Přepis řeči - P AND T WAVE DELINEATION AND WAVEFORM ESTIMATION IN ECG SIGNALS USING A BLOCK GIBBS SAMPLER

0:00:14	and then a um next page
0:00:16	it is a a a a to B and that
0:00:19	the and in um and waveform form estimation hmmm C two signals
0:00:24	um so we present by channel
0:00:33	or
0:00:34	slow
0:00:54	so morning everyone
0:00:57	but to this presentation entitled to yeah she waved iteration and form estimation in actual hardware signals
0:01:05	but using a block keep are
0:01:06	and a name in
0:01:08	this it a joint work
0:01:09	where is my
0:01:12	pitch to live there is a professor joint one of my
0:01:15	and that's only my
0:01:16	are the universal thing false
0:01:19	and also with Q O K O and professor for a large
0:01:23	are are we is to be and the university of technology also
0:01:27	so here is to you
0:01:30	i'll of from
0:01:31	addition
0:01:32	for that we would give you a brief introduction of
0:01:35	the chance you move initial problem
0:01:38	and the seems we propose to use a bayesian inference to deal with this problem
0:01:43	and was then you should use use the proposed to but more
0:01:47	and we also propose
0:01:48	to use but you the keeps mess or to estimate a and parameters of the proposed
0:01:55	i told and that would give you details of the
0:01:59	the proposed schemes that or
0:02:01	and i sure use this measure is our to the case that a where is the operation with
0:02:07	man
0:02:09	oh here have a star
0:02:11	oh have you about in some
0:02:14	but you creation problem
0:02:17	and
0:02:18	as we all know that that our ram for C G for short
0:02:22	it it for of
0:02:23	the hard to lead to a to be G after uh by the actual position on the screen server
0:02:30	so
0:02:31	the automatic
0:02:34	and the lights is of just signal as on a lot of interest in the biomedical engineering from
0:02:41	oh problem this
0:02:42	estimation
0:02:44	as we can see
0:02:51	sorry
0:02:52	but because see
0:02:53	uh a C D signal
0:02:55	consists of three distinct waves
0:02:58	she way
0:02:59	and uh the most significant part of to a caress contrast
0:03:03	and that's she way
0:03:05	oh problem
0:03:06	this figure
0:03:07	as we can see the most
0:03:09	useful clinically useful information
0:03:12	i be file
0:03:12	the wave boundaries
0:03:14	and the interval durations
0:03:17	oh these C D C
0:03:18	but that's why the deviation
0:03:21	which means
0:03:22	the um
0:03:23	determination of the with peaks with bob are used and the estimation of we've one
0:03:28	it's a very important step to the interpretation of the C D C
0:03:34	so things that or as complex as
0:03:36	are a more significant part of the each signal
0:03:39	is detection is relatively easier
0:03:42	and
0:03:43	it can be used as a reference
0:03:46	to the C and she you deviation
0:03:49	and um
0:03:51	the C and C waves
0:03:52	generally they have a low should
0:03:55	and due to the presence of the noise
0:03:58	and a baseline one three
0:04:00	if the are deviation it's more complicated than the as complex as
0:04:04	the H
0:04:06	so basically we will
0:04:09	do firstly to dress complex this and defined a search each both before and after
0:04:14	locate look S you but
0:04:17	so we can find in the three sure
0:04:22	different
0:04:23	existing methods
0:04:25	the first
0:04:26	existing mess are the first class is based on reading techniques
0:04:29	the president
0:04:30	it is to use for techniques to remove the noise
0:04:33	and the base a wiring
0:04:35	and it
0:04:36	sorry
0:04:36	and then
0:04:38	we you apply by read and it uses specialist
0:04:41	term where are the weights
0:04:43	and we are are problems
0:04:46	is that in
0:04:47	class of existing that's sort is based on
0:04:49	basis is expansion that makes
0:04:51	i mean we we use a basis is especially at least for the functions
0:04:55	and the way to use
0:04:57	some thresholds to determine where are wave peaks and where the bar
0:05:02	a certain class of existing mess or is based on a a piece and that pattern recognition that source
0:05:08	and the we can also
0:05:10	a by some realistic signal model to feed the C D signal and to estimate
0:05:17	to parameters of the model of to do the nation
0:05:20	of the D C D C
0:05:22	so some particular patient has been given to a part of you
0:05:26	you our paper
0:05:28	we we propose
0:05:29	to use
0:05:30	uh uh easy inference
0:05:32	together with is
0:05:33	some similar to ours
0:05:36	to deal with this problem
0:05:38	so now i would introduce you
0:05:40	or a mathematical model
0:05:43	the grass complex the our application are assumed to be detected
0:05:47	so a which will not be considered here
0:05:50	and that then
0:05:51	we propose to use a deep each no overlap
0:05:54	window
0:05:56	to cover the whole dct T stick
0:05:59	so basically you we have the same number of the
0:06:03	no if are asking to roles
0:06:04	we now hour
0:06:06	process we know
0:06:07	and for each not here S components
0:06:10	can be seen as a combination of all to me the to mean is
0:06:15	which are present you way and if you with
0:06:18	plus uh local based
0:06:21	so take the that parts as an example
0:06:24	this is she weights with this
0:06:27	a uh processing window can be seen as a a convolution
0:06:31	oh
0:06:31	but
0:06:32	bernoulli gaussian sequence
0:06:34	a on the one would got the sequence
0:06:35	with a
0:06:36	on the one
0:06:37	impulse response
0:06:39	so that
0:06:40	oh no impulse response present a waveform
0:06:43	and this
0:06:44	on the one really gaussian
0:06:45	to represents
0:06:47	the with locations and we want to
0:06:50	and then we can do same to the T waves
0:06:52	so basically we have some conclusions here
0:06:55	and we have a this
0:06:57	but the medical signal more
0:06:59	and this sick it represents the sequence of the baseline a local base fine
0:07:04	and this W P A a noise
0:07:06	which use that students be gone
0:07:10	and for some more we also read
0:07:12	the proposed to use
0:07:14	or a basis expansion techniques
0:07:16	to represent a on no one point four
0:07:19	so this
0:07:20	technique has to be used
0:07:21	oh C to denoising and also on C compression
0:07:26	the advantage of
0:07:27	this technique
0:07:29	i is
0:07:30	that is that we can have a used to mention though
0:07:34	on on parameters
0:07:36	and also we
0:07:38	we can seen that noise some most the version
0:07:41	oh the with that
0:07:43	for them more we also propose to use of force degree polynomial model
0:07:49	local baseline
0:07:50	but this is a patients what previous work
0:07:53	which assumes that the local
0:07:56	baseline which in each no cresting in ball is a cost
0:08:01	so
0:08:02	with all this we have to a better representation of
0:08:06	the the not is in the roles in the processing window
0:08:10	so
0:08:11	red once are all time
0:08:14	so we have the we form we've location is up to two
0:08:18	all the what she ways
0:08:19	and the
0:08:20	baseline like is the noise was far
0:08:24	so
0:08:25	here
0:08:26	i will introduce you if you model
0:08:29	that we don't know the you for realise of the product
0:08:32	all of the that it functions and priors
0:08:36	to represent the posterior distribution
0:08:39	and to estimate and have
0:08:42	so
0:08:43	by assuming the noise in our model is zero-mean gaussian
0:08:47	we can have a just like would function
0:08:50	and concerning the priors of the unknown parameters
0:08:55	they can be assigned according to the T C you signals back to this
0:08:59	just we only expect
0:09:00	at most one
0:09:02	she wave in each to thirty the both
0:09:05	we can assign a block cost rate
0:09:07	to this
0:09:08	indicator
0:09:09	and a by assuming the independence
0:09:11	oh of each individual
0:09:14	thirteen thirteen both
0:09:16	a
0:09:16	priors of the
0:09:18	the the the whole
0:09:20	she indicator vector can be to by a product of each about divided so
0:09:27	and the since we have used a run gaussian sequence
0:09:31	so when there is a way it it
0:09:33	we assign a zero-mean gaussian prior to the
0:09:36	and you
0:09:37	a store C to this petition
0:09:40	um report weeks back and not only the positive
0:09:43	i each but also the negative to every is a zero mean gaussian prior
0:09:48	and for the bit form is also a zero-mean gaussian
0:09:51	but we from partitions
0:09:53	a course we expect to not only make your of what to parts but also to part of the waveform
0:10:00	and the for the baseline for she's is also a zero-mean gaussian at the noise variance prior is you for
0:10:07	yeah
0:10:08	so
0:10:08	not that you have to all these conjugate priors because it can
0:10:13	some five
0:10:14	it it can be the computation
0:10:16	so now we have a lower posterior distribution which is the
0:10:20	the product of all this priors with the likelihood function
0:10:23	but uh it's a topic
0:10:26	distribution which
0:10:27	uh we can not
0:10:28	compute a form estimators
0:10:30	so that why we propose
0:10:32	to use a map model can estimate of a source
0:10:36	to generate samples
0:10:38	weights
0:10:38	a sounded sick leave for this posterior distribution
0:10:41	and to estimate other parameters of the um
0:10:45	the way to model
0:10:47	so this is a proposal a keep were
0:10:50	exactly is uh
0:10:52	there is a simple and S and C and C
0:10:55	simulation is or
0:10:57	uh but that's to this brought constraint
0:11:00	we have a uh as
0:11:02	a slight modification to
0:11:05	class two kids that we're
0:11:07	which that instead of
0:11:08	generate
0:11:09	that but that whole the to cater
0:11:12	there
0:11:13	we can that generate
0:11:15	brought by block
0:11:17	we have only that had a size that it is only
0:11:20	last oh the each block that's one
0:11:23	so
0:11:24	it
0:11:26	you to the company
0:11:27	patient no efficient
0:11:30	so basically is that after generate samples for C where cage or we be see if the is
0:11:37	you way
0:11:38	with is that what that dude
0:11:39	and for for you it is the same
0:11:42	and once we have a
0:11:44	finish
0:11:44	that putting the
0:11:46	indicator are
0:11:48	we was that of the waveform for block
0:11:50	or the process we do and the the baseline baselines and the noise
0:11:54	noise bar
0:11:56	and the the estimates can be a obtained by using
0:11:59	for this
0:12:00	this straight
0:12:01	i meters
0:12:02	can be
0:12:03	all all T and by using a a sample based the map estimator
0:12:07	and for this
0:12:09	the rest of the other having to get a job in by a and M
0:12:12	estimate
0:12:15	so
0:12:15	before drawing use the solution is so mention that the
0:12:19	the pre-processing step yeah our vacation
0:12:22	it's been done by has are region which is very that's
0:12:26	to you your as complex
0:12:28	and the processing window not set in set to ten part eight
0:12:32	this is a force
0:12:34	we cannot have a very large
0:12:37	uh that that thing we know last course you C you has it
0:12:40	so so though
0:12:42	stationary nature
0:12:44	and we can not have a a better little last either course
0:12:48	we have to use several observations
0:12:50	to be able to estimate a we've a problem was that
0:12:55	a if you have a christ if you are interested you can find more
0:12:58	details now journal paper sure here
0:13:02	and now
0:13:03	oh use so use some to we examples
0:13:06	for a that to database
0:13:08	the first example is from
0:13:11	uh to that it's is there a it's a three six
0:13:14	uh this example has been shown to be force it had some
0:13:17	but are we of the reasons and a long beach
0:13:22	a state the of the posterior distribution no that C N you will be cater pages
0:13:27	which means
0:13:28	the probability of having that she or to location
0:13:33	so as we can see
0:13:36	here
0:13:37	to to give sample right handed it is to locate the way so that's the interesting
0:13:42	probably a of a proposed are them
0:13:44	instead that of using a read it and you to determine whether it is a way of one
0:13:49	i one map
0:13:51	that is the map estimator
0:13:53	can tell us
0:13:54	the most probable position of having a Q a here
0:13:58	and C a small to have the peaks that the wave
0:14:02	it just tell us that peaks
0:14:03	and the well there is can be
0:14:06	turn it by
0:14:08	this is estimates of we four
0:14:12	but using different creature
0:14:14	and this is the reconstructed signal
0:14:17	and you read and estimated a baseline of be baseline glass
0:14:21	and all it no signal a don't at because you are very close
0:14:26	and and the
0:14:27	a second rate shows us to detect if it useful
0:14:33	so here is an example with a premature ventricular contraction now
0:14:38	which
0:14:39	we
0:14:40	the she where is the scene
0:14:42	and follows by a giant
0:14:43	if your as
0:14:45	as follows it follows by first
0:14:47	Q way
0:14:49	as we can see
0:14:51	most most our written can handle this situation
0:14:54	the posterior distribution of this
0:14:56	i is very low
0:14:58	so there would be no false alarms
0:15:01	and this in
0:15:03	you work that she way is about detected
0:15:06	with the battery of five was to reduce pollution
0:15:09	and this this you just is a cost that it's signal
0:15:12	and the the detected if it was one
0:15:17	another example of
0:15:18	but if physique she weights
0:15:20	and that we can see
0:15:22	we can have a a very of years
0:15:25	she we've form estimation
0:15:28	and this is reconstructed signal for this dataset
0:15:33	the last example is
0:15:36	a does that we are
0:15:37	there are some as as so you wave fred
0:15:42	so you can see you when there is no to it in the thirty two tomorrow
0:15:46	we can have we would have a very low
0:15:49	posterior distribution
0:15:50	and now or map estimator
0:15:52	we're not need any
0:15:55	for some
0:15:57	so this is a reconstructed see no we is that it that the fiducial points
0:16:02	for this is that does that
0:16:05	and now have that to show is um
0:16:08	constative C of cooperation is with are
0:16:11	we have implemented the filtering in that one of the
0:16:15	filtering techniques and the other is the bit
0:16:18	right
0:16:19	is the basis expansion time
0:16:22	as you get C in terms of of the detection is its ability
0:16:27	our our would some out performs the classical mess source
0:16:31	and also in terms of of D H arrow
0:16:36	our our in is
0:16:37	slightly better is comparable in that's better
0:16:40	and based on the compilation that like to remind you that uh
0:16:45	there are two advantages other advantages of our
0:16:49	method
0:16:50	the first is that uh
0:16:52	we can provide as well as
0:16:54	the we a estimates
0:16:56	estimates
0:16:57	and if probably is
0:16:59	and they are of and he's is that the
0:17:03	this is a ms art
0:17:04	and we can provide
0:17:06	the reliability ability information
0:17:09	such as the inter a a confidence interval
0:17:12	except for
0:17:13	uh oh of the with see which is
0:17:16	very interesting for of the medical yep
0:17:20	so here the convolution
0:17:22	the know what do here is that we have proposed a is model
0:17:26	for the not rest in the state D signals which is based on a blind deconvolution problem
0:17:32	yeah and that we have proposed a block example are
0:17:35	to estimate
0:17:36	the on parameters of space model
0:17:40	and and the as process back
0:17:42	since yeah are more the we have damage use estimation
0:17:45	so we can
0:17:47	i have to
0:17:48	to where at down that don't other nodes
0:17:51	you you some problem that's that really a detection problem
0:17:54	and it's is we can have the with form estimation
0:17:57	we made at at a it's miss detection problem
0:18:00	so all that force back
0:18:03	and since it is not yet
0:18:05	uh online application
0:18:07	uh we are
0:18:09	currently investigating the sequential mass are
0:18:12	uh us to create role what colour methods
0:18:15	to this
0:18:15	basic mode
0:18:17	so
0:18:17	set for the attention if you a is it we have a a little that that them or a personal
0:18:22	web site
0:18:57	i test your back
0:19:11	i
0:19:11	holman breeders to to the you nation problem
0:19:16	uh yes i think we have a
0:19:18	and we have calculated that better at lower our are based on the for example the cs C standard
0:19:24	the and
0:19:26	to
0:19:27	because tolerance
0:19:28	errors
0:19:29	all of the
0:19:31	deviation work
0:19:32	and we can
0:19:34	feed this
0:19:35	stand stand it is that
0:19:36	you're what is not
0:19:37	like a
0:19:38	for you
0:19:39	is is that
0:20:43	yeah i that
0:20:44	if we con
0:20:46	yeah
0:20:47	use that come back to this
0:20:50	yeah
0:20:51	i mean yeah and
0:20:53	there
0:20:54	should be some noise present in signal
0:20:56	and if the noise
0:20:59	is the very likely to that
0:21:01	with forms
0:21:02	it could be
0:21:04	uh i
0:21:06	the um how to say
0:21:08	our data are within
0:21:09	we you we should you pretty per but probably to addition of the way
0:21:14	right
0:21:15	yeah the um
0:21:17	yeah
0:21:18	why we have a sharp here but the L is not about their point five
0:21:23	since we assign the prior that to the but but but it you of having no way is is you
0:21:28	pour to the
0:21:30	a probability of having a a will be zero and with a map estimator just
0:21:34	you would not be a problem
0:21:36	yeah i
0:21:38	is problem
0:21:41	and this also is the don't know which is possible because they here we have a
0:21:47	uh uh to be
0:21:49	the noise
0:21:50	which is
0:21:50	hmmm
0:21:51	very
0:21:52	similar to a way if
0:21:54	so
0:21:54	sometimes i i just think it's a way
0:22:07	yeah yeah i i i have read to this point
0:22:10	and the
0:22:11	to to to them
0:22:12	compute
0:22:13	computational load it's um more
0:22:20	so you

P AND T WAVE DELINEATION AND WAVEFORM ESTIMATION IN ECG SIGNALS USING A BLOCK GIBBS SAMPLER

Biosignal Estimation and Classification

Přednášející: Chao Lin, Autoři: Chao Lin, University of Toulouse, France; Georg Kail, Vienna University of Technology, Austria; Jean-Yves Tourneret, Corinne Mailhes, University of Toulouse, France; Franz Hlawatsch, Vienna University of Technology, Austria