2.11.26 - Hitten Zaveri

Name: 2.11.26 - Hitten Zaveri
Uploaded: 2026-02-12T21:14:49.1Z
Duration: 1 h 4 min 20 s

February 12, 2026

Information

ID: 13833
To Cite: DCA Citation Guide

Download Transcript

00:11Hello?
00:12Hi. Good afternoon, everyone.
00:14My name is Akash. I'm
00:15one of the PGY four
00:17neurology residents.
00:19Today, I have the pleasure
00:20of introducing doctor Zaveri.
00:22He received academic
00:24training in electrical engineering,
00:27computer engineering, and biomedical engineering.
00:30He's the director of computational
00:32neurophysiology
00:33laboratory,
00:34the co director of the
00:35clinical neuroscience group for neuroanalytics
00:38here at Yale.
00:39His research interests
00:41lie at the intersection
00:43of neuroscience,
00:44engineering, and mathematics.
00:46So please help me welcome
00:47doctor Zverey.
00:54Hi, everyone.
00:55Thank you for the introduction,
00:56and thank you for the
00:57invitation to present.
00:59I hope I'm clear. You
01:00can hear me properly.
01:02And,
01:04I'll be talking about the
01:05YNN group.
01:07First, let's see. Make sure
01:09this works.
01:21I I don't see this
01:23working.
01:31Sorry.
01:37Okay.
01:44I'm curious if there's anyone
01:45here who knows,
01:46more about this than I
01:48do.
01:55Oh, it looks like one
01:56minute. Let me see.
01:59Okay.
02:00Let me try now. Got
02:01it. Okay. There was a
02:02menu on there. Okay. Sorry
02:04about that.
02:05There was a display up
02:06on on the screen.
02:08So
02:09as part of my disclosures,
02:10I'm a cofounder, one of
02:12three cofounders
02:13for Alva Health, a Yale
02:14spin out that's working on
02:16the early detection of stroke.
02:17I will not be discussing
02:18this company's work in this
02:20presentation.
02:24The presentation is on the
02:26YNN research group. This is
02:27a group that's co directed
02:29by,
02:30doctor Dennis Spencer, Tor Eid,
02:32and myself.
02:33I worked with,
02:34doctor Spencer for three decades
02:36now. I worked with Tor
02:38for more than two decades.
02:40And, about five years back,
02:42we
02:43realized we were, you know,
02:44at
02:45we we used to attend
02:46each other's meetings. We're collaborating
02:48very strongly. We decided to
02:49come together and form a
02:50research group that brings three
02:52labs together. The three labs
02:54are the computation neurophysiology
02:55lab,
02:56the Biosense lab, which is
02:58a hardware lab, and the
03:00third lab is the AID
03:01lab.
03:02We do three things in
03:04general,
03:05in this research group. One,
03:07you know, we do very
03:08clinical translation work. It's three
03:10our efforts are threefold.
03:12First, we do research on
03:13epilepsy. And here, we study
03:15epileptogenesis
03:16of the process by which
03:17the condition of epilepsy arises.
03:19We study
03:20ichthyogenesis,
03:21the process by which seizures
03:23arise.
03:24We work on localization of
03:25a seizure onset area,
03:27and we work on forecasting
03:28seizures minutes and hours ahead
03:30of time.
03:32Our works on, you know,
03:33patients with medically intractable epilepsy
03:36and in animal models of
03:37epilepsy.
03:38The second general area we
03:39work on is the development
03:41of neurotechnology. And here, we
03:42work on developing sensors, brain
03:44implantable sensors,
03:46electronics, brain monitoring system, and
03:48a brain computer interface device.
03:51And the third
03:52sort of work that, you
03:53know, doctor Spencer Tor and
03:54myself do is we mentor
03:56young investigators.
03:58We mentor a full breadth
03:59of,
04:00young invest investigators from high
04:02school students currently. And this
04:03year, we've taken on four
04:05high school students,
04:06undergraduate students, graduate students, medical
04:08students, postdoctoral fellows,
04:10and,
04:11junior faculty.
04:14I'm going to be discussing
04:15I'm gonna present five distinct
04:17areas that we're working in,
04:19and that's a fair amount
04:20to take in. And it
04:22may feel a bit disparate.
04:24And to help you organize,
04:25you know, to help organize
04:26presentation and the thoughts, I
04:28want you to keep in
04:29mind two sort of organizing
04:31thoughts if you wish.
04:32One is our focus is
04:34quite strongly on the network
04:36theory of epilepsy. We're influenced
04:38by this theory.
04:39And second,
04:41the development of neurotechnology that
04:42we engage in has the
04:44long term goal of developing
04:46a closed loop feedback control
04:47of brain networks. So keep
04:49these two thoughts in mind.
04:51First, on the theories of
04:53epilepsy, there are two general
04:55theories of epilepsy. One is
04:56the focal theory. And focal
04:58theory holds that activity in
05:00a discrete area of the
05:01cortex is aberrant. This is
05:03a seizure focus.
05:05At seizure onset, the focus
05:06recruits adjacent
05:08normal tissue in its aberrance,
05:10and this builds up into
05:11a seizure.
05:12This implicates a balance between
05:14excitation and inhibition, and that
05:16out of control excitation
05:18leads to seizure.
05:20The network theory, on the
05:21other hand, you know, was
05:22proposed by Susan Spencer, my
05:24postdoc mentor,
05:25and this was proposed more
05:26than two decades back.
05:29And it holds that seizures
05:30arise
05:31from large scale networks, aberrant
05:34brain network.
05:35And it goes on the
05:37paper goes on to define
05:38a few networks. But you
05:39can generally assume a network
05:41means you have at least
05:43two regions that are interacting
05:44with each other,
05:46and there may be more.
05:47And these are the nodes.
05:48So there's node a, which
05:50has a direct connection to
05:51node b, and node b,
05:53which has a connection back
05:54to node a. At least
05:55two nodes.
05:56But just keep that in
05:58mind. We are guided by
05:59the network theory of epilepsy.
06:01The second guiding principle is,
06:03you know, we're working towards
06:05a long term goal that's
06:06close to feedback control of
06:07brain networks. Now we're all
06:09familiar with close to feedback
06:10control. The minute you walk
06:11into a room and you
06:12adjust the thermostat because it's
06:14too warm or too cold,
06:16you are part of occlusal
06:17feedback control for controlling or
06:19regulating the temperature of that
06:21room.
06:22The occlusal feedback control system
06:24is has in general three
06:26aspects to it. There are
06:27sensors that are monitoring the
06:28phenomenon that's being controlled.
06:31There is real time analysis
06:33to achieve that feedback control.
06:34And then there's the output
06:36or intervention arm that achieves
06:37the, the the performs the
06:39intervention.
06:40It's not new to us
06:41in epilepsy either. There are
06:43two devices that we are
06:44familiar with. The NeuroPACE device
06:47is a skull implantable device
06:49that continuously meant monitors electrical
06:51activity of the brain
06:53and stimulates. So it closes
06:54the loop by electrically stimulating
06:57to try and control seizures.
06:59The NeuroVista device, which unfortunately
07:01did not cross the early
07:02feasibility stage,
07:05was chest implantable device, which
07:07monitored brain activity
07:09and
07:09gave,
07:10the patient an alert on
07:12a handheld device, sort of
07:14like a blue light, you're
07:14gonna have a good day.
07:15Red light, be careful today.
07:18Unfortunately, that did not make
07:19it past the early feasibility
07:21stage. But you can just
07:22kind of get a sense
07:23of, you know, how the
07:24sensing is done, how the
07:25computation is done, and how
07:27the the loop is closed
07:28by either alerting someone or
07:30electrically stimulating.
07:33The five projects that I'll
07:34talk about and I you
07:35know, this is,
07:39the work in the research
07:40group is quite broad. And,
07:41you know, these are the
07:41five projects.
07:43There's the NeuroProbe project,
07:45the Yale Brain Atlas project,
07:47the Seizure Forecasting project,
07:49a project on the network
07:51brain computer interface, so development
07:53of a network brain computer
07:54interface.
07:56And the fifth project is
07:57network analysis for epilepsy surgery.
08:00My,
08:01transition through these,
08:02projects is not gonna be
08:03very linear. I'll spend more
08:05time on the first couple
08:06projects, and then the remainder
08:08will go very quickly. So
08:09don't don't fret if you
08:10think I'm, you know, very
08:11slow and you're getting bored.
08:13It goes quite fast towards
08:15the end.
08:16The first project, the NeuroProbe
08:18project, this is done in
08:19the Biosense lab. This is
08:21a hardware lab, and this
08:23is currently funded by a
08:24u g three grant to
08:25doctor Spencer and myself.
08:27And,
08:28the genesis of this project
08:29here that it dates back
08:31a number of years ago.
08:32So, you know, more than
08:33four decades back, doctor Spencer
08:35invented
08:36the Spencer depth
08:38sensor probe, which is the
08:39depth electrode used for epilepsy
08:41intracranial EEG monitoring in epilepsy.
08:43And this is marketed worldwide
08:45and used by many centers.
08:46It's marketed and sold by
08:48AdTech.
08:49A second point in the
08:50development of our thoughts towards
08:52what we are doing now
08:53was around two thousand seven
08:55when we had received an
08:56SBIR grant with a small
08:58company in Colorado
09:00to add
09:01modalities to a depth electrode.
09:04And that unfortunately did not
09:05proceed because of the financial
09:07crisis at that time and
09:09the company changed direction
09:10and did not work with
09:11us on a on a
09:12phase two application.
09:14A third point in time
09:15was about ten years back
09:17when Emily Gilmore,
09:18doctor Spencer, Mark Reid from
09:20engineering, and I came together,
09:22and we applied to Connecticut
09:23Innovation.
09:24And we proposed building a
09:26depth electrode,
09:27which would have multiple modalities,
09:29pressure, temperature, oxygen,
09:31intracranial EEG, biosensing capabilities for
09:34glutamate, GABA, lactate.
09:36And we got very sensible
09:37reviews back from the reviewers.
09:40They made a lot of
09:41sense. They said we were
09:42too ambitious, too aggressive,
09:45and they suggested focusing on
09:47a single,
09:49condition and traumatic brain injury
09:51being the one they suggested.
09:53And to pick only modalities
09:54that had predicate devices with
09:56the FDA
09:57to sort
09:58of increase our chance of
09:59passing the FDA.
10:01So since then, we focused
10:02on, you know, TBI for
10:04this, probe. We will come
10:06back to the other modalities
10:08as well. We have interest
10:09in them. But our first,
10:11deliverable
10:12is sort of focused on
10:13TBI. So what's the problem
10:15we're working on? It's, you
10:15know, patients come into the
10:17NICU. They have a primary
10:19injury,
10:20and the that's not the
10:21issue. The issue is the
10:22secondary brain injury that may
10:24manifest.
10:25And, there's a need to
10:26monitor
10:28the brain to pick up
10:29signs of the secondary brain
10:30injury and to manage it.
10:32Otherwise, you know, because if
10:34it's detected early, it's reversible
10:35and managed and can be
10:36matched well.
10:38The current state of the
10:39art is to use multimodal
10:41brain monitoring,
10:42which
10:43is, you know, you place
10:45multiple
10:47sensors
10:48in typically in the frontal
10:49lobe and this illustrates, you
10:51know, the picture on the
10:52right sorry, on the left
10:53illustrates
10:54the placement of probes and
10:55the, you know, measurement of
10:57multiple modalities.
10:59The current challenges with this
11:00are, one, each probe
11:03that's placed requires special sort
11:05of separate wiring,
11:06calibration, maintenance, and then multiple
11:08monitors.
11:09And the data are not
11:11synchronized or not collected in
11:12a manner that, sort of
11:14lend themselves
11:15necessarily to easy analysis on
11:17a single computer, on a
11:18single monitor. The data collected
11:20by disparate systems that have
11:22different a to d resolutions,
11:24different clocks, different different sampling
11:26frequencies.
11:27The equipment also takes up
11:29a fair amount of space
11:30and there is an issue
11:30with cable management, which takes
11:32a fair amount of nursing
11:34staff's time.
11:36The need is for fast
11:37accurate data to guide critical
11:39sort of patient care decisions.
11:41And if this is achieved,
11:43it can help with,
11:44protecting against secondary brain injury.
11:47Our solution for this is
11:49sort of it's not very
11:50innovative. It's a single probe
11:52that integrates
11:53multiple modalities
11:55on a single implantable device.
11:57A single set of elect
11:58so, you know, the the
11:59scheme shown on the left
12:01is the brain implantable probe.
12:03The second component is this
12:05custom set of electronics, which
12:07we call the Neuralink.
12:08And the third component is
12:10a monitor
12:11that displays all the data
12:13that's received from that probe.
12:15The probe schematic is shown
12:17here, you know, measures pressure,
12:18oxygen, temperature, and integrating the
12:20EEG. It has equal or
12:22better sensitivity than predicate devices.
12:24It's a single point sort
12:26of synchronized sensor data stream.
12:29A single output connection comes
12:31out of this, goes to
12:31a single set of electronics.
12:33We don't need a surgeon
12:34or OR for placement.
12:36It can be done in
12:37the field.
12:39It can be done at
12:39the bed.
12:40And the solution can you
12:42know, it's an aspect of
12:43the solution that makes it
12:44portable, which could be of
12:45interest for the military.
12:48In addition to the modalities
12:50and what I've shown you
12:51on the probe, the electronics
12:53also measures scalp EEG,
12:55as a fifth modality. And
12:57in the future, we have
12:58in mind tying this envelope,
13:00cloud based system for advanced
13:02sort of,
13:04analysis that could be performed
13:05and data archival and research
13:07usage.
13:09The advantages of what we're
13:11proposing is, you know, multiple
13:12this table is constructed in
13:13a certain way that the
13:15metrics are listed on the
13:16left. The current practice is
13:18listed in the central column
13:20and the goal of our
13:21solutions listed on the right
13:23in the right column.
13:25And typically, in the current
13:26practice, you'll see multiple
13:28probes, multiple cranial access sites,
13:31multiple electronic devices, multiple monitors,
13:33multiple data streams,
13:35multiple sort of the location
13:36of sensors. There are multiple
13:37sensors. They're not exactly collocated.
13:40While if you look on
13:42the neuroprobe column, you know,
13:44this is all single. There's
13:45a single probe, a single
13:47access site, a single device,
13:49a single data stream, a
13:50single monitor, etcetera. So you
13:52see the advantages
13:53essentially
13:55reducing the complexity
13:56of the current practice to
13:58a single solution.
14:00Three rows from the end,
14:02you know, there is the
14:03cost per bed, equipment cost.
14:04So the cost for instrumenting
14:06a current system is roughly
14:07about hundred and seventy five
14:08k. We think we can
14:10bring this below fifty k.
14:13And, you know, as mentioned
14:14earlier, we have a portable
14:16version of our solution.
14:18In progress towards this, we've
14:20been, you know, we've the
14:21NIQG three grant has been
14:22active for three years now.
14:24We have developed pressure, temperature,
14:27oxygen, integrated EEG sensors. We've
14:29developed the three components, the
14:30probe, the electronics, the monitor.
14:33We match the performance of
14:34predicate devices. So, you know,
14:36in that sense, we can
14:37we we know we could
14:38our design beats predicate devices.
14:41We've addressed sterilization,
14:42packaging, shelf life issues.
14:45We've initiated sort of design
14:46for manufacturing, design control, QMS.
14:49These are things that are
14:51needed for medical device design
14:53and development.
14:54We've developed a patent portfolio
14:56strategy. We've completed two presubmissions
14:58with the FDA, both of
14:59which have gone quite well.
15:01We're running out of funds
15:02right now, but we still
15:03have to do work towards
15:04regulatory evaluation approval.
15:07And then at some point,
15:08if we have the funds,
15:09we will transition into an
15:11early feasibility study.
15:13A model of the probe
15:14is shown at the bottom
15:14and a model of the
15:15monitor shown at the bottom
15:16showing all the modalities and,
15:18you know, on a time
15:19lock screen.
15:21As we've done this in
15:22the lab, we've developed other
15:23sensors as well. You know,
15:25keep in mind our earlier
15:26interest in measuring
15:28glutamate, GABA, lactate on the
15:30probe as well.
15:31So we have a hardware
15:32lab. We've got,
15:34Jesus, who's here in in
15:35UA, who unfortunately we lost
15:37in fall to Silicon Valley,
15:40have have done excellent work
15:41in the lab.
15:42On the left, you see
15:44a wafer,
15:46you know, that we develop
15:47these wafers in Yale clean
15:48room, in the Yale clean
15:50rooms, and we the this
15:51is the design for the
15:52oxygen sensor.
15:54And, they come out in
15:55a wafer of this nature.
15:57On the right, you see
15:58a a chip like structure,
16:01which has got microfluidic channels
16:03in which we use to
16:04test a second device that
16:06we make in the lab,
16:07which is called the silicon
16:08it's called silicon nanowires.
16:10Silicon nanowires were brought over
16:12from engineering because one of
16:13our collaborators there, Mark Reid,
16:15unfortunately passed away when we
16:17started this project. And his
16:19lab had developed this, and
16:20there was a chance that
16:21this was not gonna be
16:22carried on at Yale. So
16:23we brought that over from
16:25engineering into,
16:26our lab. And we've been
16:28successful at being able to
16:29recreate,
16:31these nanowire sensors.
16:33So the the, you know,
16:34the the two sort of
16:35sets of images here, the
16:37top one shows multiple silicon
16:38nanowires in a sensing area.
16:40The bottom one shows a
16:41single silicon nanowire.
16:43These are very tiny
16:44sensing areas.
16:46The key aspect about them
16:47is they're built on in
16:49the back end of of
16:50the sensor is the transistor.
16:52So as soon as the
16:53sensing performed on the sensing
16:55surface, the signal essentially is
16:56getting amplified.
16:58And so we can pick
16:59up very tiny signals with
17:00this.
17:01Some, results. So this was
17:03published last year,
17:05work done by Jesus, UA,
17:07and Wahoo,
17:08showing that we can measure
17:10GABA and PBS.
17:12And, we've,
17:13you know, functionalized these sensors
17:15for
17:16measuring GABA, glutamate, and lactate.
17:18And the limits of detection
17:20are in the femtomolar.
17:22That's ten to the minus
17:23fifteen.
17:24Well,
17:25you know, well beyond what
17:27our needs are for,
17:29measurements in the brain. So
17:30this is quite good. It's
17:32a very excellent progress. More
17:34recently,
17:35Jesus presented this work
17:37with a photonic sensor that's
17:39being developed in the lab
17:40as well. So the photonic
17:41sensor is shown a schematic
17:44shown in the top left.
17:45You have a signal part.
17:47Normally in these sensors, you'd
17:48have input light going in
17:50from one edge and coming
17:51out from the other edge.
17:53We had in mind integrating
17:55this into our depth select
17:56into our neuroprobe device. So
17:58we need to bend the
18:00part to bring the light
18:01back in the same direction
18:03that you know, same point
18:04where it
18:05entered. And, hence, you know,
18:06that's the structure for
18:08it. And
18:10it looks like these results
18:12are not loading.
18:13Okay. I I'm not sure
18:15why this is not loading.
18:19Okay. For some reason, these
18:21dots are not loading. You'd
18:22have to just take my
18:23word, that, you know, we've
18:25we've achieved some level of
18:26success with detecting GABA and
18:29glutamate. The limits of detection
18:30are listed about
18:32down to fifteen,
18:33nanomolar for GABA
18:35and twenty five micro picomolar
18:37for glutamate.
18:38These are early results that
18:40were just presented last month
18:42at,
18:42Photonics
18:43West,
18:44and, this will be sort
18:46of be worked up more
18:47into that.
18:48So in this area of
18:50work,
18:51the deliver deliverables we're working
18:53towards are, one, we think
18:55embedded within our solution is
18:56a low cost scalp EEG
18:58system, which will should be
18:59ready by q three of
19:01this year as a working
19:03prototype,
19:04fully functional scalp EEG system.
19:07We would also have a
19:08a working prototype for the
19:10pressure, g, temperature, and scalp
19:12EEG version of the NeuroProbe
19:14system. That's where they'll, implantable
19:16probe, the electronics, and the
19:18monitor.
19:19And then, we're waiting for
19:21more funding to come in
19:22to add in the oxygen
19:24sensor,
19:24which we will call sort
19:26of version two of the
19:27solution. And in the future,
19:28at some point, depending on
19:30funding that
19:31we receive and the progress
19:33we make,
19:34we would include other sensor
19:35modalities.
19:37As we've developed this, you
19:38know, as I mentioned, we've
19:39developed the silicon micro nanowire
19:42sensors,
19:43and this is a solution
19:44that's looking for a problem.
19:46So if you have
19:48a need,
19:49to measure analyte concentrations
19:51in any biofluid,
19:53you know, we have a
19:54technology that seems to deliver,
19:56you know, limits of detection
19:57down into the femtomolar
19:59range.
20:00So this may be worth
20:01working out for other application
20:03areas, and we would be
20:04interested in talking to you
20:05if you've got,
20:06any any challenges in sort
20:09of doing let's say making
20:10a point of care device
20:11with those sort of capabilities.
20:13The team that's working on
20:14this is shown here.
20:16The
20:17students that have worked with
20:18us, Jennifer, Gloria, Tia,
20:22and Yue, Jesus, Wahoo, and
20:24Ronnie
20:25do a fair amount of,
20:26the
20:27the development, and Ronnie helps
20:29us with, the animal studies.
20:31The team shown in the
20:32upper right with Emily Gilmore,
20:34so Belinda May, Jennifer Kim,
20:36and Brad Duckrow are the
20:38that's the clinical team that's
20:39waiting for us to complete
20:40our tests so that they
20:42will, sort of work on
20:43this with for the early
20:45feasibility study at Yale. And
20:46the team shown at the
20:47bottom, which is,
20:49professor Belinda Srey and Kuyfmann
20:52are our data science experts.
20:54Because one of the other
20:55challenges that's going to emerge
20:56is we've got multiple modalities.
20:58We need to simplify
20:59the display
21:01and provide very clear cut,
21:05suggestions on what treatment should
21:06be followed for what particular
21:08situation.
21:09So that brings to conclusion
21:10the first project I'm presenting.
21:12I'm gonna go on that's
21:14the work of the Biosense
21:15lab. It's a hardware lab.
21:16I'm gonna go on. I'm
21:17gonna march through the other
21:18project, and then there'll be
21:20time at the end for
21:20questions and, you know, if
21:22you've got any. The second
21:23project is a multimodal brain
21:25atlas, which we're preparing for
21:27surgery
21:28planning. It's doctor Spencer, doctor
21:30Sivraj, you and myself.
21:32This is funded internally through
21:33the Swivelius Trust.
21:35We finally got our act
21:36together and submitted a grant
21:38to NIH last year,
21:39after several years of work
21:41on this project awaiting to
21:42see what the reviews are.
21:46And a few years back,
21:48we had a very good
21:49medical student from the UK,
21:51from King's College London, Harry
21:53McGrath, who spent time with
21:54us.
21:55And at that time, we
21:56challenged him, and he worked
21:57with doctor Spencer to create
21:59the Yale Brain Atlas. The
22:00Yale Brain Atlas has six
22:02hundred and ninety or six
22:02hundred and ninety six parsons
22:04depending on which version of
22:05it you use.
22:06It's
22:07to our attention no. Not
22:09that it's the highest resolution
22:11brain atlas that's built on
22:12anatomic landmarks.
22:14It's postulated. It's the m
22:16and I one fifty two
22:17brain postulated to the nearest,
22:20centimeter.
22:21We use essentially, we wanted
22:22to come up with
22:24a resolution a cortical resolution
22:26of one square centimeter.
22:28We has an intuitive nomenclature
22:30and coding structure, and colors
22:31are used to highlight sort
22:32of the gyri within specific
22:34regions.
22:37It's an atlas here. Why
22:38did we create it? We
22:39did that because we wanted
22:40an atlas based on anatomical
22:43landmarks
22:44rather than one that was
22:46generated through computation
22:47because we are interested in
22:49the network theory of epilepsy.
22:50We did not want an
22:51atlas that was created based
22:53on other network measures.
22:55We wanted one based on
22:57anatomy.
22:59And we were motivated to
23:00create an atlas at this
23:02resolution because
23:04when we place electrodes, intracranial
23:06electrodes for epilepsy surgery,
23:08we use either depth EEG,
23:10which are s e g
23:11electrodes. This not gonna be
23:13placed closer than ten millimeters
23:15apart. We're going to respect
23:16that,
23:18that resolution.
23:19And if you use strip
23:20and grid electrodes, which were
23:22popular earlier, the center center
23:24spacing for strip and grid
23:25electrodes is also one centimeter.
23:27So we wanted to respect
23:29information that was being created
23:30at that resolution,
23:31and that's the resolution we
23:33created Satis on. We believe
23:35the simplified nomenclature
23:37helps with communication of neuroanatomy,
23:39and it facilitates sharing of
23:41multimodal research findings.
23:43Once we built the Atlas,
23:45we've also built other resources
23:46around it. So we,
23:48you know, define structural data,
23:50white matter connector, and cortical
23:51thickness. We've defined functional data,
23:54and I'll go through this.
23:55And all of this is
23:56on our website, and you
23:58can interact with it. And
23:59it's it's it's a bit
24:00slow, but it works quite
24:02reasonably well. And it's worth
24:04sort of interacting and and,
24:05you know, trying it out.
24:07And all the information I'm
24:09presenting, by and large, most
24:10of it is in our,
24:12GitHub repository that's also shared
24:15and on,
24:16another open
24:18repository, which is
24:20sort of listed below.
24:21So we worked up white
24:23matter connectivity using,
24:25human connect and project
24:27thousand sixty five subject template,
24:29and we worked up parcel
24:32to parcel connectivity. So this
24:34is six ninety parcels to
24:35six ninety parcels.
24:36So imagine a matrix of
24:38that size, six ninety six
24:39ninety space. We've got white
24:41matter streamlined connection strength, got
24:43white matter distance, Euclidean distance
24:46between all of these points.
24:47We've also worked up all
24:48the major white matter tracks
24:49and their mapping onto the
24:51parcels.
24:52This was done by Omar
24:54Chishti when he was an
24:54undergrad student here at Yale,
24:57in BME,
24:58and then he stayed on
24:59in the lab and worked
25:00with us. Alex King, who
25:01came to us from UC
25:02Berkeley,
25:03worked up a pipeline to
25:04map cortical thickness on the
25:06brain atlas. And this is
25:08again, from two hundred human
25:09connective subjects
25:11and young adults.
25:12And you can see the
25:13thickness,
25:14and this color bar is
25:16shown on the right. It's
25:17in millimeters,
25:18and you can see that
25:19this thin cortex is primary
25:21motor and sensory areas and
25:23thicker court cortex in other
25:25parts of the brain.
25:27Evan Collins, who started with
25:29us as an undergrad,
25:32and is currently finishing his
25:33PhD at MIT,
25:36then took a large open
25:37source repository known as NeuroSend,
25:40which has got information on
25:42from, you know,
25:44from fourteen thousand three hundred
25:46thirty one fMRI papers
25:48and with one thousand three
25:50hundred thirty four keywords.
25:52And he mapped that. That
25:54information exists on the m
25:55and I one fifty two
25:56brain. So he mapped that
25:57over into the brain atlas.
26:00So if you wish, you
26:01know, that's coming,
26:03so we
26:05we aggregated the information in
26:06the neurocent database
26:08to our dataset at that
26:10passive resolution. And, again, you
26:11can see you go up
26:12onto our website, you you
26:14click on a parcel,
26:15you can see all the
26:16keywords that come up tied
26:18to that parcel. In this
26:19case, you know, it's tactile,
26:21touch, pain, etcetera,
26:23are coming up. There's a
26:24second large repository,
26:27you know, that was built
26:28on the Neurosyn repository. It's
26:30called parcel query, which does
26:32more
26:32word process you know, sort
26:34of semantic smoothing and word
26:36embedding
26:37on the Neurosynth
26:38data. And we map that
26:40overall. So we call that
26:42parcel query. So, you know,
26:43neurosynt got mapped into parcel
26:45synth.
26:46Neuroquery got mapped into parcel
26:48query.
26:48And then Evan asked a
26:50couple of questions on, you
26:51know, he he was he
26:52was motivated by trying to
26:54understand the relationship between structure
26:56and function.
26:57I'm I'm showing you this,
26:59you know, just skim through
27:00his findings,
27:01just to give you a
27:02sense of what can be
27:03done with this repository that
27:05we've created.
27:06So we know,
27:08that the relationship, you know,
27:09we're interested in network measures.
27:11We're interested in studying networks.
27:13We know the connectivity
27:14is, you know, measured through
27:16white matter or structural connectivity
27:18or it could be measured
27:19through functional connectivity. So you
27:21could measure connectivity from fMRI
27:23time series or EEG time
27:25series or MEG time series.
27:27But the global correspondence between
27:30structure and functional connectivity is
27:32very poor.
27:33And Evan wanted to see
27:35if using these large repositories
27:38improved our understanding
27:39of the relationship between structure
27:41and function.
27:42So this, complicated,
27:45table kind of lays out
27:47the poor
27:48relationship we see. The rows
27:50are thirteen different measures of
27:52structural connectivity
27:53pulled from the white matter
27:55connecting that we've created.
27:56So, you know, for example,
27:58SC count is, you know,
27:59structured connectivity based on a
28:01count of streamlines between two
28:03parsons.
28:04The columns, the three columns
28:06are three different evaluations
28:08of functional connectivity.
28:10The neurosynth
28:11version
28:12of of the database, the
28:14neuro query, and from resting
28:15state fMRI.
28:16And you can see in
28:17that, you you know, thirteen
28:18by three matrix there or
28:21that array that the values,
28:23the relationship between structure and
28:24function is very poor.
28:26This is again recreating what's
28:28known in the field. The
28:29global structure and function
28:31correspondence is very poor in
28:33terms of connectivity.
28:36I won't go through all
28:37of Evan's work. It's it's
28:39really it makes a very
28:40good reading. It was published
28:42in twenty twenty four.
28:43But the key takeaways are,
28:45you know, we use large
28:46scale data repositories to compute
28:48structure function correspondence
28:50and across functions, hundreds of
28:53functions.
28:54So and then we showed
28:56that global structure function have
28:59imperfect correspondence.
29:01But then we find that
29:02cortical thickness
29:03does
29:04have,
29:06an impact here.
29:07And the plot shown in
29:08the bottom left is just
29:10for the parietal lobe. On
29:11the x axis, you see
29:13the structure function correspondence, if
29:15you wish. S f r
29:17squared just as a measure
29:18of how closely structure and
29:20function are correlated.
29:22Low values means it's very
29:24poor correspondence and higher values
29:26means it's better tied together.
29:27On the y axis is
29:29cortical thickness in the parietal
29:30in different parcels in the
29:31parietal lobe. And you can
29:33see that structure function correspondence
29:36increases
29:37as the parcels
29:38are for thinner parcels.
29:40The thinnest parcels
29:42have the best structure function
29:43correspondence.
29:44The thickest parcels have the
29:46poorest structure function correspondence.
29:48Just keep that in mind.
29:50The other things and he
29:51he did a lot of
29:52exploring of, the information that
29:54he had created.
29:55And, another plot that I'm
29:57showing here in the bottom
29:58right is,
29:59structure function. Two box plots
30:02are shown. One for high
30:03structure function parcels and the
30:05other for low structure function
30:07parcels
30:08against a measure of concreteness
30:10of the word the keywords.
30:12And concreteness was scores that
30:14we got from a paper,
30:17which had, you know, asked,
30:19users or subjects to provide
30:22a sort of rate words
30:23based on the involvement of
30:25sensors
30:25and motor responses.
30:27For example, touch would be
30:29a very concrete word, while
30:31a word like justice would
30:32be very abstract
30:33or would be very low
30:34on the concrete scale. So
30:36what this is showing is
30:37that
30:38concrete words
30:39had
30:40tighter or higher structure function
30:42correspondence
30:43while more abstract words had
30:46poor structure function correspondence.
30:48The main takeaways
30:49from the paper, again, worth
30:51reading. You know, the question
30:52was, is the structure function
30:53relationship the same throughout the
30:55brain, the same across different
30:56functions?
30:57And, Evan's take home messages
31:00were it exists on a
31:02gradient.
31:03It's strongest in primary sensory
31:05motor cortical areas for perceptual
31:07and motor functions.
31:08It's weakest in association
31:10cortex for complex cognitive function.
31:12And then we speculated a
31:14couple of things.
31:15The speculation is that the
31:16evolution of the human brain
31:18might help explain,
31:21this gradient.
31:22So essentially, you know, we're
31:24seeing a gradient between unimodal
31:26cortex
31:27to cortex that supports multiple
31:30functions.
31:30And one possible reason is
31:32that while direct connection between
31:33brain regions was sufficient
31:35for faculties such as vision
31:36and movement, as the brain
31:38developed more advanced capabilities
31:40like complex cognition,
31:43these direct connections had maxed
31:45out the usefulness.
31:47And it's possible that the
31:48brain developed more indirect connections
31:50between regions to establish more
31:52advanced
31:53abilities. And we believe that
31:55these areas where this happened
31:58have poor structure function relationship.
32:02And it's a it's it's
32:03a good use of the
32:05brain atlas that we've created.
32:07We didn't create it for
32:08this purpose. We created it
32:09for epilepsy surgery, but we've
32:11mapped a lot of information
32:12onto it that can be
32:13used for a number of
32:14different studies. I'll show you,
32:16a couple other studies. The
32:18second,
32:19you know, this paper by
32:21sort of we've collaborated very
32:22strongly with, doctor Sivraju,
32:24and this is data from
32:25doctor Sivraju,
32:27his electrical stimulation mapping of
32:29language.
32:30And,
32:31Omar Chishti helped work up
32:32this analysis.
32:34And we were interested again.
32:36The BRAIN ATLAS allows us
32:37to combine data from multiple
32:39patients and do comparisons. It
32:41it it simplifies comparisons.
32:44So doctor Subraraju looked at
32:45six different tasks, auditory naming,
32:47visual naming,
32:49reading,
32:49auditory compre
32:51comprehension,
32:52written comprehension, and repetition.
32:54And this is the mapping,
32:56if you wish, across the
32:56fifteen subjects we looked at
32:58onto the brain atlas for
33:00the six tasks.
33:01And then when we combine
33:02the tasks,
33:04if you look in the
33:05center, the b,
33:07subfigure,
33:08the red parcels are the
33:10ones where all six tasks
33:13lined up or, you know,
33:14were active in all patients.
33:16So we, you know, we
33:17identified this as a language
33:19core.
33:20And then the c sub
33:22figure shows
33:23task you know, the color
33:24code is going from zero
33:25to six depending on how
33:26many tasks
33:28activated that parcel. So this
33:29allows us a very easy
33:31way to understand the language
33:33core. And as we step
33:34away from the language core,
33:36how many you know, how
33:37much variability is there in
33:39language representation across the subjects
33:41we're studying.
33:42So this was another
33:43useful example for us to
33:45sort of demonstrate the prowess
33:47of the brain atlas.
33:48A third study that I'll
33:49show you, this hasn't been
33:50published. It's it's under review.
33:52Elizabeth Watson has, been in
33:54the lab since the first
33:55year.
33:57She just,
33:58is a just joined Yale
34:00School of Medicine. She's a
34:01first year in med school
34:02now.
34:03And, she took
34:05the partial synth information that
34:07came from Evan's study
34:09from this fMRI
34:10web scrape data,
34:12and the electrical stimulation data
34:14from Aditya
34:15is a dataset,
34:17on that was shown in
34:18the last column. And then
34:20also did a literature review
34:22of expert annotation and compared
34:24the two. And you can
34:24see that going from the
34:26parcel synth
34:27dataset, which is quite broad.
34:29That's the fMRI activation data.
34:31And on the other extreme,
34:33you have the electrical stimulation
34:34data from Yale from just
34:35fifteen subjects. You can see
34:37the literature meta analysis. So
34:38we're also working up a
34:40more detailed sort of literature
34:41analysis
34:42for other lobes and other
34:44functions. But this is to
34:45demonstrate that this can be
34:46a useful,
34:47framework on which to add
34:49information from different sites, different
34:52sources, and do a comparison.
34:56I I, you know, titled
34:57this as, you know,
34:59Gale Brain Atlas for Surgical
35:00Decision Making for Hepatitis Surgery.
35:02Natalie Ackerman comes to us
35:04from Spain.
35:05She is an undergrad student
35:07who's working who's both a
35:09medical student and an engineering
35:11student at the same time.
35:12She's simultaneously
35:13doing two programs. And,
35:16as an engineer, I really
35:17appreciate that background. So very
35:18happy to have her in
35:19the lab. What she's working
35:21with us on is taking
35:22all the multimodal data that
35:23we've created
35:24and fusing it on the
35:26same on a single brain
35:27to help us with decision
35:28making. So these are two
35:30separate patients
35:31at two different points.
35:33Each one, you know, patient
35:34one is at one point
35:35and patient two is the
35:36second point during surgical decision
35:38making for epilepsy surgery.
35:40We typically have two points.
35:41So the first is we've
35:42got noninvasive
35:43measures
35:44like scalp EEG,
35:47fMRI,
35:48PET,
35:50might have
35:51MEG.
35:53And we have to make
35:54a decision on in terms
35:55of lateralization,
35:56regionalization,
35:57and understanding, you know, making
35:59decisions on where to put
36:00intracranial electrodes if the patient's
36:02gonna go ahead with an
36:04intracranial surgery,
36:06or intracranial monitoring.
36:08And that image shows,
36:10scalp EEG,
36:11cortical thickness evaluation,
36:13PET,
36:13evaluations,
36:14etcetera, collocated in the same
36:16brain. So you can kind
36:17of understand that this patient
36:19is showing primarily left temporal
36:21lobe,
36:22findings from all the noninvasive
36:24modalities.
36:25And you can see as
36:26the modalities, you know, more
36:27and more modalities overlap, we're
36:29getting closer to what was
36:30finally identified at the seizure
36:31onset area, which is in
36:33the left temporal lobe. On
36:34the right, we have a
36:35second patient where we brought
36:37in
36:38not just the noninvasive
36:39modalities, but also the intracranial
36:41measures that are collected.
36:43And, it also shows where
36:45the language mapping was before,
36:47where language,
36:48was established for this patient.
36:50It shows a prior resection
36:52area. It shows where a
36:54neuropace device
36:56electrode was placed and where
36:57stimulation is being performed.
36:59We're working up several examples
37:00along these lines to understand
37:02if this is useful
37:04for decision making and and
37:05discussion
37:06in, the epilepsy surgery conference.
37:09We are sort of at
37:11the start of this exercise.
37:12We're also doing this in
37:14a separate manner computationally, but
37:15this this is more illustrative.
37:19In other work that's coming
37:21out of the Brain Atlas
37:22project, you know, we've got
37:23several papers. You can find
37:24them, that have been published.
37:26You can find them on
37:26our website.
37:28And then, the few others
37:29that are in the pipeline.
37:30I mentioned Elizabeth Watson's
37:33literature review of language literature
37:35review.
37:36Yi Xiaow has submitted paper
37:37on quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote, quote,
37:39quote, quote, quote, quote
37:46she's submitted a small, short
37:49communication on that. Brian Bozroy
37:52is working up on a
37:52transformer based framework
37:54for fMRI data on the
37:56brain atlas, which is revealing
37:58some pretty nice structure to
38:00us and information that, we
38:02had not seen before from
38:04fMRI.
38:05This team is is quite
38:06large. What's unique about this
38:08project is that we just
38:09have undergraduate students working on
38:11it. The top two rows
38:12are under we've had one
38:14medical student. I I guess
38:15now Elizabeth's joined Yale School
38:17of Medicine. So we have
38:18two medical students
38:19and one graduate student, Jishao.
38:22The rest are undergraduate students,
38:24and they pretty much drive
38:25the project. It's a lot
38:26of fun.
38:27We meet three times a
38:28week,
38:30and they all take on
38:32different
38:33subprojects and work on them.
38:36That's the second project,
38:39that I wanted to present.
38:41The third project is, what
38:43we call our saliva project,
38:45or seizure forecasting project.
38:48And that's,
38:49you know, funded by an
38:50r o one, which we
38:51received a little over a
38:52year back,
38:54to the three of us,
38:56Torai, doctor Spencer, and myself.
38:58And our collaboration,
39:00is with Vikram Rao at
39:01UCSF and Maxine Borde at
39:03UGC Bern in Switzerland.
39:05This project
39:06is strongly influenced by this
39:08result, which was published in
39:09two thousand eighteen
39:11by our colleagues, Maxine Board
39:13and Vikram Rao.
39:14And what this shows is
39:17an you know, a patient
39:18with a neuroPACE device
39:21being monitored, and this is
39:22a twenty second epoch, and
39:23there is this one epileptiform
39:24discharge. So you have a
39:25count of one, and you
39:27have multiple discharges in this
39:28twenty second epoch. You have
39:29a count of eighteen.
39:31And what they did is
39:32they took the detector counts,
39:34hourly detector counts, and plotted
39:35them. This is over two
39:36months, February and March, for
39:38a given patient. And you
39:39can see a very strong
39:41circadian or or twenty four
39:42hour cycle in there that's
39:43modulating the data.
39:45But also modulating it at
39:47another time scale is this
39:49other cycle that happens at
39:50certain
39:51times. And then when you
39:52average the data across the
39:54whole day and you look
39:55at it, this is for
39:56twelve months or one year,
39:57and this two months, embedded
39:59within this one year, you
40:01see the emergence of these
40:02cycles. Now these cycles are
40:04not known, not understood.
40:06These are multi day cycles
40:08that are seen in men
40:09and women, and they seem
40:10to govern the occurrence of
40:12seizures.
40:13In this particular patient, the
40:15seizures are marked as red
40:16dots and you can see
40:17the seizures occur on the
40:18upslope
40:19of this cycle
40:20at relative maxima or relative
40:22minima. And you can see
40:24towards here, you know, a
40:25lot of seizures clustering and
40:26happening on the upslope
40:27of this
40:28cycle.
40:31The way we approach this,
40:33in collaboration with Maxim and
40:35and Vikram
40:36is to propose a project
40:38in which we would take
40:39daily or multiple saliva samples.
40:43Working with Tor, we put
40:45these saliva samples through a
40:46mass spec
40:47analysis and create
40:49a rich report sort of
40:50rich assay, if you wish,
40:52of organic acids, amino acids,
40:54steroids, hormones, etcetera.
40:56And then we put this
40:58through machine learning algorithms
40:59to try and understand if
41:01there are any changes in
41:02the saliva chemistry that is
41:04predictive
41:05of an episode.
41:07The two aims that we
41:09are pursuing in this application
41:12are one aim one is
41:13an at home study. It's
41:14conducted over four hundred and
41:16twenty
41:17days. And we this is
41:19the data we collect. So
41:20there are the piece of
41:21patients with RNS devices who
41:22collect the r the RNS
41:24information.
41:25They have a wearable device
41:26that's, you know, collecting acceleration
41:28data. They maintain a diary
41:30of the seizures and the
41:31mood.
41:32They maintain,
41:33information on the meals they've
41:35had,
41:35the meds they're taking, and
41:37they take saliva samples three
41:38times a day before breakfast,
41:40before dinner, and before bed.
41:42We also in aim two,
41:44we are studying if you
41:45wish more the circadian structure
41:47of these variations.
41:49And this is done in
41:50the epilepsy monitoring unit in
41:52house, and this can be
41:53up to six days or
41:54so. And here we measure
41:55EEG acceleration.
41:57We keep note of the
41:59meal times, the medications they're
42:00on, and we take saliva
42:02several times a day before
42:04meals and including a three
42:06AM sample out of sleep
42:07and post seizures.
42:10I'll give you a sense.
42:11This is still very early
42:12in terms of where we've
42:13reached in terms of analysis
42:15and working with the data,
42:16but this will just give
42:18you a sense of what
42:19may come out of this
42:20project.
42:21The
42:22figure,
42:23the the sub figures on
42:24the left show
42:25three of our subjects that
42:27have neuro paced devices
42:29and the counts
42:30the detection counts of the
42:31detector.
42:32And you can see the
42:33top two subjects are in
42:34a subject three and six
42:35have strong multi day cycles.
42:38And the top part of
42:39the cycle is colored red.
42:40The bottom part of the
42:41cycle is colored blue.
42:43And the third subject, which
42:44is subject two, does not
42:45have these cycles.
42:48So we took saliva samples
42:50from the blue part of
42:51cycle and the red part
42:52of cycle and then put
42:53them in a binary classifier
42:55to say, can you distinguish
42:57based on analyte concentration
43:00the high parts of the
43:01cycle from the low parts
43:02of the cycle? Because we
43:03know seizures come from that
43:05red part of the cycle,
43:07which we call a proictal
43:08state
43:09as opposed to the blue
43:10part, which is an antique
43:12state. Because if you could
43:13do this, then we could
43:16conceivably propose a test where
43:17you spit in a tube
43:19and you get a reading
43:20that you're gonna have a
43:21good day or a bad
43:22day.
43:23So this is very early
43:25analysis.
43:26We built a classifier. That's
43:27the ROC curve, but the
43:29numbers are here. We could
43:30correctly identify the two positive
43:32rates for the antiictal
43:33cycle is the antiictal time
43:35point is seventy eight point
43:36six percent and proactyl time
43:38point is sixty three percent.
43:39Now this is with a
43:40small subset of the analytes
43:42concentration that we expect to
43:43have. This was done very
43:45early on. We put this,
43:47we we ran the data
43:48from four control subjects and
43:50seven epilepsy subjects
43:51through the classifier,
43:53and we found that twenty
43:55nine percent of the samples
43:56from the epilepsy patients were
43:57in the procedure state, while
43:59only two percent of the
44:01samples from the control subject
44:02were in the procedure state.
44:04So I know it's a
44:04bit of stretch, but essentially,
44:06you know, the the group
44:07has a lot
44:08of interest and expertise in
44:10studying neurochemistry.
44:12We've done that for many
44:13years. Doctor Spencer, Toride.
44:15We used to do brain
44:16microdialysis
44:17in our patients at Yale,
44:19going back several decades.
44:21We've done these studies in
44:22animal models of epilepsy.
44:24And we are now taking
44:25that, you know, a couple
44:26steps. So we are arguing
44:28inherent to this is that
44:29argument that saliva, which is
44:31a surrogate for blood,
44:33contains some of the information
44:35that we'd be picking up
44:36from neurochemistry
44:37in the past.
44:38And this is, you know,
44:39what we are working on.
44:41The team we've created is
44:42quite large in order to
44:43be able to collect these
44:44saliva samples at home or
44:46in the epilepsy monitoring unit.
44:48It's again, you know, led
44:49by some, there's an undergrad
44:51medical students.
44:53Different people have worked on
44:54the project at different time.
44:56There are four or five
44:57medical students here. Ami
45:01sort of has taken the
45:02lead in creating this team
45:04that goes in and takes
45:05saliva samples for us in
45:06the EMU.
45:07And our colleagues, Maxine Board,
45:09Vikram Rao, are shown here
45:10as well.
45:12And we want to thank
45:13all the EEMU staff. That's
45:15very helpful.
45:16Amadeo,
45:18Rebecca, sort of Brandy,
45:20and, Rebecca as a leadership
45:21in this and the nursing
45:22staff that's helped us tremendously,
45:25in collecting this data.
45:28So that's, you know, the
45:29the the fourth project is
45:31a brain computer interface project.
45:34This is funded I forgot
45:35to list the second grant.
45:36We put NIH, r o
45:37one, and we've got an
45:38NSF award on this. And
45:40here we're trying to build
45:42capability
45:43for monitoring and modulating brain
45:45networks. So essentially, this translates
45:47into a question. Can you
45:48build a BCI
45:49that can monitor multiple points
45:51in the brain
45:52and
45:53perform network analysis?
45:56So, you know, can we
45:58support large channel counts? By
46:00large, we mean it could
46:01be fifty, could be a
46:02hundred.
46:03And high data rates, can
46:05you support
46:06machine learning, deep learning, connectivity
46:08analysis on a brain computer
46:10interface device? Could it be
46:11software programmable so that the
46:12same hardware can be used
46:14for multiple indications
46:15and not just be finely
46:17developed for one indication? And,
46:19you know, at the end,
46:20my colleagues threw in this
46:21one
46:22requirement. Could it last for
46:23years on a single battery?
46:27A first gen this is
46:28not vaporware. The first generation
46:30chip has been built at
46:31Yale
46:32by,
46:33Rajeet Manur and Abhishek Bhattacharjee.
46:35They call it the halo,
46:37you know, hardware architecture for
46:39low power BCIs.
46:41And this chip has been
46:42built the way our chips
46:44in our cell phones work.
46:45It's got processing elements,
46:47system of compression,
46:50discrete wavelet transforms, nonlinear
46:52energy operator, fast Fourier transforms,
46:54cross correlation, band pass filters,
46:56support vector machines. So many,
46:58computational kernels that we perform
47:01on integrating EEG
47:02have been resolved into hardware
47:04to make them more efficient,
47:05and they work on different
47:06clock domains. We could get,
47:08you know, the the, power
47:10on this is on the
47:11order of fifteen milliwatts,
47:13which is still an a
47:14few orders of magnitude much
47:16higher than what we want
47:17it to be. So we
47:18are working now on the
47:19next generation.
47:20We've got funding for two
47:21more generations of chips. The
47:23next generation of chip is
47:24going to, propose is an
47:25architecture that's very different. It's
47:27a clock free circuit. So
47:29we have, expertise at Yale
47:31on asynchronous
47:32device design, and these are
47:34clock free,
47:35chips.
47:37And,
47:38they run at a fraction
47:39of the power. The key,
47:41person for the asynchronous
47:43design is Rajit Manohar. Our
47:45other colleagues are shown here
47:46and the students and,
47:48who help with the development
47:50of devices
47:51and programming it, and Ronnie
47:53who helps us with the
47:53animal studies.
47:55So the last project, and
47:56this is just a couple
47:57of slides. So is, work
47:59that's been completed on network
48:01analysis that Tor and I
48:02worked on. And here,
48:05this is, you know,
48:08sort of just demonstrating information
48:10that we found. This is
48:11a patient with left anterior
48:13superior lateral temporal onset of
48:14seizures. There's a grid that's
48:16been placed. You don't see
48:17it. And the connectivity measures
48:19that we found
48:21are highlighted
48:22in in in the anterior
48:23part of the you know,
48:25in the right temporal lobe.
48:26This is where the highest
48:28and these this algorithm that's
48:29running up there is showing
48:31you that from any part
48:32of the area that's monitored,
48:34you can trace a path
48:35to these parts that have
48:36the highest connectivity. So there's
48:38a graded structure over a
48:40large part of cortex
48:41that shows that this area
48:43that's sort of connected
48:45and tied to the seizure
48:46onset area is several centimeters
48:48in size. And so we're
48:50we're working that up. And
48:51separate sort of aspect here,
48:53we have, created special issue
48:56on, you know, it's been
48:57twenty years since more than
48:58twenty years since Susan's paper
49:00on the network theory.
49:01It's a frontier special issue
49:04led by Klaus Leonard from
49:05University of Bonn,
49:07doctor Dan Spencer, and myself.
49:09It's close to wrapping up.
49:10Sixteen papers have been accepted.
49:12One remains in the review
49:13stages, and we expect an
49:15ebook to be generated from
49:16this,
49:18in in twenty twenty six.
49:20Finally, this is a list
49:21of collaborators whose work I've
49:23not touched on,
49:24collaborators in the epilepsy program
49:26in pediatrics and electrical and
49:29computer engineering.
49:31But otherwise, you know, these
49:33are the five projects I
49:34walked you through. I'm sorry
49:35to have, you know, taken
49:36you through so much material,
49:38but I hope you get
49:39a sense of the breadth
49:39of the work we're doing
49:40in the lab.
49:42It ranges from hardware,
49:44computation,
49:46work with patients,
49:47seizure forecasting,
49:49neurochemistry.
49:50It's a fairly broad set
49:52of work.
49:53Thanks very much for your
49:54attention.
50:03Doctor Mattson.
50:04That that was very impressive.
50:06I wanted to ask a
50:07little bit more about the,
50:09sensitivity and the method of
50:11analyzing the GABA and glutamate.
50:15Is that compared to microdialysis?
50:17Can you amplify that a
50:19bit? Yes. So,
50:22we,
50:24you know, the target for
50:26us was to achieve nanomolar
50:29limits of detection.
50:30And we with the first
50:32method that we've developed, the
50:33silicon nanowires,
50:34we are orders of magnitude
50:38better than that.
50:40So it would I don't
50:42know. I mean, Tor would
50:42be best placed to tell
50:43us how well that would
50:44compare with microdialysis
50:46and,
50:48mass spec methods.
50:49But it would be sufficient
50:51for our purposes
50:52in terms of monitoring
50:54analyte concentrations,
50:55those three analytes in the
50:57brain. It's much more sensitive.
50:59It's like hundred times more
51:00sensitive with this method than
51:02with mass spec, actually. So
51:04it's it's really messing me.
51:07Yeah. So, it's it's very
51:08good technology, and we are
51:10very grateful that we could
51:11bring it over from engineering,
51:13recreate it, and achieve these
51:15results.
51:16So so real time measure
51:18versus versus
51:20Yes.
51:21We we we have we
51:21have when we write grant
51:23applications, the reviews force us
51:24to say real time in
51:25quotes because they say it
51:26takes five to ten minutes.
51:28But compared to one hour
51:30plus the offline analysis that's
51:32done for mass spec,
51:34which can take you know,
51:35sometimes it takes us weeks
51:36or months before we see
51:37a result.
51:38This, we can propose experiments
51:39where we are monitoring an
51:40animal.
51:41So what's your yeah. So
51:43can you speak a little
51:43bit more about the application
51:45side of things? How how
51:46would you apply this to?
51:48So as I mentioned, there's
51:49a lot of interest in
51:50the research group, particularly with
51:51doctor Spencer and Torreid on
51:53understanding neurochemistry.
51:55In animals,
51:56Torre and Ronnie
51:58do brain microdialysis and subcutaneous
52:00microdialysis.
52:01And we've established certain changes
52:03that happen with during epilepogenesis.
52:06And we have established in
52:07patients certain changes that we
52:09expect to see,
52:10in the seizure onset area
52:11and well connected areas in
52:13terms of glutamate concentrations.
52:15And in the lab we've
52:16observed in animals
52:18changes in gaba concentration levels
52:20and the occurrence of seizures.
52:21Now all of that is
52:22offline analysis.
52:24This allows us to bring
52:26capability into the lab where
52:27we can monitor in real
52:29time cause the seizures happen
52:31spontaneously. But at the same
52:32time we can work in
52:33an intervention arm
52:35based on neurochemistry,
52:37you know, analysis,
52:39not just based on electrophysiology.
52:41We're also very interested in
52:42tying electrophysiology
52:43and neurochemistry better together.
52:46The electrophysiology
52:47measurements are down to the
52:48millisecond, but we can we
52:49can work on them, you
52:50know, at the at the
52:51seconds or minutes or hour
52:52resolution.
52:53And it would be very
52:54valuable for us to be
52:55able to tie our electrophysiology
52:57measures to neurochemistry.
52:59Because when we measure spikes
53:00or we measure sharps or
53:02we measure other EEG phenomenon,
53:04we don't quite understand what
53:05that means in terms of
53:06excitation and inhibition.
53:07We make certain statements that
53:09we're not absolutely sure of.
53:11So we we want to
53:12bring the two closer to
53:13one. And in terms
53:15of real world application, you
53:16know, the microdialysis
53:18done over,
53:19you know, twenty five to
53:20to thirty years,
53:23demonstrated
53:24particularly elevated glutamate in the
53:27network.
53:28And so this ties together
53:30with
53:31what a tennis talking about
53:32with
53:33network analysis because we know
53:35that if we can detect
53:39glutamate or
53:40the glutamate GABA
53:42ratio
53:43in areas that are highly
53:45connected,
53:46then
53:47that leads you to thinking
53:48about the brain computer interface
53:50modulation of those particular
53:53points in the network.
53:57Thanks, Doctor. Spencer.
54:07In terms of ICU
54:08monitoring, have you applied this
54:10to I mean, is that
54:11is that something that can
54:12be So that's what we're
54:13working with, with Emily Gilmore
54:16and Jen Kim and Bulent.
54:18Yeah.
54:19That's the target.
54:20Essentially, initially, we'll introduce
54:22a
54:23probe with pressure, temperature, oxygen,
54:25and integrating the EEG. But
54:27in the future, we would
54:28love to be able to
54:29integrate these modalities in there
54:31as well.
54:34Hi, Hal. Great. Two and
54:36of course, amazing stuff.
54:38The brain atlas, does it
54:39include support?
54:41So the shortcomings of brain
54:42atlas, we've done the new
54:44cortex,
54:45the hippocampus, amygdala, and insular.
54:47We have not done,
54:49other structures.
54:51Our thought you know, there
54:53there's a question that we've
54:53been asked about the thalamus
54:55is to take an existing
54:57thalamus sort of atlas and
54:59integrated within this. But we,
55:01you know, we we started
55:02with that and we worked
55:03with that.
55:04It's also sort of a
55:05project that's not funded. So
55:07we're waiting, we're hoping to
55:08get funding to build in
55:10these other components.
55:11We also want to understand
55:13white matter if you can,
55:15you know,
55:16parcel white matter in some
55:17manner, some logical fashion.
55:19You'd like to do that.
55:24A lot of very interesting
55:25projects.
55:26For the seizure forecasting,
55:29you know, you have, an
55:30individual's data over a long
55:31period of time, and you
55:32have these peaks where they
55:33have a change into the
55:34number of spikes that they're
55:35having. And sometimes they get
55:36seizures with the upslips, and
55:38sometimes they don't. Has is
55:39there any way to sort
55:40of extrapolate or add in
55:42patient specific data? Maybe they're
55:43sleeping less, maybe they're intoxicated
55:45in some way or any
55:47like input that would show
55:48why one spike would get
55:49a seizure type one.
55:51Right. So I'd recommend reading
55:53work by our collaborators,
55:56Maxine Baud and Vikram Rao,
55:57who've looked into this a
55:59bit more than us.
56:00For our part, what we
56:01are hoping by capturing, we
56:03ask them to maintain a
56:04diary. We have an accelerometer,
56:06so we pick up the
56:07sleep,
56:08the amount of time, yes,
56:10the sleep. We keep track
56:12of the dietary intake.
56:15We keep track of the
56:16mood. We're hoping to pick
56:18up other information that will
56:19inform the seizure forecasting we
56:21want to do. So we
56:22have in mind building a
56:23model with
56:25all the data that we're
56:26collecting and then dropping the
56:27intracranial
56:28r and s data and
56:30seeing how well the non
56:31invasive measures can also forecast
56:34seizures. We also want to
56:36mention that that project
56:37could lend itself very well
56:39to other episodic,
56:40neurological, and psychiatric disorders, and
56:43we are looking for collaborations
56:44in other areas that may
56:46want to use the framework
56:48that we've created.
56:49We collect daily multiple samples
56:52of saliva per day. We've
56:53solved it to the point
56:54we can do this in
56:54a patient's home, bring them
56:56in to the lab,
56:58put them through the analysis,
57:00mass spec analysis,
57:01which creates a very rich
57:03array of data, time series
57:05data of saliva analyte concentrations
57:08which are then put through
57:10machine learning. We built that
57:11entire architecture
57:13and it can be used
57:14for other disorders
57:16And, you know, we'd be
57:17open to talking to anyone
57:18who may have a use
57:19for it. But thanks for
57:21the question.
57:29Any other questions?
57:36Maybe can you talk about,
57:37sort of ischemia applications just
57:39for some of the stroke
57:40people in the audience for
57:41the microdialysis
57:43and molecular
57:45I was thinking about. Well,
57:46it it you know,
57:48we would love to get
57:49input in terms of design
57:52and what else we may
57:54want to measure.
57:55We've taken this project on
57:57as if we are
57:59going to produce a delivered
58:01product.
58:01And that requires us to
58:03do very clear intake in
58:04terms of user needs, what
58:06requirements there are. Because every
58:08change that we make to
58:09it reverberates throughout the design
58:11in terms of how the
58:12electrode is gonna be placed,
58:14what modalities are going to
58:15be studied, how long it's
58:16going to be used,
58:18and, you know, who's going
58:19to place it, what sort
58:20of skill set exists.
58:22And
58:23so all of those things
58:23are taken into factor. But
58:25we've we've got a tremendous
58:27engineering team
58:28and we've got we've solved
58:30many challenges.
58:31You know, I skimmed through
58:32this but there are multiple
58:33challenges that we solved here
58:35in terms of integrating these
58:36sensors.
58:37So we'd be open to
58:39other suggestions in terms of
58:40using what we've already built
58:42which
58:43could be useful.
58:44You know, the prototype pressure
58:45temperature EEG
58:47probe will be ready soon.
58:48So and the EEG measurements
58:50could be of value in
58:51terms of electrical activity.
58:53It's a full resolution intracran
58:55EEG
58:56and the pressure and temperature
58:57could be a full use.
58:59We drop the blood flow
59:00sensors,
59:01and that's something that we
59:02could bring back into the
59:03picture
59:04if need be.
59:06And, you know, at another
59:07point in time, we're bringing
59:08in oxygen sensing
59:10back into this.
59:11Yeah. So we use quad
59:12volt monitoring already in our
59:14subarachnoid hemorrhage patients.
59:16But if we if we
59:18adapted
59:18some of these monitors for
59:20some of our larger stroke
59:21patients, for instance, and think
59:23about the number monitoring, but
59:25put some of these microdialysis
59:26like, I think the chemistry
59:28for some of these patients
59:29could be really fun to
59:30think about. So I think
59:32we should brainstorm.
59:34I'm curious, like first of
59:35all, I just wanna say
59:36again what everybody else said,
59:37which is it's really, really
59:39cool to see all the
59:40amazing work you all have
59:41been doing together, and so
59:43collaborative
59:44and important.
59:46I was just curious in
59:47the landscape, is there sort
59:48of a raise to the
59:50winner? Are there other groups
59:52with similar devices,
59:54or are we kind of,
59:56you know, at a different
59:57place right now than just
59:59kinda in the broader
01:00:01global landscape on this? Right.
01:00:04So the competitive landscape is
01:00:09is interesting.
01:00:10There are a few players
01:00:12in the field.
01:00:13We are proposing a probe
01:00:14with the most modalities,
01:00:17and the solution
01:00:18in the sense that we
01:00:19are going from sensors all
01:00:21the way to display and
01:00:22analysis
01:00:22from a single
01:00:25source. For the all the
01:00:26other solutions need to piece
01:00:27together the probe with electronics,
01:00:30with a monitor,
01:00:31and it can be difficult
01:00:33for centers that don't have
01:00:34expertise,
01:00:35that don't have the expertise
01:00:36that we have.
01:00:38The probe that does have
01:00:40three modalities is made by
01:00:41a very good it's very
01:00:42good work. It's by a
01:00:43company called Raub Medic. It's
01:00:45from Germany, and it has,
01:00:47you know, it it has
01:00:48a presence in the US.
01:00:51What's unique about us is
01:00:52we have integrating the EEG
01:00:54and scalp EEG
01:00:55right from the get go,
01:00:56and it's all integrated throughout
01:00:58the solution.
01:00:58And we have a pipeline
01:01:00where we are thinking for
01:01:01other modalities, and we're working
01:01:02on other modalities.
01:01:05Yeah. This is also an
01:01:07area where the number of
01:01:07NICUs around the world are
01:01:09increasing. Their use is increasing.
01:01:11So
01:01:12the market's not very big
01:01:14if you look at it
01:01:15from a perspective of commercialization,
01:01:18but the market is unfortunately
01:01:19growing
01:01:20and,
01:01:21it's something that is a
01:01:22worldwide market and tour.
01:01:24How easily can you introduce
01:01:26the chemistry in the world?
01:01:28I think that yeah.
01:01:30So that that is a
01:01:31bit difficult because of the
01:01:33path through the FDA.
01:01:34We've got to evaluate the
01:01:36the the regulatory strategy.
01:01:38We've got predicate devices for
01:01:39every other modality that we've
01:01:41incorporated.
01:01:42But when it comes to
01:01:43neurochemistry,
01:01:44there are no predicate devices
01:01:45for the kind of sensors
01:01:46we're building. And that's something
01:01:48we are evaluating. We are
01:01:49that's the, you know, reason
01:01:50that we are taking the
01:01:51electrochemistry
01:01:52approach using silicon nanowires,
01:01:54but Jesus is also building
01:01:56up an optical approach with
01:01:58the photonic sensor because that
01:01:59might provide an alternate path
01:02:01through the FDA
01:02:03and that's something we're working
01:02:04up. And we're at a
01:02:06pretty early stage there, but
01:02:07we're hopeful.
01:02:10And what's the last out
01:02:12of curiosity,
01:02:13these
01:02:14your sort of novel neurochemistry
01:02:16sensors,
01:02:17are they at all impacted
01:02:19by the presence of heme
01:02:21and iron or other things
01:02:22that may happen in some
01:02:24brain injured regions but not
01:02:25others?
01:02:26So,
01:02:27we use
01:02:29antibodies and aptamers
01:02:32to make very specific binding
01:02:34sites,
01:02:34and that helps make it
01:02:36more sensitive
01:02:37and specific.
01:02:39But it's not to say
01:02:39that we don't have to
01:02:40test more fully.
01:02:42And,
01:02:43that's both approaches have used,
01:02:45you know, functionalization
01:02:47using aptamers and antibodies. Use
01:02:49aptamers for glutamate and lactate
01:02:51and antibody for GABA.
01:02:53We are exploring
01:02:55an approach that would be
01:02:57free of this sort of
01:02:58functionalization,
01:02:59but that would mean that
01:03:00the testing would have to
01:03:01be much more thorough and
01:03:03yeah. So yeah.
01:03:04And then,
01:03:07if I may, a second
01:03:08question.
01:03:09The the size of these,
01:03:12obviously, you're thinking about clinical,
01:03:15implementation and so they're human
01:03:17sized. But do you have
01:03:18preclinical
01:03:19sized devices as well to
01:03:20work in to test in
01:03:21models?
01:03:24Yes.
01:03:26Yes. So, you know, that's
01:03:27one of the strengths of
01:03:28the research group is that
01:03:29we've got Toraide and Roni
01:03:30Dar, and the eyed group
01:03:32does, you know, studies with
01:03:34rats.
01:03:35And the oxygen sensor we've
01:03:36developed, for example, has been
01:03:37tested in rats.
01:03:39And the
01:03:40silicon nanowire sensors, we've discussed
01:03:43testing it in animals as
01:03:44well, but we've not done
01:03:45that.
01:03:46Those can so we can
01:03:47make up, sort of
01:03:50holders and we we can
01:03:51we can come up with
01:03:52a way to introduce them
01:03:54into an animal. Yeah.
01:04:04Well, thank you very much.
01:04:06It's a few minutes past
01:04:07the hour. Thanks very much
01:04:08for your attention and your
01:04:09questions.