0:01:07	um
0:01:07	so i i mean miss amy
0:01:09	is the map rotation any he's done most of the work on the details the beamforming
0:01:14	so i look a fast for that
0:01:16	we also working with um the pulse like a tree
0:01:20	um i an engineering
0:01:22	that's made
0:01:23	and speaker i Z is is one of the D sections
0:01:26	work in they uh that if possible source just or not
0:01:29	in the brain stimulation
0:01:32	this is a the based in a sense
0:01:35	we've got a clinical problem
0:01:36	um which is to do with the brain stimulation
0:01:40	and we are trying to image it's using a make and and stuff a
0:01:44	so i'll start with a base
0:01:46	discussion about um
0:01:47	and then all want to the beamforming and then out of use some results
0:01:53	a just a fish
0:01:59	that that stimulation is technology
0:02:02	web i
0:02:03	oh they plan electrodes into the brain
0:02:07	so that the here
0:02:09	you can see a
0:02:12	that that coming in a top
0:02:14	coming coming down
0:02:17	it's used
0:02:18	white and not your the sources
0:02:20	in particular we you it's not with problems
0:02:22	see
0:02:24	for the most common
0:02:25	uh those that eating with M are that is press it's channel
0:02:29	and eating with a a a a a motion problems slide stand here like can easy easier
0:02:34	the
0:02:35	case not can be talking about that is pain
0:02:39	and we say that a uh is that you've a person
0:02:41	and
0:02:42	well pain is for the the perception of pain is all sixteen is subjective
0:02:47	some people see it could very well with pain or not find a right
0:02:50	handle
0:02:52	um it's also use some of the
0:02:54	um areas such as much as its is not in your
0:02:59	so no one yeah
0:03:00	stands for it works
0:03:02	what you do is you plot the election in the brain
0:03:05	and will be here brain stimulation the of signals actually coming out of N be very small
0:03:10	the electrodes implant
0:03:13	they then put on a five vote
0:03:15	sine wave normally five votes like very bits
0:03:18	at a frequency to about fifty hz and the functions that a
0:03:23	um and then a a a go through period of this a a a lecture it's by search
0:03:28	that's then a period where the external eyes
0:03:32	well you might access to the stimulates to
0:03:34	and S the signals coming off the implanted electrodes the local field potentials
0:03:38	and during that time they tried rate
0:03:40	so they one the site that's signal of the patient
0:03:43	and then off to that they employ a battery
0:03:45	and everything missing the was down inside everything becomes internal
0:03:50	the what we do is very trying to get a of that text last period
0:03:54	so we can actually the feel potential of the electrodes
0:03:57	and use the
0:03:58	to try to improve a
0:04:02	so mad ninety two and stuff lot of a
0:04:04	a thirteen new technology
0:04:06	um
0:04:07	picking up
0:04:07	a magnetic fields of you were have to um
0:04:11	gauss
0:04:12	so that that's
0:04:13	about three more three um eight or the magnitude smaller than the spec
0:04:17	feel
0:04:19	so essentially put a number of my comment isn't but you might as rather here
0:04:23	got
0:04:23	doesn't
0:04:24	a she
0:04:25	you can see what here
0:04:28	and for these very small signals
0:04:30	then we tried to um we can sort the sources within the brain
0:04:35	and a point of that may is to so
0:04:37	uh one is to improve
0:04:38	the surgery so they it that what in four
0:04:41	and the other one is to try to understand more about the you it's
0:04:48	so that's what are difficult to use
0:04:50	with D B S
0:04:53	one is way looking at that very small signals
0:04:56	we expect most of the
0:04:58	excitation of the brains to come from the a region around the a lecture
0:05:03	well fortunately
0:05:04	that's quite good separation between the frequency at which are the electrodes
0:05:08	and the frequency that which principle
0:05:11	so it is possible to try to access that's small signals want to me like sing
0:05:16	so we looking at about say fifty to a hundred thirty hz stimulation
0:05:20	and the low cell
0:05:21	um a a range between about five and fifteen a
0:05:24	for the range from
0:05:28	yeah that that quite big it's you
0:05:30	is when you bring the wires out
0:05:32	they are have to do a lot and skull
0:05:35	uh the wires a magnetic you got cost of wires are yeah the whole of the skull than not done
0:05:40	inside
0:05:41	but those double for distortion
0:05:44	a that region in terms of your source
0:05:46	or
0:05:52	so that think in this paper a is where
0:05:55	um got two things it using a placement of an now now that a whole
0:06:00	we
0:06:00	a a well on that show how to improves
0:06:02	the recovery of the of the
0:06:04	spatial source
0:06:07	what we think to take that in this paper
0:06:10	it's it's used in the idea that want all simulation is on
0:06:14	then it's
0:06:15	you would expect most of your signal
0:06:17	so yeah the stimulation can see
0:06:19	be coming back from a region
0:06:21	and which stimulating
0:06:23	and hence by splicing the cross correlation
0:06:26	between the signal you putting in
0:06:28	and the signal you're thing
0:06:30	at my comment is
0:06:32	then any fact you can improve your beamformer
0:06:35	especially one region of interest
0:06:37	which is found the region whether you like to tune
0:06:41	so the idea here is way looking at stimulation on and trying to get the best stuff beam as we
0:06:46	can
0:06:47	and then but using
0:06:48	time and stimulation is all
0:06:50	to try to verify that
0:06:53	so what i do i'm elements is they stimulate the time
0:06:58	um
0:07:00	people don't seem to get a
0:07:01	uh used to that's
0:07:02	T where king
0:07:04	but there are it's use a at that time
0:07:07	so that just cost about ten thousand pounds
0:07:10	um
0:07:10	and they lost from a couple of years that's what made thing not to take infection
0:07:15	the um
0:07:16	put in you new battery
0:07:18	yeah that is you
0:07:19	which is
0:07:20	coming more to like that does seem some evidence that stimulation at
0:07:23	that some of the of the range emotional areas
0:07:27	and that's some showing that actually and not to stimulate send march
0:07:31	because of things that a
0:07:32	strange
0:07:34	so a of my and fast which should welcome foundations support
0:07:37	is trying to actually um
0:07:39	find an adaptive method where we actually you know way to simulate when
0:07:43	and
0:07:45	once tried in former to you
0:07:47	very quickly um Z say that's very high meets work and i don't think that yeah
0:07:52	that you
0:07:53	and then began to demonstrate using training data
0:07:56	from a a a a a patient
0:07:57	pain
0:07:58	with P B S i mean have them
0:08:09	so a techniques
0:08:12	yeah shows um is look something like
0:08:15	so that that's uh how a bit of each
0:08:18	and to
0:08:19	um
0:08:20	a few centimetres long
0:08:22	they have a number or electrodes or may have about four electrodes on where you can stimulate
0:08:27	and
0:08:28	these are the to that the field of getting around each one is in fact a let's
0:08:32	approximately
0:08:35	so we can to assume the N F P data comes from a small cool volume
0:08:39	a bound the implant electrode we know that where that is from looking at um
0:08:43	ct a oh i want to implement
0:08:46	you can use
0:08:46	see
0:08:48	um we can we that now
0:08:49	a a a a a whole of skull
0:08:52	and then make a to use a robust minimum variance beamformer
0:08:55	um what's the to work by element set out in two thousand and five
0:08:59	but the obvious didn't use the particular aspects which relevant to um make imaging
0:09:07	yes or forward model
0:09:11	so the first time here
0:09:15	so so white state wide C R be of measurements
0:09:18	at skull
0:09:20	um S
0:09:21	is the stimulation
0:09:23	so that's from the electrode
0:09:25	and a and is the forward model leave but a vector
0:09:29	then we have a second term
0:09:31	which includes the
0:09:33	uh oh S to try to now but i
0:09:36	that's a N
0:09:37	as T
0:09:40	and we assuming a um but the N F P data is going to a small volume only got here
0:09:46	i can the location of the whole is known
0:09:48	so we can deal with that
0:09:50	and then got some noise
0:09:52	there's also in fact times day
0:09:54	going from be um lot because the instrumentation
0:09:57	between the
0:09:59	um excitation signal
0:10:00	and am signals with a
0:10:05	so just think they are a problem formulation
0:10:07	is to optimize are beam
0:10:10	for
0:10:12	um so we've got a why we want to estimate S
0:10:17	use optimisation on it
0:10:20	or or first time here is the difference between that's essentially
0:10:23	thank expectations
0:10:25	um with a value you out for a penalty factor
0:10:29	um which there is we be and one
0:10:31	so you put note you want you take in and that is
0:10:34	of of the product of a correlation
0:10:36	between the source
0:10:38	i'm the measurements
0:10:40	um
0:10:40	and then subject to all source vector
0:10:44	so that's well
0:10:46	and it to it's all
0:10:47	so we know we got source that
0:10:49	is the only condition
0:10:51	okay and then we find that
0:10:57	so that if we take the um the variances of the and take the expectations
0:11:02	then a first term does becomes
0:11:05	this one down here
0:11:07	so a why S yeah is the first good that is
0:11:10	between the source
0:11:12	the excitation we put on and the measurement
0:11:17	it's a a a a solution essentially
0:11:20	well we're trying to maximise or or correlations so
0:11:23	something to minimizing constraints such that constraints
0:11:26	you can find exact
0:11:27	use the W
0:11:29	this for increments as work that E
0:11:31	which comes to stop one
0:11:33	yeah yes one of the parameters which says what which is the described in that
0:11:37	where you putting on the excitation
0:11:40	now to me a of once month i
0:11:42	constant
0:11:44	um if we substitute are right if W back into the a constraint
0:11:48	and in fact
0:11:49	we find you
0:11:51	this is but
0:11:54	we then that no as that's
0:11:56	using the
0:11:57	normalization low
0:12:00	so be yeah so i put in this
0:12:02	i
0:12:03	um we
0:12:04	oh of the um diarization is with that i've the factor
0:12:08	are a it's a half sheet
0:12:11	a a heart sounds
0:12:12	pose
0:12:13	so that's the vector which later
0:12:15	the region of what we know the excitation is
0:12:18	to the
0:12:19	um close correlation between the measurements and the excitation
0:12:23	we can do that i was that get a secular equation
0:12:27	yeah
0:12:29	which late
0:12:30	a a a a me we found this relates number of than our on unknown
0:12:35	to the eigenvalues yeah i
0:12:38	of all normalized
0:12:40	matrix
0:12:41	or a of how cute
0:12:46	so the details of that are in the paper
0:12:48	um and the was in the paper and so i am of a fast of it that that's
0:12:52	time is and results
0:12:53	um and i have got some papers is here no once
0:12:56	but a big advantage the way that actually be normalized this
0:13:00	so we've normalized here
0:13:02	um
0:13:03	talk about V before
0:13:05	so E and
0:13:07	uh
0:13:11	i
0:13:12	with the to some of those covariance matrices
0:13:14	and also the parameters of the of is what we put X
0:13:17	i
0:13:21	is actually uses as a nice solution which then amenable able to an effective
0:13:25	solution to have find
0:13:27	um and uh
0:13:28	or as some previous work has a a few a given bounds for them the
0:13:32	this kind of normalization as she allows as to optimize and you know
0:13:39	yeah
0:13:39	a out a which allows us to do that
0:13:41	which are not been to go to i think
0:13:43	time
0:13:44	that that used to some themselves
0:13:47	so we've done this in simulation one of the problems of course as we all know made doing things well
0:13:52	mel by medically
0:13:53	um it is very hard to try to validate
0:13:57	um
0:13:57	so we done validation using a simple simulation of a spherical head model
0:14:02	put in the deep source
0:14:04	simulates the excitation station
0:14:06	we put see in the source of the but of interference
0:14:09	um and with that's noise
0:14:12	and to sure that we can to look up on and all conditions
0:14:16	we've allowed to all source
0:14:18	um which would only be
0:14:20	i dominate dominated
0:14:21	by the stimulation we on
0:14:23	or we should be seen was once the brain when off
0:14:26	a different frequencies you sign "'cause" on way
0:14:30	so yeah that was with the start of wiener filter a
0:14:33	and we've compared them with the filter so in both whole now without correlation
0:14:40	so yeah it is just the interference and their noise
0:14:44	i i mean is
0:14:46	um because it's that we've
0:14:48	simulate interference long that thing else be be estimated it
0:14:52	um we can see the wiener filter
0:14:54	is that you not
0:14:55	um a well
0:14:58	um um
0:14:59	as well
0:15:01	improves things the the S I
0:15:03	but that but
0:15:04	using the be
0:15:09	so if that the S noise as well as a fair comparison
0:15:13	that we can see a method that she does give us advanced just
0:15:16	oh the the other two men said putting in this post correlation so um
0:15:19	um between of the source
0:15:21	and the measurements since we know sources is doing that frequency as she does hell
0:15:28	a technical data is most interesting
0:15:30	um of the forty or or one with a body pain wanting pain had separate yes
0:15:36	um in but in the pack
0:15:38	which is close down to them
0:15:41	oh and fifty hz
0:15:43	so the it and sell them on magnetic
0:15:46	so they are a problem
0:15:47	for the mac
0:15:49	uh
0:15:50	but in fact
0:15:51	the
0:15:52	say whether whether wow clusters then you can still
0:15:57	do you have an all i can just and that's
0:15:59	before four we assessed the lectures
0:16:01	for sequences like to
0:16:03	um so i are you can use for but not often
0:16:06	you got inside you
0:16:08	and that just can't of is that um
0:16:11	using the he was first right
0:16:13	this my
0:16:17	the things we do
0:16:19	first of all
0:16:22	we
0:16:23	recall yeah that P data in all conditions
0:16:27	so that's that she recording straight of the electrodes
0:16:29	the that are too long used a beamformer
0:16:32	um
0:16:33	and then we were what happened all
0:16:35	so we use the beam forming one number you know
0:16:39	we then be constructed using a P point how much is what we would expect to see from
0:16:45	the yeah it is maybe a filter
0:16:48	yeah it is
0:16:49	using a a a a as well
0:16:51	a use of the um i three
0:16:53	not very much
0:16:54	and here it is
0:16:56	using you a
0:16:57	B
0:16:57	and you be
0:16:59	so you can see you press that i think that that i
0:17:01	um i think this is a much better
0:17:03	the U B
0:17:07	finally
0:17:08	the a point of this
0:17:09	try to find out
0:17:10	a in the brain
0:17:12	so this is rate just the fit it interest as suppose
0:17:15	and this is a get the difference
0:17:17	between the
0:17:18	stimulation of
0:17:19	one
0:17:20	so this woman that
0:17:21	a once it was turned off
0:17:23	and you didn't
0:17:23	on on
0:17:25	um not an entirely goods
0:17:27	um experiment because she stopped yes you knows this awful
0:17:30	do
0:17:32	you a you do get some differences between the of a young condition
0:17:35	and the way in which are response
0:17:38	this a you looking at part should known to be sensitive to pain
0:17:41	um um rate standard high
0:17:44	a i'm the A C C is the and tear single call
0:17:47	yeah right
0:17:49	and the three and i is the um of the email
0:17:54	so this technique can see that point
0:17:56	a
0:17:57	spatial layers grain
0:17:58	um where we getting sports
0:18:00	to be um nipple stimulation
0:18:04	so we we've also a of using correlations that all closely and identify more closely spatial regions in right
0:18:11	um i i a and the results if five using elliptical volume of the yeah that's P
0:18:16	to improve things of using circular volume
0:18:18	still provide
0:18:19	solution
0:18:20	and we've shown some results
0:18:22	um
0:18:23	improvements using simulation
0:18:24	i mean and you
0:18:26	and you
0:18:40	yeah
0:18:41	and
0:18:42	yeah
0:18:44	well
0:18:47	G
0:18:49	as much
0:18:50	so
0:18:51	or
0:18:52	or a more you are still going on
0:19:01	or you are you
0:19:03	two more
0:19:04	yeah
0:19:05	well
0:19:08	and
0:19:10	i
0:19:11	oh
0:19:14	or like to call
0:19:16	or
0:19:20	yes may just the problem
0:19:21	say that's that
0:19:22	for my my kinetic model of the brand such a simple
0:19:26	"'cause" you can assume that is a of a two D is
0:19:29	is one is what
0:19:30	to
0:19:31	so
0:19:32	the
0:19:33	oh well as for a um but it's not a context could you what a much more complex model right
0:19:38	so my a standard are not experts what they don't E G my understanding is make use gives you what
0:19:44	oh
0:19:45	that is to discuss that with you for are actually that if anyone and any information on that
0:19:49	that's a my hands and you i mean it is a recent technology is come about lost five to ten
0:19:53	years
0:19:54	well a you also
0:19:55	to machine
0:19:56	and not let me
0:20:04	oh
0:20:06	what should you signals
0:20:10	tricks
0:20:10	of
0:20:13	what was is
0:20:15	are you sure from your speech and so
0:20:21	or
0:20:23	but
0:20:24	for a the S T N was a packet is next to it
0:20:28	is
0:20:30	but pain yeah
0:20:31	the use S the end from it
0:20:33	oh so you're right
0:20:37	trying to be your with peace
0:20:40	for four
0:20:42	what what you saying
0:20:43	for the final are we measure yeah that peaks as X to i and then we tried to estimate as
0:20:47	well
0:20:48	yes using the mac
0:20:50	in a compared
0:20:51	so
0:20:54	you
0:20:55	for point
0:20:57	well
0:20:59	yep
0:21:00	so what is
0:21:02	right
0:21:03	for
0:21:04	she
0:21:05	a
0:21:06	all
0:21:08	the marks is
0:21:10	i five to
0:21:11	i i mean that sense that to the signal
0:21:14	we receiving see in the of
0:21:15	or you
0:21:16	a frequency
0:21:17	a point where we stephen it is that you want place that
0:21:21	is quite hard to know how you try to optimize
0:21:24	world i mean would really strong
0:21:27	so i
0:21:29	one Q one
0:21:31	so
0:21:32	still
0:21:34	from
0:21:35	and
0:21:36	estimate
0:21:36	um
0:21:37	right
0:21:38	which still
0:21:38	what we are we are doing a spectral analysis as well so we looking
0:21:43	so a beamformer is that you
0:21:46	i i think that if
0:21:47	for
0:21:49	so we all by some that to look at
0:21:51	maybe to right
0:21:52	the B two
0:21:53	of what's it's yeah was either
0:21:58	and yeah the
0:22:06	source you right hmmm part
0:22:11	to
0:22:12	ooh
0:22:13	yeah a just that even that's one
0:22:15	but
0:22:15	that's what if you they get
0:22:18	sure one
0:22:20	uh_huh four
0:22:23	do
0:22:24	we would
0:22:25	you
0:22:27	hmmm
0:22:28	it's a that's not in a sense
0:22:30	that are not quite so what you're looking for
0:22:32	seven seven sense i mean it is that's what we try to estimate we getting and then measure we get
0:22:36	"'em" and i suppose way we can possibly indicate
0:22:39	still
0:22:40	one thing to get a good results or
0:22:42	but it's good question and very hard to know how about it
0:22:46	from
0:22:47	see
0:22:49	you try to the um
0:22:51	so i mean
0:22:55	okay well thank you everyone
0:22:57	uh i think you again maybe

A FAST SOLUTION TO ROBUST MINIMUM VARIANCE BEAMFORMER AND APPLICATION TO SIMULTANEOUS MEG AND LOCAL FIELD POTENTIAL

Biosignal Estimation and Classification

Presented by: Penny Probert Smith, Author(s): Hamid Mohseni, Morten Kringelbach, Mark Woolrich, Penny Probert Smith, Tipu Aziz, University of Oxford, United Kingdom