"Machine Learning in OSA: Dreaming Towards the Future" Miranda Tan (11/03/2021)

November 12, 2021

ID7154

To CiteDCA Citation Guide

00:11Alright, so good afternoon everyone.
00:12I'm going to start with a
00:14few quick announcements before
00:16I introduce our speaker today.
00:18First, the Sleep seminar lectures are
00:20available for CME credit when viewed
00:23in real time and to receive credit,
00:25just text the ID for the lecture to Yale
00:27Cloud CME by 3:15 PM today and there'll
00:30be more information on how to do that.
00:32Showing up in the chat and recordings
00:34of these lectures are available
00:35approximately 2 weeks after other
00:37lecture at the site noted in the chat.
00:40On these viewings are not
00:41available for credit.
00:43If you have questions during the talk,
00:44please just use the chat and we
00:46will address them at the end,
00:47and please keep your
00:49microphone muted otherwise,
00:50so now it is my pleasure to
00:53introduce today's seminar speaker,
00:54Dr Miranda Tam.
00:55Doctor Tan is the director of
00:57Sleep Medicine at Memorial Sloan
00:59Kettering Medical Center and
01:01an instructor of medicine.
01:02At Weill Cornell Medical College?
01:05She received her medical degree from
01:06New York College of Osteopathic
01:07Medicine and then moved to New
01:09Jersey for internal medicine,
01:10internship and residency at Rutgers.
01:13After that,
01:14she served as chief resident for
01:15quality and Patient safety and
01:17internal medicine at Dartmouth.
01:19And she completed her pulmonary and
01:21critical Care Medicine Fellowship
01:22at Thomas Jefferson University
01:24and her Sleep Medicine Fellowship
01:26at University of Pennsylvania.
01:28From there she moved to Memorial
01:29Sloan Kettering,
01:30which she founded the Sleep Medicine
01:32program within the pulmonary service,
01:34and her work has led to the
01:36revision of the Presurgical sleep
01:38apnea screening guidelines there.
01:40Plus,
01:40she has established several clinical
01:42patient pathways with other programs,
01:44including with the male
01:46sexual reproductive program,
01:47the Integrative Medicine program,
01:49and the bone marrow Transplant
01:51Survivorship clinic,
01:52among other awards that she's received,
01:54she received a junior Faculty Development
01:56award at Memorial Sloan Kettering.
01:58She's an active member of the ATS and
02:00is currently serving on a working
02:02group to develop an official ATS
02:03research statement regarding cancer
02:05related fatigue and lung cancer.
02:07She is also a member of the American
02:09Academy of Sleep Medicine and a Fellow
02:10of the American College of Chest Physicians.
02:13Her research focuses on the
02:14predictors of obstructive sleep
02:15apnea and cancer patients,
02:17specifically using machine learning
02:18and on the prevalence and severity of
02:21obstructive sleep apnea in men with
02:24polycythemia on testosterone therapy.
02:25She presented 4 abstracts at
02:27the most recent ATS,
02:28including one entitled Machine Learning
02:30validated screening tool to predict
02:33obstructive sleep apnea in cancer patients.
02:35So we are so fortunate to have
02:37Doctor Tan joins us today to
02:39discuss machine learning and OSA
02:41dreaming towards the future welcome.
02:46Thank you for the introduction. Dr.
02:48Hilbert and good afternoon everyone.
02:50Thanks for tuning in.
02:51I'm delighted to present my talk,
02:53machine learning and OSA
02:55dreaming towards the future.
02:58So as the title suggests,
02:59the purpose of this talk is to provide
03:01an overview of machine learning and
03:03discuss how we can integrate Emil to
03:05improve our understanding of OSA and
03:08hopefully move the needle closer towards
03:10personalized medicine for our patients,
03:13specifically the objectives of our
03:15discussion will be to review basic
03:17approaches of machine learning,
03:19described potential data sources
03:21for mill in obstructive sleep apnea,
03:23and identify opportunities for mill in OSA.
03:29As for obligatory disclosures,
03:31I have none, but I can offer you
03:33the code to obtain CME credit.
03:35That's 28436.
03:37That's again, that's 28436.
03:41As for disclaimers,
03:42I am not a data scientist.
03:45I did some programming in college
03:47where being referred to as novice
03:50would be an overstatement.
03:52The information I am presenting
03:53to you today is to the lens of
03:56a clinician with a background
03:58through collaboration with data
04:00scientists and computer engineers.
04:02And my second disclaimer for
04:03you is that this is going to be
04:06a Prezi presentation format,
04:07so in contrast to traditional
04:10Microsoft PowerPoint,
04:11this has an interactive zoom
04:13which may induce nausha.
04:14So sorry in advance if you
04:16develop motion sickness.
04:20So the applications of AI and milk
04:22are ubiquitous throughout society.
04:24Then say I, Netflix knows which shows to,
04:27recommend to us it can redirect my
04:30car when it's drifting out of lane,
04:32and Alexa can understand my mumblings
04:35after I've trained it to understand me
04:38over the past two years. But you know,
04:41before discussing the prospects for OSA,
04:43we should start with some basics to ensure
04:46everyone in our audience is on the same page.
04:49So AI and Emily are often used
04:53interchangeably similar to COPD,
04:55chronic bronchitis, asthma, COPD etc.
04:58Yeah, we all know there
05:00are subtle differences.
05:01Artificial intelligence is defined as the
05:05simulation of human intelligence by machines.
05:08Picture of this AI is is like this
05:11large umbrella term within AI.
05:13There is a machine learning machine
05:16learning or the algorithms that can
05:18recognize trends and patterns from data.
05:23And then there's deep learning.
05:25So deep learning is a subspecialized
05:28form of machine learning.
05:30And what it does is it is it's an
05:33artificial neural network with
05:35many layers to adapt and it learns
05:38through complex patterns of this
05:41from this high volume of data.
05:43And when I say high volume,
05:44I mean like 5:50, that's a lot of data.
05:50So here is a rudimentary example
05:53of machine learning,
05:54so we can tell we can tell the machine
05:57that we want to know what a dog is,
05:59and then we can train it by saying
06:01this is what a dog should look like.
06:03And then what we do is we can
06:05present a new image.
06:07And then the machine can learn and say,
06:08well, I think this is also a dog.
06:13Then we can do.
06:13Next is give this a this picture of a
06:16cat and then the machine will tell us no.
06:18This is not a dog.
06:19It can't tell if it's a cat,
06:20but it knows that this is not a dog.
06:25So even though our talk is not to
06:29focus on deep learning, I really do
06:32think I should digress for a minute.
06:34To show you this image of
06:35deep learning and action.
06:37This is a really hot topic in the
06:39tech world and I I do think that
06:41this can penetrate Sleep Medicine
06:43sometime in the near future.
06:45So there are two distinct differences
06:48between machine learning and deep learning.
06:50Although traditional machine
06:52learning does not require hardcoding,
06:54it does require defined features of interest,
06:57such as age, race, etc.
07:01Deep learning, however,
07:02leverages these neural networks to
07:04learn the relevant features or patterns.
07:07So in deep learning,
07:09the first layer not shown here may
07:12just be a series of circle and dots.
07:14OK, then what deep learning will
07:16do is then look at the next layer.
07:18So for example here on this first
07:21column we have that top row that what
07:23deep learning will do is look at that
07:26top row and say that's a face and
07:28then in the second column it'll say
07:30well this could be a car and then the
07:32third column could say this is an elephant.
07:34And then we'll do it.
07:35Apply another layer after that,
07:37and then it'll start understanding
07:39that while this is a face,
07:41this is a car.
07:43This is a chair.
07:45This is my deep learning,
07:47requires 5:50 amount of
07:48data to make inferences,
07:50whereas in machine learning this
07:52can be done and performed on a
07:55smaller pool of patients i.e.
07:56You know 200 to 300 patients for data.
08:03OK, so we Speaking of data,
08:05you know what is big data right?
08:08We frequently hear the term in
08:11the context of AI MLDL big data,
08:14it described by the 5DS.
08:16That's volume velocity,
08:19veracity, variety and value.
08:23So volume speaks for itself.
08:25You need large volumes in order
08:27for it to qualify as big data.
08:29Velocity speaks to the speed.
08:31I wish the data is.
08:33Acquired so if it takes you two years to
08:36extract the data then it's not big data.
08:38It should be instantaneous turnover.
08:41Veracity describes the quality right?
08:44So we really need good quality data,
08:46so if any if I learn only one
08:49thing from the data scientists
08:50and engineers that we work with
08:53is that garbage in garbage out.
08:55If we feed the machine with bad data,
08:59we are most likely going to
09:01receive a poor algorithm,
09:02so we really must enforce a quality data.
09:06Next just as important we have variety
09:09in data in order for the machine
09:12to determine the best algorithm.
09:14It requires an ecosystem of multiple
09:17different interactions simultaneously.
09:19This kind of what this kind of like
09:21diverges from what we are familiar with
09:23in traditional medicine where it's like OK.
09:25Let's focus and hyper focus on
09:27this one facet whereas this one
09:30heavily relies on variety and
09:32understanding multiple different
09:33interactions to understand what is
09:36prominent and what is significant?
09:38So now I feel like I've been
09:41belaboring the point of quality and
09:43unfortunately loss in the term.
09:45Big data is the importance of quality
09:49data and the volume of data often
09:52overshadows the quality of the data,
09:54which is the perfect for the
09:56for the poor algorithm.
09:57So I really wanted to drive
09:59home this point of quality data.
10:02And so you'll see my akawa.
10:04He is the editor in chief
10:06of the Molecular Brain.
10:08When my account did is that he reviewed
10:10181 AI manuscripts this past year in 2020.
10:1540 of them were deemed too good to be true,
10:18and he questioned their authenticity and
10:20what he did is he requested the raw data of
10:24a of the 40 studies to assess its integrity.
10:27Half were withdrawn because
10:29the data couldn't be provided.
10:31Of the other twenty remaining 19
10:34were rejected either because of
10:36insufficient raw data or the data
10:39mismatched with the prediction results,
10:41or the output could not be reproduced.
10:44When we tested with the validation test sets,
10:46so ultimately only one of the 40 were
10:49accepted. So don't waste your time.
10:51Just really make sure that the quality of
10:53the data is is the best that you can provide.
10:58OK, so now we're armed with the
11:02definitions of AI, ML and DL.
11:04We know what big data is and we
11:07stress the need for quality data.
11:09Now, how does this all work together?
11:11So here's a bird's eye view
11:14of the process itself.
11:15So first we start with the big data,
11:18either obtained via HR.
11:20This could be imaging could
11:22be genomic sequencing,
11:24whatever have you and we input it into AI
11:28UMSL and then finally we get our outcome.
11:30We get the diagnostic accuracy.
11:33We get prediction models.
11:35We get workflow efficiency and
11:38precision medicine simple enough.
11:40OK, maybe not that simple,
11:42so let's get a little bit more granular.
11:46So how does machine learning work?
11:49So Machine learns through training.
11:53Training is an iterative learning
11:56until the best model is found.
11:59So let's say we want to predict OSA.
12:02OK first, what we need to do is
12:06to classify OSA via features.
12:09In order to do this,
12:10we will tell the machine.
12:12That OSA isn't hi greater than five
12:16and non OSA isn't ahi less than five?
12:20And that's our label of interests
12:22with the definition of interest.
12:24The next thing we will do is we will feed it.
12:27We will feed our machine with a variety
12:30of patient characteristics such as age,
12:33BMI, gender, next size, anything.
12:37From there on,
12:39these features will be analyzed.
12:43I use that term features interchangeably
12:46with patient characteristics because
12:47features is a term that's used in the world.
12:50So anyways,
12:50these patient characteristics or
12:52features will then be extracted.
12:55One example of feature extraction
12:57is principal component analysis,
13:01principal component analysis.
13:03What this does is it determines
13:06what is the best fit of this feature
13:09to match our label of interest.
13:11So for example.
13:12If we were to use PCA to extract
13:15the features and create components.
13:18What we will find is that maybe age
13:21greater than 50 will match closely with OSA.
13:24Maybe BMI greater than 35 will
13:27match closely with OSA.
13:29After employing this feature,
13:31extraction of the most salient components,
13:34then vectors are formed.
13:37These vectors, what they are,
13:39is really an interaction of
13:41all the key components.
13:43So another example is in my
13:46identify an old obese man.
13:48Or am I identify a post menopausal
13:53woman with history of radiation
13:56to the head in the neck.
13:58So basically the more vectors we
14:01can create based on the more data,
14:03the better our algorithm will be,
14:05because these vectors will then be fed
14:07into the machine learning algorithm.
14:10I'm at the machine learning algorithm
14:12will do is it will process these vectors
14:15together form clusters and then it
14:17will develop the best classifier model.
14:20So with this best classifier model will do?
14:23Is that it'll say, well,
14:25we can most likely predict OSA to this
14:29maximum degree 98% using this model.
14:32But we need to know that this
14:34model is indeed accurate,
14:35right?
14:35So then what we will need to do
14:38is to see if we can apply this
14:41best model to actually predict
14:43our desired outcome being OSA.
14:46So then what we do is we give it like
14:48what we call the validation set,
14:50otherwise known as a test data set.
14:52What this is is something unseen,
14:55completely naive.
14:55We feed it again to the machine
14:58via data input.
15:00Again,
15:00the features are extracted and
15:03then feature vectors are formed.
15:05Then finally we want to note if
15:07the predicted label is achieved.
15:09If the predicted label in this
15:12case OSA is achieved,
15:14then we know that this is a very good.
15:16Algorithm.
15:21So to summarize, we start with data we
15:25start first with a data training set.
15:28It gets fed into the algorithm.
15:29There's an evaluation,
15:31and a model is formed.
15:33Then we repeat with a test data set,
15:35otherwise known as a validation data set,
15:37and essentially it's just naive
15:39information fed into the machine to
15:41see if the model that we built was
15:43correct in achieving our prediction.
15:47So how do we know that this is not by chance?
15:50How do we know that this algorithm is indeed
15:53reliable outside of this one data test set?
15:56Well, let me just repeat just we want
15:59to make sure we have reproducibility
16:01and we want to make sure that we
16:03don't have what's called overfitting.
16:05Overfitting is essentially.
16:09Us saying that yeah,
16:10this is a perfect study,
16:12but only for our small cold water.
16:14So the more data that we feed
16:15them or test datasets to show
16:17that if the prediction can again
16:19be achieved and achieved then we
16:21know that we're not overfitting.
16:27So based on the shared
16:28process of machine learning,
16:29all of us can appreciate that
16:32machine learning is vastly different
16:34from evidence based medicine.
16:36One is really not better than the other.
16:38It's almost like comparing
16:40apples and oranges, right?
16:41But acknowledgement of the differences will
16:44allow us to embrace machine learning as
16:46one of our future options and Sleep Medicine.
16:49So, specifically, PBM,
16:51as we know, is hypothesis driven,
16:53versus machine learning,
16:54is data driven right?
16:56EBM we have to account for confounders,
17:00we have to eliminate confounders.
17:01So then there's tends to be less
17:03variables and a lower diversity,
17:05whereas machine learning heavily
17:07predicated itself on a high diversity,
17:10so we need more variables
17:12to sort out the noise.
17:15EBM what it does is it compares groups to
17:19infer causation whereas machine learning
17:23relies on clustering to infer causation.
17:27Next, a medium,
17:28its success susceptible to bias
17:31because of evidence hierarchy.
17:34Whereas in machine learning
17:35there is no hierarchy,
17:37right?
17:37So no hierarchy and we're possibly
17:40eliminating the risk of bias outside
17:43of the data collection process itself.
17:46And then finally in EBM.
17:48We have confidence in the evidence
17:50with more studies and also
17:54replication of the study outcome.
17:57Similarly to some degree in machine
18:00learning we have confidence in
18:03repetition by increasing the
18:04training datasets and I just
18:06mentioned and then also feeding
18:08it with multiple validation sets
18:10to ensure that we didn't overfit.
18:12So Apple and orange they're different,
18:14but they're both good for you.
18:18OK, so that was fun.
18:20Now let's transition into
18:22machine learning in OSA.
18:24Here we have an English bulldog.
18:26Can anyone unmute and tell me
18:27why we have the English bulldog
18:29as my mascot for Melon OSA?
18:34Anyone?
18:39OK, well I'll tell you so the bulldog
18:43snores probably has sleep apnea, maybe.
18:46Well, my internal machine thinks
18:48I've always say every time I see
18:51a bulldog because of this story.
18:53And also I think of this.
18:55So the the field of OSA is.
18:58It's truly right for machine
19:00learning algorithms and recognition.
19:02It is prime for mill and possibly a
19:04shoulder ahead of other subspecialties
19:06due to its abundance of data sources
19:08both in the medical level for the
19:10patient in the form of EHR data on
19:12diagnostic sleep study path data and
19:15at the digital health product level for
19:17the patients as a consumer in the form
19:20of Fitbit and sleep apps and bed sensors.
19:23So the options for data or boundless,
19:26the hint of its success relies on quality
19:28data collection that I've been mentioning.
19:31Integration, transformation,
19:32and interpretation.
19:34Biostar clinicians only clean,
19:37usable and meaningful data.
19:38Can create value for our
19:40patients in the future.
19:42So what types of sleep
19:43data do we have access to?
19:47So here's the first class of data.
19:49We have state technology, so to the non
19:52sleep physicians in the audience we have
19:54a polysomnogram on the left and we have
19:57a home sleep apnea test on the right.
20:00So the polysomnograms are very
20:02rich in raw data, so here.
20:05We are able to note their their neural
20:07status were able to see limb movements,
20:09were able to see heart rate and oxygenation,
20:12but you know what?
20:13We use less than half of this
20:16to determine what is OS.
20:17What is OSA we use the hi.
20:20How much of the inside is gathered
20:23from this sleep study right?
20:26So the PSG itself lends itself to,
20:29you know, dynamic phenotyping like why
20:31can't we use symptoms in addition to the
20:35raw data to help us understand more of OSA?
20:38How do we subtype patients
20:40based on this data?
20:42I mean, if we truly think about it,
20:44the the ISR,
20:45like the interscore reliability,
20:47it just shows us that we cannot best
20:49determine what N one and two is like.
20:51We need more people,
20:53so why can't we just push this a
20:55step forward into machine learning
20:57and have the machine tell us
20:59this is truly an one based on
21:01all of these previous findings?
21:04And even on the right we had
21:06the home sleep apnea test,
21:07so age sites have been accelerated into use.
21:11I'd say definitely during the pandemic,
21:13and even with the age that we have
21:15a reasonable amount of continuous
21:16data for just one night of sleep.
21:21I'm in terms of.
21:23You know objective long term monitoring.
21:26I think we're really in a good
21:28position compared to other types
21:30of chronic diseases because we have
21:32the ability to watch patients every
21:35night and gather objective data
21:37regarding their PAP compliance.
21:40Also, throughout the course
21:42of their chronic disease.
21:43So for those that are unfamiliar,
21:46here is an example of a
21:48PAP compliance report.
21:48Were able to determine their dates of usage,
21:51their hours of usage,
21:53their average pressure per night,
21:55even if they had a leak there.
21:57Hi, this is we have access to all of this,
22:01we just need to analyze it together
22:04and figure out what's most meaningful.
22:10Another huge data source for
22:13Sleep Medicine, particularly OSA
22:16as a consumer sleep technology.
22:19We are swimming in data
22:21and the patients love this.
22:23I can't say that all clinicians office,
22:25but definitely the patients love this.
22:28The most common consumer sleep tech that
22:30we are probably familiar with is the Fitbit.
22:33So the Fitbit on utilizes these
22:36Tri axial accelerometer sensors
22:38to determine like heart rate.
22:41It's basically like actigraphy
22:42where we uses motion sensor,
22:44so it's determining like,
22:46well this is activity so the patients
22:49likely not sleeping at this time
22:51and there are other flavors of
22:53this two in terms of wearables.
22:55There's also the oral ring,
22:56which is the upper right.
22:57Diagram.
22:59And then now recently there is
23:01a bin new rebels.
23:02So what new rebels do is they have
23:05this lower energy radar that detects
23:08movement and breathing the diagram
23:10to the bottom left hand corner.
23:12That's the Google Nest hub.
23:14So the Google Nest Hub recently
23:16launched as a sleep sensing feature,
23:18which estimates when you went to bed
23:20when you woke up and how long you slept.
23:23It can also detect sounds like
23:25snoring and coughing as well
23:27as environmental features such
23:29as wide in room temperature.
23:31What these sensors are designed to
23:33do is to help assess sleep quality
23:35and help identify potential causes
23:37of sleep disruption in the morning.
23:40The Nest Hub will show your sleep
23:42summary and then the sleep data
23:44syncs with Google Fit app.
23:46So after learning your sleep
23:48habits and patterns through AI,
23:50the sleep sensing will then give
23:52personalized recommendations
23:53as a matter of fact,
23:55Google partnered with the ASM to deliver
23:58these recommendations tailored based on this,
24:01the prior night's sleep.
24:03So,
24:03for example,
24:04it might say you are not sleeping
24:06this much considered going to
24:08bed at a regular time each night.
24:10Pretty crazy.
24:13Then finally the third class
24:15of data for sleep that is near
24:18and dear to us all is the EHR.
24:21From HR we can gather a
24:24demographics socioeconomic status,
24:26even comorbidities, lab data,
24:29imaging, hospitalizations,
24:31health care utilization,
24:32self reported questionnaires,
24:34and then we can multiply the data by
24:36combining it with the sleep technology
24:38that we previously mentioned,
24:40as well as consumer sleep technology.
24:44So we have all this data,
24:46so here's a comprehensive summary
24:48slide of all our big data sources,
24:50challenges and opportunities.
24:52I borrowed this from Peppin.
24:56So to recap, we talked about the Patella
24:58monitoring as well as a sleep study,
25:01but also social media.
25:02I know social media can also affect
25:05sleep if you are tweeting as our former
25:08president at 2:00 o'clock in the
25:10morning 3:00 o'clock in the morning,
25:124:00 o'clock in the morning.
25:14Then most likely you are going
25:15to have a sleep disturbance.
25:17Now, how will this affect your
25:18PAP usage if you're supposed
25:19to be on C PAP for tweeting at
25:203:00 o'clock in the morning?
25:22This is all valuable data for
25:24us to help identify patients.
25:26Even before they start therapy.
25:29Other things to consider.
25:30The omics and I think we may have
25:32mentioned and lifestyle activities so
25:33we know in the sleep world that patients
25:36that are more active during the day,
25:37they're they're more likely to
25:39get a better night's sleep.
25:43And Geo localization.
25:44We mentioned socioeconomic status,
25:47how that can affect OSA, but then
25:49also access to care or pollen counts.
25:52Imagine that.
25:53So now on your phone you can detect
25:56what is the pollen count outside.
25:58Should I bring my Flonase in the morning?
26:00How will this affect OSA patients?
26:03Well patients with OSA and allergic rhinitis
26:06may have decreased pain here and right there,
26:10congested at night.
26:11They don't want to use their nasal.
26:13Ask they're congested at night.
26:14They can't breathe through their mask.
26:15We can identify this,
26:17and we can refine our therapy based on
26:19what we know from all of this data,
26:22and consolidating every piece together.
26:28But life isn't easy, right?
26:30This data is heterogeneous.
26:32How do we ensure interoperability
26:35of all this data?
26:36So we have one Amar?
26:38Why can't Sloan talk to Yale?
26:40Combine our data and then
26:42figure out the best solution.
26:43Once the interoperability like unfortunately
26:45we have Allscripts I'm not sure about Yale.
26:48If you guys have Epicor,
26:49Allscripts but that's always comes into
26:52question interoperability also data privacy.
26:55So recently Google health.
26:57I think they tried to work with
26:59EMR and there was a lot of
27:01pushback in terms of data privacy.
27:03So their project has stalled.
27:07Next, we want to think about,
27:08you know the the natural inherent flaws
27:11of the observational studies right?
27:14And then finally,
27:15if we have the proper
27:17infrastructures for data sharing.
27:22But you know, we should definitely
27:24commit to moving big data and
27:26machine learning and OSA four.
27:28This can really help us reshape our OSA
27:31through integrated care and could help
27:33us a partner with other institutions.
27:36And then also maybe even use this
27:38information that we learn through big data
27:40and machine learning to create proper
27:42risks for a therapeutic interventions.
27:47And that all of this is that
27:48patients so patients at the
27:50center so patient centered care.
27:54OK, so now that we have the data,
27:56how can we make meaningful use of this
27:59data so one common tool is automation.
28:02Automation is basically the
28:04extrication of humans and piles
28:06of paper as much as possible.
28:09So we can achieve this through continuous
28:12remote telemonitoring and feedback messaging,
28:15so that's one form of automation.
28:18And this is an example of
28:20it via the telly OSA trials.
28:23This is crying and Co.
28:24He sought to examine the
28:26effects of telemedicine,
28:27delivered OSA education and CPAP
28:30telemonitoring with automated
28:32patient feedback messaging on CPAP
28:34adherence is a four armed a randomized
28:37factorial design clinical trial and
28:39he enrolled about 14150 patients.
28:43So for all intensive purposes,
28:45this figure it collapses the four
28:46treatment arms into three to directly
28:49compare the automated feedback messaging
28:50and what he found was that here,
28:52like you can't see that arrow here.
28:54OK, so here what he saw is that the the
28:57patients who received messaging from
28:59the onset of CPAP and throughout were
29:02more likely to be compliant compared to
29:04those without messaging after one year.
29:08This was proven to be
29:10statistically significant,
29:11so we learned that there's a positive impact
29:13on those who receive continuous messaging.
29:15Interestingly,
29:16those who initially received messaging.
29:20Right,
29:21but they stopped after 90 days.
29:24Had similar outcomes compared to
29:26those who never even received a text.
29:29If I were to extrapolate this
29:31to my real life,
29:32then I should nag my husband to
29:34the dishes every day if I want
29:36him to wash the dishes right?
29:38So.
29:40Next up, instead of you know,
29:44maybe continuous remote telemonitoring.
29:45What if we use self reported
29:48questionnaires and link them to
29:50sleep studies or appointment types?
29:52How can we repair this automatically
29:54to remove the paper waste and to better
29:57understand our parents our patients?
29:59And then finally we can consider alerts
30:03based on consumer technology data.
30:05So for example,
30:06during the pandemic what we did
30:08is we gave these little pulse
30:10oximeters for our patients,
30:12sent them home and we would call them.
30:13We had like an army of nurses called patients
30:15every day, checking in on their symptoms.
30:18You know now.
30:20Shouldn't we think about?
30:22Advancing shouldn't we think
30:24about Bluetooth technology right?
30:25What about home oximeters?
30:27An alarm system can be prepared,
30:30and what we can do is we can be alarmed
30:33every time their auction dips below 85%.
30:36For example,
30:37this already exists for blood pressure.
30:39This already exists for glucose monitoring,
30:42would it benefit the sleep apnea
30:45patient to watch all of these metrics
30:48while they're being treated at home?
30:50Hopefully you know if we start
30:52using Bluetooth technology more
30:54transporting the information to a cloud,
30:56automating alerts we can interact
30:58with our patients better,
30:59provide better feedback,
31:01better engagement,
31:02and hopefully better outcomes.
31:07OK, let's see.
31:12OK, so here is another example
31:14of an integration tool HL seven.
31:17OK so HL 7 otherwise known as health level 7,
31:21but this is a conduit between
31:24platform A and platform B.
31:26In terms of Sleep Medicine,
31:28it can create a bidirectional
31:30passage of information of sleep
31:33data and HR data and vice versa.
31:36So when I was a fellow at Penn,
31:38we utilized the HL 7 framework to
31:41identify patients with insufficient
31:43PAP usage in high risk patients,
31:46commercial transport operators,
31:48and also we use this data to identify
31:51who had high HI despite path usage.
31:54What we did is this is the prequel era.
31:56By the way,
31:57we generated letters not trusting
32:00regenerated letters to send them to patients.
32:03Then what we did is we wanted to
32:05see if the patients actually called,
32:06we measure call volumes to see
32:08if patients actually came back
32:11to clinic and to understand if
32:13this affected their a clinical
32:15outcomes such as blood pressure.
32:17And lo and behold,
32:19we found that patients with higher PAP
32:21usage was linked to lower blood pressure,
32:24diastolic blood pressure.
32:27So this is just one example of
32:28what we did a couple of years ago
32:30and other things to think about.
32:32Easy integration tool is a
32:35population management database query.
32:37You can query info and combine it
32:40with general sleep diagnostics,
32:42consumer health technology
32:44and development algorithms to
32:45predict outcomes and performance.
32:52So currently at MSK we are
32:55completing a study of ML and OSA.
32:57So we had about 300 or 400 patients in
33:00sleep clinic that referred for home sleep
33:03study and we wanted to determine the
33:06predictors of OSA in cancer patients.
33:09The first step was do we really need
33:12machine learning data scientists are
33:14not cheap and can we simply use our
33:17beloved logistic regression to determine
33:19predictors of obstructive sleep apnea?
33:22So we answered our first question by
33:24testing to see if there is a linear
33:28relationship between characteristics
33:30otherwise known as features and OSA.
33:32So the first three subplots we have here
33:35depict the features scattered in 3D space.
33:38What this shows us is that there's a
33:42nonlinear relationship between each of
33:44the features and obstructive sleep apnea.
33:47So we've tried using chronic kidney
33:49disease with stop bang score,
33:51diabetes, all the comorbidities you
33:53could think of COPD metastasis and
33:57we found no linear relationship.
34:01We this was an iterative process.
34:02We kept feeding it more data and more data.
34:05Still no relationship.
34:07So I suppose we can still continue
34:10with logistic regression,
34:11but that would be obviously flawed
34:13since we confirmed the nonlinearity of
34:16features we sought to use unsupervised
34:19machine learning and employed advanced
34:22techniques such as PCA combined with RF,
34:25so its principal component
34:27analysis and random forests.
34:29So the PCA,
34:30what it did is it performed feature
34:32extractions to determine the relevant
34:35components and then these components.
34:37Were processed to yield clusters or feature
34:40vectors that are most relevant to our
34:43population and there we have the 4th subplot,
34:45the 4th.
34:46So plot is a projection of the
34:48features as principal components
34:50in three dimensions that contribute
34:52to the maximum variance of OSA.
34:54We mentioned that the algorithm
34:55will do is it will try to help us
34:58predict OSH the best of their ability
35:00based on the data that we fed it.
35:02So within 93% we were able to determine
35:04this within this Max variance,
35:06we will have OSA.
35:09And then finally the best classifier
35:11was determined for subsequent datasets.
35:17So I will share our preliminary findings
35:19with you because you sacrificed your
35:21Wednesday afternoon to join me here today.
35:24So through mill we learned that the
35:26strongest predictors of obstructive
35:28sleep apnea and cancer patients were
35:31stopping score radiation therapy to
35:33the head and neck and cancer type.
35:35Meaning. Specifically,
35:36it was long head and neck and prostate.
35:40So here is a diagram that
35:42illustrates the airflow limitations
35:44in obstructive sleep apnea.
35:45So in figure a,
35:47this is the normal sleep we have
35:50air that enters through the nose
35:52and then down the posterior or
35:54fairings and trade into the lungs
35:57and figure B is that of sleep apnea.
35:59So in figure B we can see the error
36:01is trying to enter through the
36:03nose and then there is relaxation
36:05instruction and the posterior or Franks,
36:07thus causing the chronic
36:10intermittent hypoxia.
36:11In obstructive sleep apnea.
36:13And finally,
36:14in see this is a diagram of a patient who
36:18received radiation to the head and neck.
36:22Alright,
36:22so similarly air will enter
36:24through the nose and then it
36:27it will counter obstruction.
36:28This obstruction can be from
36:30the posterior or pharynx,
36:31or it can be in the form of
36:33fibrosis or scarring or stenosis.
36:35Even after radiation therapy.
36:37So this radiation therapy
36:39can increase risk for OSA or
36:42exacerbate pre-existing OSA.
36:48Alright, so this just goes
36:49to show you our results.
36:51We have 100% sensitivity and 90%
36:54specificity on the PCA plus RF was
36:57able to determine the Max variants
36:59of OSA through clustering and
37:01this was indeed superior to the
37:04traditional techniques of LR or RF.
37:07Individually we've said it multiple
37:09tests thereafter and were able
37:11to reproduce the same result
37:13of OSA accurately and again.
37:15This was unsupervised learning.
37:20So an ML is applicable from diagnosis
37:23to intervention ING to you know
37:26long term monitoring, right?
37:28So we tend to fixate on diagnosis,
37:31but let's shift towards this
37:33chronic management, right?
37:34Like how do we improve compliance
37:36for our patients using milk?
37:38How much compliance is even necessary?
37:40So we have this large advantage of the path.
37:42And I think we should try to leverage
37:44this in the future for studies to
37:47improve management of our patients.
37:49Next thing to appreciate that
37:51you know the purpose of mill.
37:53It's a compliment that physician
37:55not replace a physician.
37:56IBM Watson failed.
37:58This is daily and everyones memory right.
38:01But you know we just have to work with it,
38:02right?
38:02We have to be able to translate
38:04the structured data instead of
38:06something valuable and actionable
38:08information for our patients.
38:09And then finally, you know,
38:11we should always strive for
38:13aiming for clinical value,
38:15establishing these maybe
38:17decision support systems.
38:19For future discovery.
38:23So some general opportunities
38:25we can redefine OSA.
38:27This is clearly low hanging fruit.
38:29The HI consists of Hypopneas and apneas.
38:33Well, what about?
38:35What about restless legs?
38:36What about arousals?
38:37What is it about the brain waves
38:40that can affect hi and you know,
38:42machine learning will help
38:43us understand this better.
38:45What about the ODI O2 nedir time
38:47auction saturation less than 88%?
38:49How does this factor into OSA?
38:52And then can we combine this with clinical
38:55symptoms and atrophy to understand
38:57what OSA is better so that we can
39:00personalize therapy for these patients?
39:02There's a very large knowledge
39:04gap in Sleep Medicine.
39:06But we have to figure out how
39:08to move forward using ML.
39:10And other opportunities with the
39:12proper definition of OSA treatment.
39:14As I mentioned earlier,
39:16Pop success is the 70%
39:17compliance is very archaic.
39:19We have to kind of reevaluate that
39:23morbidity and mortality outcomes.
39:24I think there was a jerk
39:25you earlier this year.
39:26That said, maybe you know,
39:27treatment for OSA won't help
39:29our cardiovascular patients
39:30when we know in our hearts.
39:31It's definitely can help some patients.
39:34And how do we combine this with
39:36subjective measures as well?
39:37And then the personalized
39:38medicine that we talked about,
39:40the different.
39:41Algorithms based on phenotype cluster
39:43analysis and targeting clinical outcomes.
39:49And then also you know health care
39:53disparities is always important.
39:55How do we address all the races?
39:57So, for example,
39:58Asian Americans and African Americans,
40:02they are known to develop OSA at
40:05earlier ages and at lower BMI's,
40:08but they're frequently
40:09being missed by their GP.
40:12So how do we better address these groups?
40:13And of course, women post menopausal
40:15women who can develop sleep apnea
40:17at the same rate at a higher age.
40:18These again are.
40:20Frequently going missed, and of course,
40:22how do we improve access to care?
40:24So one of our long term goals, right?
40:27ML and OSA study is to develop a website.
40:29This way we can post this algorithm so
40:32you know clinicians and oncologists
40:34in rural communities can adequately
40:37define R OSA like sleep studies.
40:39Excuse me, sleep physicians,
40:41there's a dearth of us like
40:43we need more clinicians,
40:44but that's pretty much impossible
40:45in the near future.
40:46So how do we improve access to care and
40:49help our colleagues that are defined OSA?
40:52And then finally, you know,
40:55we have to validate consumer technology,
40:57right?
40:57So we had to find a way to meet in
40:59the middle to engage our patients.
41:01And also we have to recognize that
41:03OSA does not occur in isolation.
41:05So we have to consider insomnia.
41:07Their CBT apps like how do we work with
41:10that to improve sleep apnea treatment?
41:16So AI is a comment. It can be applied
41:20throughout the essay journey.
41:21It can start with cleaning by screening.
41:23How do we improve screening based
41:25on patient related features?
41:27How do we improve our diagnostic testing?
41:29Should we create an algorithm and
41:30have it available to our colleagues?
41:32How do we redefine the definition of OSA?
41:35How do we improve therapy
41:37like tailoring therapy?
41:38Who would best be suited for
41:40the oral mandibular device?
41:41You know now there's an L
41:44for mandibular movement.
41:45At night,
41:45so this way these patients may be
41:47best fitted for an oral appliance
41:49rather than going through straight
41:51to see PAP and also for follow up,
41:53can we predict those who will not
41:56adhere to CPAP initially if we have
41:59a patient in the middle sleeping
42:01in the middle of Central Park,
42:02are they less likely to be accurate
42:04with PAT because they have chronic
42:06allergies and then finally with
42:08the longitudinal of follow up?
42:10Are there ways that we can predict risk
42:13of hospitalizations in our OSA patients?
42:15Can we create an electric system?
42:19And as mentioned earlier,
42:20you know it's important to note that
42:23there there should be a way where
42:25this ML should not replace RTGS.
42:28And could this mill help us better
42:31target what RCTs are needed in the
42:34future for Sleep Medicine? It's.
42:38But you know, we shouldn't overshoot
42:40our expectations and remember the
42:43challenges of machine learning, right?
42:45Realistically speaking,
42:46can we have a data scientist
42:49in every institution?
42:51Or maybe even better yet,
42:53on every medical team, maybe in the ICU?
42:57Second thing is you know the academic
43:00partnership with industry, right?
43:01How do we you know industry is way
43:04ahead of us in terms of data analytics?
43:06Is there a simple way to do identify
43:09and hash the data so that we can start
43:12collaborating with external companies?
43:14Third, democratization of data
43:16from all healthcare ecosystems,
43:19like we have the consumer technology.
43:21How do we get access to that?
43:22We have the HR data.
43:23How do we get access that
43:24we have all the app data?
43:26How do we get access to that?
43:27And then finally, you know,
43:29designing prospective studies.
43:30How do we improve sensitivity,
43:32specificity and accuracy
43:35with leveraging mill?
43:40So sorry you know here.
43:42I just want to close this out with the
43:44artificial intelligence Sleep Medicine.
43:46So GSM has noticed that you
43:49know we have to address AI,
43:52but it makes a point to highlight
43:54that the goal of AI integration
43:56should be to augment not replace
43:58expert evaluation of sleep data.
44:03So it's OK to be different.
44:06Machine learning will not replace EBM.
44:08We should use it in conjunction with it.
44:11There are multiple data
44:13sources for Mill and OSA.
44:14Quality data will require integration,
44:18transformation and clinical interpretation to
44:20create a compelling value prop for AI and ML.
44:24And then finally,
44:25you know an animal can't solve every problem.
44:28But we can start in a field such as
44:31OSA that has sufficient problems.
44:33To begin with and for us
44:36to tackle one by one.
44:38And finally, you know,
44:39I hope we can empower our patients
44:41with this more personalized
44:43approach through continuous care.
44:46And I like to close off with this.
44:49I know we're all excited for ML
44:50and ready to take it by storm,
44:52so please do read this primer article.
44:54It's by JAMA.
44:55It's how to read articles that use
44:57machine learning so we can identify
44:59that this would be good and this
45:01is considered bad garbage data.
45:05And I I definitely want
45:07to thank my dream team.
45:09It's these are wonderful software
45:11engineers and physicians that are
45:13well versed in data analytics and they
45:16really guided me through this process.
45:20So thank you all for your time.
45:22I know we have a hard stop at 3:00 PM.
45:24I have my contact information
45:26there on the screen tube you'd
45:27like to email me and in private.
45:40Thank you so much Doctor Tan,
45:42that was wonderful really.
45:43A great, great overview of machine
45:45learning and how it applies to sleep.
45:47And I agree with you.
45:49I think that sleep is really an excellent.
45:51You know it's the perfect field
45:52'cause we have so much data and there
45:54is so much that we have to learn.
45:56I'm going to moderate the chat so I want
45:58to see first if there's any questions in
46:01the chat already that I will ask you about.
46:04I have a thank you already.
46:06If people want to unmute, that's
46:08certainly fine to ask your own question.
46:11I mean, I think you brought up a lot
46:13of the great questions actually in
46:14one of your final summary slides is,
46:16you know there's so much opportunity.
46:18But how do we get around some
46:19of these issues?
46:20You know the privacy issues?
46:22How did different institutions
46:23like ours don't talk to each other
46:25and the platforms that the PAP
46:26devices use don't talk to each other
46:28nor do they talk to us easily.
46:30What's your thought?
46:31Should we be collaborating and
46:33doing this together as a team?
46:35Should each institution try to
46:36do it its own way,
46:37what how should we move forward?
46:40Oh, that's a great question,
46:41so you know why reinvent the wheel
46:43if you're going to start the wheel,
46:45I'll continue it with you, right?
46:47So, for example, what we did is.
46:51We just leverage Redcap,
46:52so we're able to leverage all of the
46:56HL seven information and then deposit
46:58it into Redcap just by hashing it.
47:00So there are ways to create a better
47:03interface to consolidate the data,
47:05so I'd say like the hard question
47:08of Epic not talking to Allscripts
47:10that's very difficult,
47:11but an easy stepping stone be
47:14identifying someone else with epic and
47:16creating a platform of almost like a
47:19registry of all the data to deposit.
47:21And then analyze and maybe sharing
47:23the cost of a data scientist.
47:25Great
47:26thank you doctor clear. I think.
47:28I think I did. Was able to unmute
47:29you if you want to ask your oh
47:31OK, so let me just OK so. One
47:36of the things that you pointed out,
47:38and it's certainly true is that
47:41consumer wearables are like taking
47:43the world by storm. And and there are all
47:47sorts of devices, the aura ring and you
47:50mentioned some of them.
47:52But a lot of the devices out there
47:55we have no idea about what they do,
47:59whether they are accurate and so forth.
48:01And there are some devices that have
48:03come out that are and more are coming
48:06out that are going to be able to.
48:08For example, on a ring be able
48:10to measure oxygen saturation,
48:12blood pressure, temperature,
48:14you name it and and and it's
48:17very important for that.
48:19The validation data to be made available.
48:23Before anybody uses it
48:25for anything and and and
48:27this is something that people
48:29need to pay attention to.
48:31Yeah, I couldn't agree with you more.
48:34Unfortunately, there is this conflict
48:35of interest by these companies that are
48:37creating these consumer technologies.
48:39They're almost like praying on
48:41the patients like little pariahs.
48:43Yeah yeah, they're just like OK.
48:44This will generate a lot of revenue
48:46and they almost don't care how.
48:48Like if it's proper if it's appropriate,
48:51because then if they did,
48:52they'll need FDA approval,
48:54which creates many obstacles more time,
48:56more money.
48:57So I'm wondering if there's a way that
48:59we can almost meet in the middle and.
49:02They want to work.
49:03Google Nest Hub did with the ASM because
49:05Google could have easily said, OK,
49:07you're having this sleep disturbance.
49:09Try drinking chamomile tea at night,
49:12but instead what it did is
49:13it tried to partner?
49:14You know there's a valid attempt.
49:16It tried to partner with the ASM
49:18to say to deliver like proper sleep
49:21hygiene based on external factors.
49:23So something as consistent something
49:25as simple as consistent sleep
49:27times which we all know will
49:29help improve insomnia and sleep.
49:31So that's something that I think
49:32that's a good stepping stone.
49:33But I couldn't agree with you more that
49:35there's no way that we can control
49:37this wrath of consumer technology.
49:43Great thank you. Other
49:44questions from the audience.
49:46If you'd like me to unmute
49:47just if you send me a message,
49:49I can do that for you or otherwise if
49:51you type your question in the chat,
49:52I'd be happy to ask it.
50:07The question.
50:14Our team is unusually quiet today.
50:16I think you answered probably
50:19everyone's questions.
50:19I think it was terrific.
50:23I'm just looking to see that there
50:25was a question about how this
50:27is different from discriminant
50:29analysis or factor analysis.
50:31How does machine learning
50:32different from that differ from
50:34that that type of analysis?
50:36Any plots there
50:37now? Sorry, I don't quite know. Yeah,
50:40OK, you did specifically say that at
50:42the beginning you give a disclaimer,
50:43so that seems like a tough one.
50:46So what's your next step
50:48in your in your project?
50:50So I think that's really interesting
50:52that that that radiation therapy
50:54you know is a potential risk factor.
50:59So you're obviously gonna publish that.
51:01Where are you going to go from
51:02here with that? Would that work?
51:05Thank you, that's a great question.
51:06So I do want to establish a site so you
51:09know right now we we use a stopping,
51:11but there are times in
51:13the stopping is imperfect.
51:14It's a simple measure,
51:15but I I think the best step is to find
51:18like designate a website and then have
51:21this algorithm available for kind of
51:23like a scoring tool that clinicians
51:25can use an input for their cancer patients.
51:28And I've spoken to a lot of oncologists
51:30and just being at Sloan Kettering
51:31and they find that sleep disturbances
51:33are highly prevalent in their group.
51:35And now they're coming to realize that maybe
51:38insomnia isn't just related to their disease.
51:40It could be,
51:41but there are other disorders
51:42that they can consider,
51:43but they just don't have that access.
51:46So something as basic as sleep apnea
51:48that they can think about,
51:50especially if their patient had radiation
51:52to the head and neck or other patient,
51:54was obese.
51:55And but they're just not quite sure
51:56they can just easily use this tool.
51:59Terrific, I think that's wonderful.
52:01Thank you. Alright.
52:04So if we don't have other
52:06questions from our group,
52:07I'm just gonna thank you for a
52:08really terrific presentation.
52:09It was really wonderful and I'm sure
52:12you'll get some other questions.
52:13As with your email,
52:14probably some private questions
52:15about how to set up some of these
52:17interfaces and how to actually do
52:19things on a on a nuts and bolts basis,
52:21but thank you so much for your time.
52:22Really appreciate it.
52:24Take care. Bye bye.
52:28Bye bye everybody. Thank you.