"Automation, Big Data, and Artificial Intelligence in OSA Management: Future (and current) Implementations" Dennis Hwang (04.21.2021)

April 26, 2021

ID6530

To CiteDCA Citation Guide

00:00I'm Lauren Tobias,
00:01an I'd like to welcome you to our
00:05yield sleep seminar this afternoon.
00:07I have a few quick announcements
00:09before I introduce our speaker.
00:11First, please take a moment to ensure
00:13that you are muted in order to
00:16receive CME credit for attendance,
00:17please see the chat room for instructions.
00:20You can text the unique ID which is
00:22listed there anytime until 3:15 PM and
00:25if you're not already registered with
00:27Yell Simi you will need to do that first.
00:30If you have questions during
00:32the presentation, please.
00:33Make use of the chat room and I
00:35will also invite everybody to unmute
00:37themselves at the end of the hour.
00:40We do have recorded versions of
00:42these lectures that will be available
00:44online within two weeks at the
00:46link provided in the chat,
00:48and finally feel free to spread
00:50the word about our seminar series.
00:52We still have at least three talks left.
00:55Two or three talks left for this year
00:57and then we take a hiatus for the
01:00summer and will resume in the fall,
01:02but we always love to see.
01:05New names and faces on our list.
01:07So now I have the great pleasure
01:10of introducing Doctor Dennis Wong
01:12as our Speaker Doctor.
01:13Huang's training has taken him
01:15across the country from being an
01:17undergrad at Princeton to a medical.
01:19His medical degree at the University of
01:22Illinois Residency and Chief Residency
01:24in Internal Medicine at Rush in Chicago.
01:26Fellow and pulmonary critical
01:28care medicine at UCLA,
01:29and finally a Sleep Medicine
01:31fellowship at North Shore,
01:32Long Island Jewish in New York.
01:35For the past 11 years he served
01:38as medical director of the Kaiser
01:41Permanente Fontana Sleep Disorder Center.
01:43He is an actor.
01:45He's very active in the American
01:47Academy of Sleep Medicine and
01:49currently serves as member of the
01:52Presidential Tele Medicine Committee.
01:54His research and clinical expertise
01:56includes the implementation of
01:58health information technology to
02:00improve the cost effectiveness of
02:02sleep care and developing artificial
02:04intelligence systems to enhance.
02:06Personalized care and provide
02:08clinical decision support.
02:09He has a strategic research award from
02:12the American Sleep Medicine Foundation
02:15to study artificial intelligence
02:17to optimize diagnosis of OSA and he
02:20also has an institutional award from
02:22Kaiser on the use of AI to improve triage.
02:26Ng of patients who have suspected sleep
02:29disordered breathing into appropriate
02:31diagnostic and testing pathways.
02:33So with that he has a very
02:36exciting topic for this afternoon.
02:39Anne Anne please join me in welcoming
02:41him to speak about automation,
02:44Big data and artificial
02:45Intelligence and OSA management.
02:47Future and current implementations
02:49that I'll turn it over to you, Dennis.
02:53Thanks so much, Lauren,
02:55and thank you for the invite.
02:57It's you know really.
02:59My pleasure to you know to be
03:02here and to present to you all.
03:04I certainly don't consider myself to
03:07be an expert and I will, you know.
03:10Certainly go into that a little bit.
03:13You know just what's influencing me.
03:15You know what my perspective is,
03:18you know as a relates to the
03:21use of automation, big data,
03:23artificial intelligence and so forth.
03:25So I am now sharing my screen.
03:30And please let me know if you can't see it,
03:33but I think you should be able to.
03:37There we go.
03:37OK, so this is my title slide and let
03:40me just kind of start flipping through.
03:42I do want to make sure that I
03:45leave enough time at the end.
03:47You know for us to have a discussion to be
03:50able to take questions you know and so forth.
03:53A lot of what I'm presenting there are,
03:55you know, vision,
03:57kind of things and.
03:59Into in some respects not
04:00a lot of concrete things.
04:02You know, perhaps that have been
04:04developed so kind of sharing.
04:06You know people's ideas about
04:07vision about the future,
04:09I think is certainly very valuable.
04:11You know, Lauren did a great job of,
04:14you know,
04:14talking about my research
04:16support in some of this stuff.
04:18A lot of this stuff that I'm
04:21presenting is going to be.
04:23Relevant Eno to the funding
04:24agencies you know for my research.
04:27So I did want to just throw that out there.
04:31OK, some things that I looted too and
04:34this is talking about my perspective.
04:36You know,
04:37to this work into this presentation.
04:39It is not as a data scientist.
04:41I don't have that kind of expertise.
04:43What I do is I have a team of data
04:46scientists and I tell them this is
04:48what I need and this is what I need
04:52from a clinician perspective to help
04:54me take care of a large volume of
04:56patients that are very different,
04:59very complex for a unique in many ways.
05:02And then my data scientists,
05:03you know, will go and try to.
05:06Start up your ****.
05:07You know that are necessary for
05:09me to achieve my goals.
05:11So you know I'm not going to
05:13spend too much time on this.
05:15Yeah, just a very quick background in
05:17regards to what is, you know, big data,
05:20you know you need to have a lot of it.
05:23It has to be diverse data.
05:25It has to be data that is actionable.
05:28Veracity is really critical.
05:29You know, we spend I spend so much effort.
05:32You know on data cleaning because you know,
05:34without good, proper, accurate.
05:36You know,
05:36I data,
05:37you know the outputs you know are just
05:40not accurate and they have to have value.
05:43And really the idea here is that
05:45we need to use big data to really
05:48help us make better decisions.
05:49So you are the objectives of
05:51you know this talk.
05:53The first is to discuss the different
05:55approaches to healthcare related big data.
05:57Secondly to talk about integration
05:59of data from various sources.
06:01How we go about doing this
06:03talking a little bit about data
06:05structure and so forth and Thirdly.
06:07What are the tools that we
06:09envision in our developing to make?
06:11Excuse me, big data useful for clinical care.
06:14So these are you know very well
06:17known challenges you know to
06:19the health care delivery system.
06:21You know it's reactive rather than proactive.
06:23It's intermittent. You know.
06:25If patient may see their doctor,
06:27you know once every three months,
06:29six months or even less frequently,
06:31you know it's less patient centric
06:33and more provider sent centric.
06:35You know we're incentivized
06:37by being more productive.
06:38Rather than, you know,
06:40trying to achieve good outcomes and when
06:42we do try to achieve good outcomes,
06:45you know that the system.
06:48You know, you know,
06:49really emphasizes you know.
06:50Achieving a good positive average
06:52population outcome rather than
06:53trying to personalize personalize
06:55outcomes for each individual patient.
06:57So this is an example of,
06:59you know, kind of what I mean.
07:01And this is a real life example
07:03from my own health care system,
07:06in which we I believe we take
07:08care of populations very well,
07:10but we don't do a very good job.
07:13I believe I've actually taken
07:15care of individual patients.
07:16So what happens here is.
07:18You know you have administrators
07:20who takes a look at our data,
07:22finds an Association with the
07:24use of hypnotic sleep medications
07:25with an increase in Ed.
07:27That's it's, you know, for the elderly.
07:29So they create these mandates and
07:31these strategies that trickle down
07:33to the provider level to reduce
07:35these prescription rate of these
07:37types of medications.
07:38By what ends up happening is that
07:40this makes my work more difficult.
07:43You know,
07:43it's just an additional metric that
07:46I have to that I have to meet and
07:49I end up having to take certain
07:51patients that are doing very well
07:54on these medications in order to
07:56in order to improve the population
07:58average outcome at the expense of
08:00some good outcomes for certain individuals.
08:03Rather,
08:03my approach to you know big data or not,
08:07my approach by.
08:08You know,
08:09I believe that philosophically,
08:11the approach to big data should be
08:14able to impact the grassroots provider.
08:16To really enhance individual physicians
08:18to care for individual patients to
08:21achieve good personalized outcomes.
08:22So here's a little bit left.
08:25A case study,
08:26and this is my health care system,
08:28Kaiser Permanente.
08:29And no,
08:30I really kind of supervised
08:31our Southern California region.
08:33We have a network of nine or
08:36ten sleep centers.
08:37You know, we are a busy,
08:40you know,
08:40center with basically about 2000
08:42referrals a month,
08:43and you know over on the right
08:46side are just some philosophical,
08:48foundational approaches to how we
08:50approach care.
08:51You know the first of which is.
08:53We certainly believe in launch to
08:55no end to end care,
08:56so we tend to be very good at
08:58the orange part, which is getting
09:00patients tested and diagnosed.
09:01You know we're even decent at
09:03initiating therapy and perhaps
09:05a little bit of follow up,
09:06but we're not very good at
09:08trying to emphasize that the
09:09care of patients actually starts.
09:11You know, before the patient is referred,
09:13even in coming in for,
09:14you know their test,
09:15and we also don't do a very
09:17good job of really trying to
09:19achieve good long term follow-up
09:21in achieving good long term.
09:23Now comes the second philosophical
09:24or foundational you know,
09:25approach to our delivery of care
09:27is really a team based approach.
09:29Even though we are very very busy,
09:32we have only really three
09:33positions and a lot of my time is
09:36actually taken up with research
09:37and administration and so forth.
09:39And so the way that we have to
09:41deliver care is through a team
09:43based approach with case managers
09:45and each of our case managers.
09:47You know they are part of a
09:49team that had their where their
09:51primary purview is one of these.
09:53Different types of services
09:54that we provide within our sleep
09:56center so we have an ambulatory,
09:58you know, testing team PSG team.
10:00We have a C pap follow up team.
10:03Alternative therapy team.
10:05We have a pediatric case management
10:08team and also a group arrested Tory
10:11therapist that provides long term
10:13lanja tude no care for our patients
10:16with chronic respiratory failure and
10:18so forth so you know it's a pretty
10:21diverse set of services that we provide,
10:24even some impatient types of services.
10:27Bedside,
10:27Poly sonography and so forth.
10:30So we recognize that you know,
10:32with this complexity of how we
10:34deliver care for a large volume
10:36of patients that we needed help,
10:39and we thought that you know big data.
10:42You know technology was going to
10:44be a key component of our strategy,
10:47and so we adopted a sleep data
10:49integration or sleep technology
10:50integration system.
10:51The system that we use is somewhere,
10:54and it's integrated about five or six
10:56of our diagnostic sleep study platforms,
10:59both PSG as well as HS 80.
11:02We've integrated, you know,
11:04the two main manufacturer,
11:06Pat manufacturer,
11:07you know platform so all the
11:09daily CPAP data is flowing into
11:12our integration system.
11:13We have integrated patient reported
11:16data from electronic questionnaires,
11:17and each of these patients have
11:20also had their entire electronic
11:22health record integrated together
11:24into a common data source to
11:27really provide a large volume.
11:29And what we believe to be a
11:32sufficiently diverse set of data.
11:34In which we can then make it actionable.
11:37We've also actually worked to
11:39integrate consumer Health Technologies,
11:40which includes Fitbit.
11:42We've integrated even something I
11:44think the the group at Yale is involved in,
11:46which is the body metrics ring.
11:49You know, oximetry and so forth,
11:51and there's no additional.
11:52You know, items that are on the road map for,
11:56you know, integration,
11:57or at least to be considered for
11:59integration. So I wanted to share it.
12:02Maybe you know, a couple of
12:04concrete examples in regards to.
12:06You know just the power of having
12:08you know data integration in a
12:10automated fashion that includes all
12:12of the patients that essentially
12:14come into one of our sleep centers.
12:16They immediately become,
12:17you know, a data point.
12:19So for example, when Covid hit,
12:21we were able to pivot actually
12:23very quickly to say, hey,
12:25we've got this really robust data set.
12:27Let's take a look at covid outcomes related
12:30to sleep apnea and where you can see here.
12:33And this is statistically significant
12:35that patients who have untreated.
12:36You know, sleep apnea.
12:39You have a higher risk of getting
12:41covid compared to those with
12:43no obstructive sleep apnea,
12:45and as the pap appearance increases,
12:47their risk actually will then decrease.
12:49We also have found that as a sleep
12:52apnea severity increases, you know.
12:54So, for example,
12:55severe obstructive sleep apnea,
12:56how it compares to to mild sleep apnea,
12:59that severe sleep apnea has the
13:01highest risk for patients to
13:03actually contract COVID-19 infection.
13:05Here's a forest plot which actually shows
13:07after you know all the various adjustments.
13:10Bird characteristics and so forth.
13:12You know that you know there's
13:13a number of different you know,
13:15patient characteristics that are associated
13:17with a higher risk of getting kovid,
13:19and you can see here that
13:20those who are well treated,
13:22meaning they're using their path at least
13:244 hours a day during the pandemic period,
13:26have a lower risk of getting coded.
13:29Interesting Lee lower age.
13:33Also has a reduced risk of getting covid.
13:38We found certainly racial disparities.
13:40You know, Blacks and Hispanics have a
13:43higher risk of getting covid higher BMI.
13:47More existing core abilities,
13:48all of which were both of these
13:50conditions or characteristics,
13:51increases risk of getting covid,
13:53so there's a lot more to this,
13:55but I just wanted to share with you.
13:57You know, the power of, you know,
14:00being able to collect data,
14:01big data, you know,
14:03through a real world clinical environment
14:05and and how that data can be quickly
14:07utilized to really be able to,
14:09you know, you know,
14:10for research as well as you know,
14:13we believe for clinical purposes.
14:14So now that we have all this data.
14:17I'm gonna pivot,
14:18you know,
14:19to talking about big Data Tools
14:21and this is really important and
14:23really a large you know portion of
14:25my efforts because if we have big
14:27data you know and we don't have
14:29the tools to make it actionable.
14:31We just end up drowning in data
14:33and we and we get paralyzed because
14:35we have so much data.
14:37We don't know what to do with it.
14:39So the four that I'm going to
14:41mention here are number one,
14:43automation and more specifically
14:44to improve patient engagement,
14:45#2 population management tools,
14:46#3 remote patient monitoring.
14:48And #4 artificial intelligence.
14:49So what do I mean by automation?
14:52I mean there is a zillion different
14:54examples of how
14:55we can and already are using automation,
14:58but this is just one example
15:00that I wanted to present.
15:02This is a a diagram that you know
15:05reveals kind of a manual process of,
15:08you know, you know RPM.
15:10Remote patient monitoring.
15:11The CPAP data goes to the cloud wirelessly.
15:14You have a provider who
15:16reviews the data manually.
15:17If the patient is not doing
15:19well on half therapee,
15:20then will initiate and manual encounter.
15:22This, however,
15:23is a very very labor intensive process,
15:25and so we look to see whether we
15:28could automate this and so the
15:30data still goes into the cloud.
15:32We apply these automated algorithms
15:34with customizable thresholds and you
15:36know if the patient is not doing well.
15:38For example,
15:39three nights in a row.
15:41They get a text message and
15:43completely bypasses any kind of
15:45manual intervention unless necessary
15:46by the sleep center provider
15:48and we further personalize this
15:51based on other characteristics,
15:52for example,
15:53whether their cardiovascular risk
15:55score is elevated or whether their
15:58effort is effort is is elevated and
16:01you know a different set of messages
16:03you know can then be tailored and
16:06delivered to the patient to really try
16:09to address you know their specific.
16:11Risk profile and so we did do a study.
16:14This is a Tele OSA trial that was
16:17published in the Blue Journal about
16:19two or three years ago and not going
16:21to spend too much time on this.
16:24But you can see here and I'll just
16:26have you follow this line above
16:27that those who receive messaging
16:29and continue to receive messaging
16:31through this dotted line compared to
16:33those who don't receive messaging
16:35have significantly improved Pap use.
16:37You know, we just this very simple feedback,
16:40patient engagement,
16:41type of you know mechanism without
16:43any additional provider intervention
16:45and so we determined this to be
16:47what we thought was a very cost
16:49effective type of intervention.
16:51So there's a zillion different
16:52examples of automation,
16:54but wanted to move on to tool #2.
16:56How do we use?
16:58You know what kind of tools can we use
17:01to make big data more manageable and
17:03so the second tool that I wanted to
17:06present is population management dashboards.
17:08And this is a way for us to organize our,
17:12you know,
17:13Brazilian patients you know with Brazilian,
17:15you know pieces of different
17:17information and what it does is
17:19this dashboard organizes you,
17:21know for us the different
17:23characteristics identifies those
17:24who are risk and tells us which
17:27patients need to be followed up with.
17:29And, you know,
17:30here's a kind of an example or a diagram
17:33of what our follow-up protocol is.
17:36C Pap is prescribed over here.
17:38On the left side,
17:39and we have various.
17:41Checkpoints get them each patient
17:43meets certain kinds of checkpoints.
17:44They will pop up onto our task list
17:46when they are determined to be at risk,
17:49and then we have a different
17:51set of algorithms for the post.
17:53You know,
17:53three month period.
17:54So this is a way you know for the system
17:57to tell us who do you need to follow up with,
18:01because you know these are the patients
18:02who are at risk while allowing us
18:05to manage patients by exception.
18:06So for those patients who are doing OK,
18:09we just passively follow them and.
18:11And do not initiate an encounter now.
18:14One of the things that we do is
18:16we've created different algorithms
18:17for different cohorts of patients.
18:20So for example,
18:21those who are commercial drivers.
18:24You know, have a very different
18:26instead of algorithms you know.
18:28And even to the point where we've
18:30created a an automated process where
18:33commercial drivers every three months,
18:36they have A and they get an email
18:38with their three month package
18:41here and report delivered,
18:43and it's completely automatic,
18:45it's completely automated.
18:46So they can take this to their occmed
18:49physician and to be able to continue driving,
18:51and so you know, we've created a
18:53system here that is fully customizable.
18:56And you can develop different algorithms
18:59for different cohorts of patients.
19:02Uhm?
19:04You know we did a time motion,
19:07you know, study where we looked at,
19:10you know,
19:10trying to do population management
19:12using a manual process universe
19:14versus this more automated and
19:16even semi intelligent process.
19:18And we were able to find that there was
19:21a 83% improvement in efficiency by being
19:24able to utilizes system that enhances
19:27our ability to manage patients by exception.
19:30We do have a trial that is incorporating a
19:33lot of things that I you know talked about.
19:37This is,
19:37you know,
19:38we're hoping to submit this to the NIH,
19:41a collaboration between US and and Upenn.
19:44We're developing A3 arm randomized
19:45control trial in which we are looking
19:48to see whether automated management,
19:50which is an enhanced version of
19:52the Tele OSA study that has a
19:55new wants delivery of education,
19:57engagement, motivational enhancement,
19:58and even CBT over two years is going to.
20:01Improve long term adherence,
20:03you know, compared to usual care,
20:05and then the third arm is combining this
20:07automated management as well as Savior.
20:10A nuanced case management strategy,
20:12whether that also is going to be
20:14a superior to usual care in both,
20:17will see how the outcomes compared
20:19to automated management.
20:20So that's trial that is in.
20:23You know, that is in development,
20:25but kind of incorporates the first two tools.
20:28The big data tools that I presented.
20:31Tool #3 is remote patient monitoring,
20:33so you know we've integrated sleep
20:35activity trackers near those data.
20:37Now that even suggests that you know
20:39Fitbit is more accurate than actigraphy,
20:41and so we've you know pivoted and
20:44switched over to Fitbit as our
20:46as kind of our primary method of
20:48being able to track sleep.
20:50If we do need some type of
20:52actigraphy information,
20:53and this is the body metrics ring,
20:55which is a oximetry ring and
20:58we've been using this.
21:00Even during covid you know tracking
21:02our patients with chronic respiratory
21:03failure and being able to provide
21:05remote care for these patients.
21:06You know, I even had a couple of friends.
21:09You know who end up getting kovid and I
21:12gave them this oximetry ring and for two
21:14weeks you know I logged on every day to
21:17take a look at their data and to make sure
21:19that I didn't need to end up, you know,
21:22escalating care you know for my friends
21:23we can embed that information you know
21:25into our population management dashboard.
21:27You know we can superimpose path usage
21:29with sleep durations, so now we not
21:31only know how much they are sleeping.
21:34But we can actually look at the
21:36relative amount of PAP usage compared
21:38to their overall sleep,
21:39you know and so forth, so you know.
21:42Again, being able to incorporate this into
21:44the patients population management profile,
21:46you know it really organizes, you know,
21:48the display unit of information for
21:50us in a way that you know makes it,
21:53and you know, hopefully more efficient.
21:56In terms of, you know,
21:58other ways that we're you know.
22:00Looking at the use of RPM in,
22:02you know this is a study that is
22:04in development where we're going
22:06to be combining.
22:07You know,
22:07this is just a proof of concept here,
22:10but you know combining path usage.
22:13You know, use longitudinally overtime.
22:16In comparing it directly with
22:18simultaneous integration of blood
22:20pressure monitoring and you can see in
22:22this example as the pack usage decreases,
22:25the blood pressure starts to trend
22:27upwards and that creates more
22:29actionable and perhaps more engaging
22:31types of information for the patient.
22:34With the, you know,
22:36effective really trying to directly
22:38improve hypertension management over
22:39on the right side is an integration
22:42of various types of RPM.
22:44You know,
22:45you know.
22:45Technologies and when all the data seems
22:48to be trending in the wrong direction,
22:51perhaps we can intervene before the
22:53patient ends up in the hospital.
22:57So in putting this together,
22:59you know this is kind of our philosophical
23:02framework or backbone of how we approach
23:05care using big data you know to help
23:08us at the at the top to identify
23:10patients who are risk for you, know,
23:13obstructive sleep apnea before the patient
23:15comes in for their diagnostic testing.
23:17We will, through our system,
23:19automatically send web education as well as
23:22an intake questionnaire that's auto linked.
23:24The appointment patient gets diagnosed.
23:26Now PAP is initiated and then the
23:29first three months our efforts are
23:32really trying to get the patient
23:35successfully on boarded with Pap therapy.
23:38After the first three months,
23:40we transition to perhaps a more balanced
23:43approach where we're still obviously
23:45trying to maintain Pap adherence,
23:47but we are also trying to then implement
23:50additional strategies to really be able
23:52to optimize comorbid clinical outcomes
23:54and to have an interdisciplinary impact,
23:57and so you can see there's multiple areas
23:59in which automation is implemented where
24:02we use population management dashboards,
24:04and you know where we are able to
24:07provide remote patient monitoring.
24:10So that's this leads us to tool number 4.
24:13You know artificial intelligence.
24:14You know if you want a second opinion,
24:17I'll ask my computer and we're
24:18probably getting to that point.
24:20Actually,
24:20pretty pretty soon and wanted to kind
24:23of dig into this a little bit further.
24:25You know, for those who attended,
24:27you know ASM Destructors you know kind
24:29of shared this example a little bit.
24:31That AI has been oversold.
24:33You know, very good at chess,
24:35you know Watson very good at Jeopardy,
24:37but not so good at health care, you know?
24:40And there's a really great article
24:42by Lisa Strickland.
24:43Talked about how IBM Watson really
24:45overpromised and in this article you
24:47know really talked about, you know,
24:49here's some key catchphrases, I suppose,
24:51in terms of why Watson failed.
24:53You know it's now being sold because
24:55it just really hasn't made the kind
24:58of impact that they had envisioned.
25:00It's difficult to build an AI doctor.
25:03You know the bulk of its,
25:06you know information that it used,
25:08was unstructured information,
25:09such as doctors,
25:10notes and literature articles.
25:12It does not have a product to
25:15analyze medical images in.
25:16Pattern recognition is something
25:18that AI tends to be very good at,
25:20and it hasn't proven that will
25:22actually do something useful and very
25:24little in the way of demonstrating
25:26that AI can improve patient outcomes
25:28and save health system money.
25:30Or I can distill it down into this,
25:32which is application of AI has been
25:34too focused on mimicking rather
25:36than complementing provider work,
25:38and so how is Sleep Medicine trying
25:40to address these barriers effectively
25:41in order to make AI really effective?
25:44Impactful,
25:44useful in number one is creating a
25:47standardized framework for structured data.
25:49I can assure that with your ready.
25:52With all the technology integration
25:54that you know we've been able to
25:57develop to create specific clinical
25:59decision support tools that complement
26:01and enhance clinician work.
26:02To leverage pattern recognition
26:04to transform our definitions of
26:06sleep disorders in outcomes.
26:08To enhance patient interchange
26:09experience to increase engagement and
26:11also to automate clinician tasks such
26:14as documentation which tends to be
26:16very tedious and very labor intensive
26:17so you know I've already demonstrated
26:19this for you but you know Sleep
26:22Medicine is in some respects in an
26:24enviable position because, you know,
26:26we are so very much data driven.
26:29We have a lot of data and in were able
26:32to collect a lot of this data even from
26:34our manufacturers you know wirelessly.
26:37So we are.
26:38You know, overall you know in a
26:41pretty good position to be able
26:43to create a standardized set of
26:45structured data that can be useful.
26:47#2 wanted to talk a little bit more
26:50about clinical decision support,
26:51so over on the left side is
26:53our longitudinal care pathway
26:55components that I shared earlier,
26:57and in each of these components is
26:59it is an inflection point in which
27:02a clinical decision has to be made,
27:04and each of these components is
27:07very amenable to having big data.
27:09And artificial intelligence provide
27:12us support.
27:14You know guidance,
27:14recommendation in terms of what
27:16to do about a particular patient,
27:18so here's some areas that we're working on,
27:20and this is actually a little
27:22antiquated 'cause we're working
27:23on more than just the four that
27:25I've highlighted here.
27:26But you know, some examples are who's really,
27:28who's really at risk for sleep apnea.
27:31And in different populations.
27:32For example,
27:33the surgical population and so forth.
27:36You know who should be triaged and
27:39undergo home testing versus PSG.
27:41You know what is the real risk
27:43of cardiovascular disease in
27:44developing novel metrics?
27:46Doing a better job of phenotyping patients.
27:49Can we tailor what type of
27:50therapy that patient should be on?
27:52Perhaps the patient you know is
27:54better off going directly to an
27:56oral plants right away?
27:57Can we individualize pep adherence
27:59targets for patients so that we're not
28:01fixated on 70% four hours for everybody?
28:03Can we predict when a patient is going
28:06to fall off the tap before they actually do?
28:09So how would we utilized this?
28:11And here's one example you know
28:13over on the left side is you know
28:15a patient who is green and we
28:17predict good papet hearings.
28:19But if they are not adhered,
28:20we're going to spend the
28:22effort to troubleshoot CPAP,
28:23whereas the patient over on the right
28:26side we predict that this patient is
28:28going to struggle with pap therapy.
28:30And if they are not inherent instead
28:33of really trying to bang your head
28:35against the wall and risk losing
28:37time and losing patient engagement,
28:39this is a patient that will divert to
28:42some other type of therapy more early on,
28:45or what we can do is create
28:48individualized management prescriptions
28:49for patients you know through.
28:51You know a series of you know,
28:53characterizations.
28:54You know,
28:54a different characteristics and prediction,
28:56you know metrics.
28:57So, for example,
28:58patient number one you know.
29:00Has a high risk of cardiovascular disease,
29:03is predicted to do well with C pap and
29:06this is a patient that we're going
29:09to initially path therapy in target.
29:11A high adherence and provide
29:13the necessary follow up,
29:15whereas patient #2 is mostly symptomatic.
29:17Should respond to an oral appliance.
29:19It has a low PAP adherence prediction
29:22score and this is a patient that we will,
29:25perhaps, you know,
29:26divert to oral appliance therapy.
29:28You know right away instead of.
29:31Wasting our effort without therapy.
29:34So that's, uh,
29:36approach #2 approach #3 to making
29:39artificial intelligence useful is to,
29:42you know,
29:43leverage something that AI is very good at,
29:47which is pattern recognition.
29:49So here's one example.
29:51You know, daily stepapp data.
29:54You know,
29:55can we use AI to help us preemptively predict
29:58when a patient is going to lose adherence?
30:01An engagement with pap therapy so that
30:03we can maintain engagement rather than
30:06you know what I think is probably
30:08harder to have them lose engagement
30:11and then have to re engage them.
30:13Or can we use pap therapy data?
30:16To be able to preemptively determine
30:18when a patient is at risk for
30:21hospitalization so that we can intervene
30:23before the patient ends up in the hospital.
30:26Another I said of effort that we
30:29are engaged in in regards to,
30:31you know,
30:31pattern recognition is by including
30:33the use of raw tracings.
30:35You know,
30:36and our initial step is the raw PS3
30:39tracings and the raw hsat tracings and
30:41throw it into our machine learning
30:44data set and create new metrics
30:46that are much more informative
30:48about cardiovascular risk.
30:49You know,
30:50neurocognitive impairment
30:51and response to therapy,
30:52and so enter data is a group that
30:55we're working together with a call.
30:57This dynamic phenotyping and,
30:58you know, an effort that you know
31:00we're very much engaged in,
31:02and this is what it would kind of look like.
31:05This is kind of a very famous image of,
31:09you know, some.
31:11You know artificial intelligence?
31:13You know, experts that have you know,
31:16identified various clusters of
31:18diabetes subtypes.
31:20And in a way,
31:21we're looking to be able to do something
31:23very similar so that instead of relying on,
31:27you know the a try to infer.
31:30Risk you know,
31:31and we know that the HY does a
31:33very poor job of that.
31:35Whether we can actually be able
31:37to identify different subtypes of
31:38obstructive sleep apnea that are much
31:40more meaningful for clinical management.
31:42So we work together with enter data and I
31:45don't want to take any credit for this.
31:48'cause this is all enter data here is,
31:51they've actually taken,
31:52you know, you straw tracings.
31:53You know, e.g.
31:54Tracings and company and you know,
31:56provided some complimentary EHR
31:58information they have come up with
32:00something called the Brain H-index.
32:02So in the middle here you can see that
32:04there is a lot of different clusters.
32:06You know a different.
32:08Subtypes of EG.
32:09Waveforms,
32:10patterns and this brain age index.
32:13After adjusting for,
32:14you know a number of different factors,
32:17you can use it to predict.
32:22What the person's gender is.
32:24Whether they have depression,
32:25whether they're sleepy,
32:26have impairment in concentration and memory,
32:29whether they have sleep apnea,
32:30and you know other.
32:32You know clinical conditions,
32:33and so you know this is just an
32:36example of some plima Neri work.
32:39You know that really looks at.
32:42You know?
32:43Artificial intelligence to
32:45identify patterns of different
32:48sleep disorder subtypes.
32:51Uhm?
32:54I'm going to move on to a fourth tool
32:56or approach to artificial intelligence,
32:58but I wanted to present this to you.
33:01You know quickly as a you know,
33:04kind of as a transition into the next
33:06slide I mentioned previously how we
33:08were able to use our big data set to,
33:11you know, Pivot really quickly
33:13to looking at covid outcomes.
33:15The other thing that we've done is
33:17we started to look at the risk of
33:21acute cardiovascular event. You know,
33:23over even less than a one year period,
33:26you know with about 46,000 patients.
33:30And so acute cardiovascular event was defined
33:34as an episode of either a heart attack.
33:38Unstable angina stroke and one
33:40other thing can't quite remember,
33:42but anyways those were the key conditions
33:44that defined a cardiovascular event and
33:46where you can see here is that for the
33:50mild obstructive sleep apnea patients,
33:52that hazard ratio really wasn't too much.
33:55Change the P value you know showed that there
33:58was really no significant relationships,
34:00but from moderate to severe obstructive
34:03sleep apnea you can see that those who.
34:07I have moderate to severe sleep
34:10apnea but are not using Pap.
34:13Compared to those who had moderate
34:15to severe obstructive sleep,
34:16and I'm sorry, compared to those
34:17who have no obstructive sleep apnea,
34:19the hazard ratio is.
34:20You know quite a bit higher,
34:22and that was statistically significant.
34:25When the patient uses path,
34:27but to a moderate degree,
34:29the hazard ratio drops quite
34:31substantially and significantly,
34:32and the certainly the group that
34:35you know did the best for those who
34:38use PAP for at least four hours.
34:41In the hand hazard ratio drops
34:43you know to an even greater.
34:46Uhm?
34:46You know,
34:47yeah,
34:47FX size and when we had the greatest
34:50effect size and this was adjusted for
34:53a number of baseline characteristics.
34:56So you know we were able to determine
34:58from our, you know, big data set.
35:01You know pretty quickly you know
35:03some of these really significant
35:05associations and so that leads to.
35:08Approach #4 how can we use
35:10artificial intelligence?
35:11Well, our next step is to look at, you,
35:14know, some of these relationships again,
35:17but using machine learning
35:18and then being able to,
35:20you know,
35:21develop cardiovascular risk profiles
35:23for patients and then be able to
35:25model for the patient that if you
35:28use PAP you know your risk of
35:30cardiovascular disease is going to
35:32go from 60 to 35 and if you use or
35:36appliance it's going to go to 45.
35:39Not real numbers that you just made
35:41up by just for conceptual purposes.
35:43And so, let's say the patient,
35:45does you know select C pap as
35:48the patient meets their target.
35:50You know, based on our machine learning,
35:52you know relationships that have
35:54been identified,
35:55that risk or is going to constantly
35:58update and decrease as the
36:00adherence targets targets are met.
36:02But if they're doing poorly,
36:03then the score will,
36:05you know will then increase,
36:07and so being able to use API
36:09to gamify patient
36:10engagement, you know for
36:12patients is another approach.
36:14That you know is,
36:16you know is just part of our one of
36:20the efforts that we're engaged in.
36:23Another example of patient interchange
36:26engagement using an AI interface
36:29is by utilizing an automated bot.
36:34And in this case you know this is.
36:37You know what has seen Globat,
36:40but it's a kind of a CBT for depression.
36:45No AI bot. That I,
36:48you know, it's pretty slick.
36:49It it's pretty easy to engage with,
36:51you know, and there are some people you
36:54know many people and and I've been reading.
36:56Read an article about
36:58how some people in China,
36:59you know they work very hard and
37:01they would actually rather engage
37:03with an artificial companion
37:04than a real life companion,
37:06so I'm not sure if that's
37:07really a good thing, you know.
37:09But the point here is that you know you can,
37:12you know?
37:13Utilizing an AI bot that collects a
37:15number of different holistic inputs and
37:17provides a number of different holistic.
37:19Outputs to provide a constant
37:23interactive health companion.
37:26You know that complements near the
37:28intermittent nature you know of.
37:30You know,
37:32provider encounters and so that's another
37:35kind of effort that we are engaged in.
37:39Uhm?
37:42How can a I, you know help us?
37:46Care more efficiently, you know,
37:47provide care more efficient,
37:49efficiently and effectively,
37:50and in the fifth thing that
37:52I'll mention here is to really
37:54help us with Commission tasks.
37:56So another thing that we are working
37:58on is helping us generate auto generate
38:01a clinician note based on the patients.
38:05You know electronic questionnaire
38:06information as well as you know,
38:09you know other sets of
38:11information from the patient.
38:12For example there are EHR
38:15comorbidity information and so forth.
38:18You know Epic is trying to work on.
38:20You know these advanced,
38:22you know voice recognition software
38:23you know to be able to generate notes,
38:25but I just have a hard time seeing
38:27that that's going to be successful.
38:29You know,
38:29anytime soon you know when I'm in clinic,
38:32there's a lot of stuff you know
38:33that we're talking to patients
38:35about and and don't want that.
38:36You know most of that to be
38:38in the clinician note.
38:39And how would you organize it?
38:41I don't think this is a much better approach,
38:43which is effectively to effectively to
38:46have the patient generate the note for us.
38:48I I kind of mentioned this.
38:50You know being able to take a
38:52diverse set of information to,
38:53you know,
38:54automatically triage patients to what
38:56type of testing do they need with what
38:59type of protocol you know and so forth.
39:01Another example of clinician tasks,
39:03and I'm not going to spend too much
39:05time on this 'cause I already talked
39:07about population management dashboards,
39:09but what I've really mentioned,
39:10you know too.
39:13You know,
39:14some teams is what I need is a
39:17system that is going to tell me.
39:20Who do you need to care for and how?
39:23You know, and you know.
39:25Obviously it's it's meant to be a compliment.
39:28You know we're not going to dumb down,
39:31you know provider,
39:32you know work and expertise and so forth,
39:35but having this set of you know
39:38automated intelligent processes
39:39to to automate. You know patients who
39:41are at risk to provide suggestions you
39:44know based on in overwhelming set of
39:47data that a human just is not able to.
39:50You know that you know to you know to.
39:53You know view and process all that data.
39:56You know, for patients it you know,
39:58we certainly envision that having this
40:01type of system to tell us you know who's
40:04at risk and how to manage a patient is a
40:07you know is in clinical decision support
40:09tools that we believe can be very,
40:11very useful and effective, and make our
40:13care delivery a lot more efficient.
40:15So there's a kind of summarize the
40:17pension potential of big data here.
40:19You know over on the left side,
40:21our current challenges, and I think you
40:24know we can use big data to really address.
40:27Each of these challenges.
40:29So that our care is more proactive.
40:33You know it can be, you know more
40:37continuous rather than intermittent.
40:39We can personalize the management
40:41you know better for each patient.
40:44Uhm?
40:44You know,
40:45within a population management framework
40:47in to achieve optimal outcomes,
40:49you know that are tailored for
40:52each individual patient.
40:53So big Data has the potential to transform,
40:56you know,
40:57Sleep Medicine requires application.
40:59The tools which currently exist.
41:01How to make data useful
41:03for sleep specialists.
41:04They can empower sleep specialist
41:06with organizing patient data,
41:08providing clinical decision
41:09support to improve outcomes,
41:10provide care more efficiently.
41:12Thirdly,
41:13it has the potential empower patients
41:15with a more personalized approach
41:16with continuous care in fourth,
41:18artificial intelligence can potentially
41:20enhance virtually all aspects of
41:22of care along the end to end care.
41:24You know, spectrum.
41:25I I had to speak with the mirska Brown from,
41:29you know the NIH yesterday who can
41:31help a little meeting about big Data.
41:33And you know what the NIH can
41:36do and so forth.
41:37And one of the things that I you
41:40know really kind of emphasized in my,
41:42you know,
41:43kind of follow up email to her
41:45is that you know while you know
41:47some of this really advanced,
41:49you know big data machine learning
41:50stuff to look at various kinds
41:52of things are really interesting
41:55and scientifically necessary.
41:56You know,
41:57I cannot emphasize that.
41:58I thought that it was necessary
42:01that it's necessary to.
42:02You know to have a balanced approach you
42:05know to really be able to look at how AI.
42:09Can be supported.
42:10The development AI can be supported in
42:13a way where it can directly impact.
42:16Clinician work you know and and
42:18patience you know and and make
42:21you know I work more translatable
42:23into real world settings and so
42:26in a way that I think summarizes.
42:28You know my philosophical approach.
42:30You know, to this entire area,
42:33and you know, here's a number of different,
42:36you know,
42:37partners that have been very
42:40important in just.
42:42Collaborating together on
42:43this particular work,
42:44so I'll go ahead and stop
42:46right here and I'll
42:48turn this over to like probably you learn
42:51and happy to take questions. Great,
42:54yeah, thank you so much and everyone
42:56feel free to post questions and chat.
42:59We can open it up to
43:02questions to Doctor Wong.
43:03That was a fantastic overview.
43:05I think it really illustrates
43:07how much potential there is in
43:10Sleep Medicine for improving our
43:12care processes for patients and.
43:14Really, how uniquely suited sleep
43:16is with the amount of data that
43:19we're getting off devices to
43:21really tailor care to to patients?
43:23You know, one of the things that you
43:26alluded to with use of somewhere
43:28which got me excited 'cause the
43:31VA is about to be transitioning
43:34to somewhere nationally.
43:35As you probably know,
43:37for across the country was.
43:39And that we are good at identifying
43:43the patients who are not using
43:46PAP or at least a lot of the.
43:49The device manufacturers have been
43:51incentivized already to develop
43:53that on local device levels and
43:55to tell us who is not doing well
43:58and and what their adherence is,
44:00but I think for a lot of us the
44:03challenge is what to do with
44:05that at that point and when does
44:08intervention make the most difference,
44:11and what should the intervention look like?
44:14An obviously you're studying
44:15a lot of this stuff,
44:17but I guess just from a
44:20practical standpoint and.
44:21And you know,
44:22do you have any any particular insights,
44:25not necessarily data driven,
44:27but you know from your experience
44:29about what you recommend to centers
44:32that don't have as much data easily
44:34accessible and streamlined, is Kaiser.
44:37But we can get this kind of
44:40superficial level of hey,
44:41these are the patients who are non adherent.
44:44We can't build an app or a bot to
44:48remind all our patients by text,
44:51you know.
44:51That they should be using it and
44:54calculate their STD score.
44:56But should they be getting a phone
44:59call from an RT or or?
45:01What is the highest yield intervention?
45:03I don't know what would you
45:05recommend as a consultant,
45:07say to practices with with less integrated
45:09data than Kaiser.
45:11Yeah, I really good question and you
45:13know and and I'll just throw out.
45:16You know, maybe some you know some you know.
45:19Key examples you know that could be useful.
45:23You know the first of which is
45:25even just this automated process of
45:27delivering text messages to patients.
45:29You know automated automatically
45:31every three days.
45:32If they're not doing well,
45:33you know from our study we determined
45:36that there is a substantial,
45:38and you know statistically significant
45:39improvement in happen here,
45:41and so overtime you know without any
45:43additional provider intervention you know,
45:45so you know,
45:46I think that is certainly one tool
45:48you know that can be utilized.
45:51Part of our work,
45:52also in terms of developing machine learning.
45:55You know clinical decision support tools.
45:57We are actually also,
45:58you know creating prediction models
46:00based on different sets of information.
46:02So for example,
46:03you know trying to determine if a
46:05patient is going to do well on pap therapy,
46:08or if you know timing of when
46:11they're going to not do well and
46:13lose engagement with pap therapy,
46:15you know we're using different
46:17sets of information.
46:18So for example,
46:19you know we may just isolate
46:21the PAP adherence.
46:22You know, you know daily CPAP data right?
46:25Or you know we can,
46:27you know take that along with
46:28a sleep study data right?
46:30Or we can take that and even
46:32add to an EHR data right?
46:34And so we can have these different,
46:36you know prediction models based
46:38on different sets of data.
46:39So for assistance that have a fully
46:41integrated you know data set,
46:43you know some of these you know you
46:45know being able to add in available
46:48data you know very well could be useful.
46:50But for some systems in which
46:52maybe the data is more isolated.
46:54And maybe a little bit less diverse.
46:57You know,
46:57prediction algorithms may still be useful,
46:59you know, because of,
47:01you know our approach to.
47:02You know,
47:03creating different prediction
47:04models based on what the you know
47:07available data you know might be
47:08for you know any given practice.
47:12Gotcha, thank you.
47:16Hey Dennis, nice talk.
47:18This is Andres in truck from Yale.
47:22I also like your house,
47:23looks really great.
47:25It's a virtual background, 'cause
47:27if I turned it off you'd see that
47:28I'm actually standing in my closet.
47:32You got it. You got an awesome closet.
47:35I I think it's you know I.
47:38I agree with you wholeheartedly.
47:40That I think that we probably
47:42overshot expectations for AI in
47:44the last five or ten years or so.
47:46And maybe it's really best suited to
47:49make us cyborgs rather than replace us.
47:53As as you're suggesting,
47:55and so you know,
47:57I guess I did want to ask a question
48:01about the slide where you had looked
48:05at brain age with the ENSO data folks.
48:09And just curious,
48:11is that metric generated through?
48:13What feature is it selecting some
48:16specific features from the EG?
48:19Or how is that brain age generated?
48:22One of the challenges with AI is in
48:26machine learning is that sometimes the.
48:30The mechanisms or the pathways by
48:32which marker is determined are not
48:35very clear and still I'm wondering,
48:37is there some physiological
48:39possibility to that brain age?
48:41At that we can glean from what
48:43you guys have done within so data.
48:46Yeah, that's a good question.
48:47Really. Great question as always.
48:51And in I'll kind of emphasize a little bit.
48:55Again, my ignorance and in regards
48:58to actual machine learning work.
49:01And I and I and then sedated.
49:04People are really the people
49:05that you know, did this work?
49:07And so I don't want to claim, you know,
49:10too much or really if any credit you
49:13know for this now you know what they did
49:16was you know I'm I don't know whether
49:19they were able to identify specific patterns.
49:21You know that they were able to
49:23recognize through this, you know,
49:25dynamic dynamic phenotyping,
49:26you know, type of process, you know.
49:29But it was e.g EMG, EOG.
49:31And in select you know HR information.
49:33Essentially, Unicorn abilities,
49:34age, demographic and so forth.
49:36They did a bit of supervised,
49:38you know learning you know to,
49:39you know,
49:40create their initial model and
49:42then they you know did a bit of
49:44clustering and so that can that that
49:46middle graph that I showed you with
49:48those you know pretty colors and
49:50almost would even looks like a brain
49:52stem you know was the clustering.
49:54You know that looking at the
49:56different phenotypes but you know
49:58by itself you know we don't know
50:00whether it's useful or not.
50:01You know we need where?
50:03Proceeding with trying to validate this,
50:05you know,
50:05with within a new and much larger data
50:08set you know within you know within
50:10Kaiser and then being able to you
50:12know do that type of validation and
50:15then additional validation work to
50:16see whether you know how to implement
50:18this or whether it's even going to
50:20be useful for actual implementation
50:22in a real world clinical setting
50:24I think is still up in the air,
50:27so it's really just kind of,
50:28you know step one of I think
50:30a much longer journey.
50:32Oh yeah
50:33no, it's just a very cool concept of looking
50:35at something that is a better marker of.
50:38Outcomes in patients.
50:40Then what we look at traditionally?
50:43And I guess the last one ask for more
50:46question about the oximetry sensor
50:47that you were using and whether you
50:50found that to be useful at predicting
50:52who's going to compensate and who from.
50:54You know your product failure.
50:55Patients needs to be addressed ahead of time,
50:58and if So what was your experience with them?
51:02Yeah, we have. I think about and I think
51:04you you know you've been playing around
51:06with the device as well and so I think
51:09at this point we have about 60 patients.
51:11Who are, you know, in this particular pilot.
51:14And we found it to be useful.
51:16I don't. I wish I, you know,
51:18it's been awhile since I've
51:19presented the data presented to,
51:21you know, one of our committees,
51:22you know to get formal.
51:25Engagement with you know you
51:26know contract ING formally.
51:27You know for this you know device
51:29that we're still under a pilot status
51:32before we can replicate it around.
51:34You know the country really?
51:35You know for Kaiser, but when we do,
51:37I think we'll have, you know much better.
51:39You know data to be able to work with.
51:42But in terms of our preliminary,
51:43you know, kind of experience,
51:45you know, I wish I had the numbers,
51:47you know, easily pulled up,
51:48you know, for you,
51:49you know we found it to be #1
51:51very successful.
51:52You know, in that you know 90 plus
51:54percent of patients were able to.
51:56I think 95 plus percent of patients were
51:58able to use it and to be able to put on
52:01their cell phone and to have you know,
52:03data coming through.
52:05Secondly,
52:06we found it to be useful in a wide set of,
52:10you know type of you know
52:12clinical circumstances, you know.
52:14So for example,
52:15being able to use a short term
52:18just to determine if the patient is
52:20responding adequately to either oxygen
52:23or noninvasive ventilation you know.
52:26You know, you know that that's been useful.
52:28We've also found it useful.
52:29You know,
52:29for long term management and
52:31so for some of these patients,
52:32we just, you know,
52:33even though we call it an indefinite loner,
52:35you know there's just code for saying,
52:37you know, you get to keep it just
52:39you just have to use it, right?
52:41You know.
52:42And so if you don't use it,
52:43we're going to take it back.
52:45You know, 'cause I into providing
52:46incentive for them to actually use it.
52:48So we were protocol in which you know,
52:50we asked the patient to use it
52:52three times a week at least,
52:54and then obviously, if they are.
52:56Looking like there are risk,
52:57you know, a little bit borderline.
52:59You know we asked him to wear it everyday,
53:02you know and so forth.
53:04And we found it.
53:06You know very,
53:07very useful for things like
53:08titrating how much oxygen they need.
53:10You know,
53:11for determining or feeling more secure
53:13that the patient is not at risk and
53:15for whatever reason I cannot recall
53:17any patients that started to decompensate.
53:20And we're like, oh boy,
53:22you know your oxygen.
53:23Your ring data is starting to
53:25turn in the wrong direction.
53:27You know we need to come in.
53:30And we need to intervene 'cause you're
53:32about to end up in the hospital.
53:34And I don't know whether
53:35it's just because we got
53:36lucky or I don't know whether it's
53:38maybe using this ring, you know type
53:41of thing and and work close TPN,
53:43you know with the ring were able
53:44to optimize their treatment and so
53:46they're they're not decompensating
53:48as as frequently or maybe so I.
53:50I think it's probably probably
53:51more than former who were just
53:53getting lucky at this point.
53:55But we certainly anticipate that
53:56you know we're going to be able to
53:58eventually recognize patients who are
54:00decompensating be able to, you know,
54:01provide early intervention and to keep
54:03them keep them out of the hospital.
54:08Great, thank you so much.
54:09I think given the time we
54:12will will wrap up with that.
54:14I just want to let everybody
54:16know about our last three talks
54:18for the year before we have our
54:21internal sleep Jeopardy in June.
54:23So we have three talks that are all
54:26actually on hypersomnia coming up
54:28and next week we're going to hear
54:31from one of our clinical fellows,
54:33Doctor Otukolo who's going to
54:35speak about RBD in narcolepsy.
54:38And then stay tuned for.
54:39I'll announce the the talks for
54:41the final two weeks after that,
54:43but again, thank you so much.
54:45Doctor Huang.
54:45That was really incredibly
54:47informative inspirational,
54:47I think.
54:48Gave all of us a lot of ideas for
54:50how we can be taking better care
54:52of our patients and processes.
54:55We can consider in our own practices.
54:57So thank you so much.
54:59Yeah, thank you everyone for the invite.
55:02Great to connect with you all.
55:05Thank you, thanks so much.
55:07Great take, care see you
55:08ever see you next week?