Latest / Becker’s Healthcare Digital Health + Health IT / Dr. Nigam Shah, Chief Data Scientist at Stanford Health Care

Transcript

Dr. Nigam Shah, Chief Data Scientist at Stanford Health Care · Becker’s Healthcare Digital Health + Health IT

• 0:00
  Hello, everyone. This is Jacob Emerson with the
• 0:02
  Becker's Health IT podcast.
• 0:04
  Thrilled today to be joined by doctor Nigam
• 0:07
  Shah, who is the chief data scientist at
• 0:09
  Stanford Health Care. Doctor Shah, thank you so
• 0:12
  much for taking the time to be with
• 0:13
  me on the podcast today.
• 0:15
  Oh, great to
• 0:16
  be here. Yeah. And before we dive into
• 0:18
  everything we wanna talk with you about, Nigam,
• 0:20
  can you tell us a little bit more
• 0:21
  about yourself, your background in health care, and
• 0:24
  what it is that you do today at
• 0:25
  Stanford?
• 0:27
  Sure. Sure. So I often joke I've never
• 0:29
  had a real job. I got my PhD
• 0:32
  from Penn State, came to Stanford as a
• 0:34
  postdoc in 02/2005,
• 0:35
  and never left.
• 0:37
  Been on the faculty since 02/2011,
• 0:41
  tenured in 2015.
• 0:43
  And then I was doing the usual, you
• 0:44
  know, writing papers, real world evidence, machine learning
• 0:47
  on EHRs and so on. And, got sucked
• 0:50
  into working on things during the pandemic in
• 0:52
  the health system
• 0:54
  and found it very rewarding to do things
• 0:56
  in an operational environment. So in 2022,
• 1:00
  we set up this new role of chief
• 1:02
  data scientist for the health care system where
• 1:04
  the job basically is to translate things
• 1:07
  into practice such that patient care directly benefits.
• 1:13
  Fantastic. So, Nigam, we're here to talk with
• 1:15
  you today about an announcement that Stanford HealthCare
• 1:18
  made earlier this year saying that the system
• 1:21
  was going to be launching an Ambient AI
• 1:23
  pilot with Atropos Health to integrate real world
• 1:26
  evidence within the electronic health record
• 1:29
  via Ambient AI listening,
• 1:31
  from from Microsoft decks. So talk to us
• 1:34
  a little bit about this.
• 1:36
  How is this gonna be working, this pilot,
• 1:39
  and and how is this gonna help your
• 1:40
  physicians
• 1:41
  access real world evidence to make better decisions
• 1:44
  for your patients?
• 1:45
  Mhmm.
• 1:46
  So real world evidence has been part of
• 1:48
  my work
• 1:50
  almost since 2011,
• 1:51
  2012.
• 1:53
  And as you probably know, I mean, most
• 1:55
  of the times, real world evidence is an
• 1:57
  observational study that takes anywhere from six to
• 2:00
  nine months or maybe one year. It gets
• 2:02
  published in a journal like JAMA or New
• 2:04
  England Journal or Lancet or what have you.
• 2:06
  And maybe a physician reads it or doesn't
• 2:08
  read it. And the chance that it affects
• 2:11
  patient care is
• 2:13
  very low.
• 2:15
  There's some very shocking numbers out there saying
• 2:17
  where physicians tracked the number of decisions they
• 2:20
  make, they go back and ask how often
• 2:23
  was some evidence available
• 2:26
  particularly relevant to that case at hand. And
• 2:28
  often the answer is, you know, under five
• 2:30
  percent, under three percent, things like that.
• 2:33
  So that's what I've been researching.
• 2:35
  And we did a project here called the
• 2:38
  green button project that
• 2:40
  took that nine month time frame down and
• 2:43
  got it to under two days. So at
• 2:45
  the bedside, a physician can make a request
• 2:48
  and a team produces a written report,
• 2:51
  in two days.
• 2:52
  Once that was done,
• 2:54
  we spun out Atropos Health,
• 2:56
  which took those forty eight hours, turned around,
• 2:58
  and got it down to under three to
• 3:00
  four hours.
• 3:02
  So at the bedside, you have a question,
• 3:05
  you ask something, what happened to similar patients,
• 3:07
  and you get a written report before the
• 3:09
  end of the day.
• 3:11
  So then Ambient AI comes around,
• 3:15
  and number of physicians
• 3:17
  are super excited about how it's now saving
• 3:19
  them time and freeing up their sort of
• 3:22
  cognitive
• 3:22
  bandwidth to think more about the patient.
• 3:25
  On the other hand,
• 3:27
  Atropos adoption is going fine. About a thousand
• 3:30
  or 1,200 consultations
• 3:32
  are requested and reports being produced within within
• 3:35
  a day.
• 3:36
  Now separately,
• 3:38
  not at Stanford,
• 3:39
  Atropos had been developing this solution, which they
• 3:42
  call CHAT RWD,
• 3:44
  which essentially automates this
• 3:47
  procedure of producing
• 3:49
  observational studies on demand,
• 3:51
  takes the human mostly out of the loop,
• 3:54
  and can get an answer,
• 3:56
  for a lot of cases, in under five
• 3:58
  minutes.
• 3:59
  So now you look at these two things
• 4:01
  together and you ask, alright.
• 4:03
  We have a patient visit that's fifteen minutes
• 4:06
  now getting transcribed, and in about a minute
• 4:08
  or two, you get the full transcript.
• 4:10
  We have this separate ability
• 4:12
  that Stanford has already purchased
• 4:14
  where the physician, instead of a day, can
• 4:17
  get an answer in five minutes.
• 4:19
  What would happen if we put the two
• 4:21
  together?
• 4:22
  So that led to the current project you're
• 4:23
  talking about.
• 4:25
  Wow. That's that's some incredible numbers you decided
• 4:28
  there in terms of,
• 4:29
  five minutes and getting an answer back. That's
• 4:32
  incredible.
• 4:34
  I I wonder, have you heard from clinicians
• 4:37
  across the Stanford Enterprise in terms of how
• 4:40
  this is being received, what their thoughts are?
• 4:42
  Is this, being received enthusiastically?
• 4:45
  So lots of enthusiasm
• 4:47
  to pilot it and be, you know, the
• 4:50
  first to get access to it, 100%.
• 4:53
  The Atropos service, the one that produces this
• 4:57
  report
• 4:57
  has been live on our campus since almost
• 5:00
  2022
• 5:01
  actually or 2021.
• 5:03
  Okay.
• 5:05
  And that typically is used in the inpatient
• 5:08
  setting because there you can tolerate a one
• 5:10
  day or a four hour turnaround. Right?
• 5:13
  Ambient scribes are used in the outpatient setting.
• 5:16
  And so this will be the first time
• 5:17
  that real world evidence will be offered in
• 5:19
  the outpatient setting on demand
• 5:22
  at the point of care. So everybody's super
• 5:24
  excited about it.
• 5:26
  Wow. No. It's incredible. And is it my
• 5:28
  understanding that this is going to be continued
• 5:30
  to be rolled out this year? Is that
• 5:32
  the general timeline for this pilot?
• 5:35
  Absa absolutely. Definitely before the 2025
• 5:38
  ends. You know, most likely, the first actual
• 5:41
  user will get it in the next sixty
• 5:43
  days or ninety days. Wow. And and this
• 5:46
  is all your physicians across Stanford?
• 5:48
  No. We'll be actually piloting this with what
• 5:51
  we call our, community practice providers or UMP,
• 5:55
  university medical partners, or sometimes called Stanford medical
• 5:58
  partners. Sure. So these are physicians who are,
• 6:02
  like, not at the university. They're in the
• 6:05
  community practices all over the Bay Area. And
• 6:08
  they're the ones who are super excited that
• 6:10
  finally,
• 6:11
  technology is at a at a stage at
• 6:13
  ease of use that they can incorporate into
• 6:16
  their care. Wow. That's amazing. And and I
• 6:18
  know we did briefly just touch on this,
• 6:20
  but if from the patient's perspective,
• 6:23
  how would how would this,
• 6:25
  on the ground? What does this look like
• 6:26
  to me? I come into to the office
• 6:29
  and I receive those answers within five minutes,
• 6:31
  or can you can you detail that for
• 6:33
  us a little bit? Mhmm. So imagine before
• 6:35
  all of this technology, you go to a
• 6:37
  physician.
• 6:38
  You're speaking. They're looking at a computer screen,
• 6:40
  typing away furiously so as to not to
• 6:42
  miss anything.
• 6:44
  They maybe glance at you once or twice.
• 6:46
  And then, you know, once their note is
• 6:48
  done, they look at you, tell you three
• 6:50
  things you should be doing, and off you
• 6:51
  go. Right?
• 6:53
  Ambient scribes come around, and now the doctor's
• 6:56
  actually looking at you.
• 6:58
  And
• 6:59
  there's this thing in the ether that's transcribing,
• 7:03
  and by the time the the, the conversation
• 7:06
  ends in about a minute, a note shows
• 7:08
  up, which they can turn the screen around
• 7:10
  and and show it to the patient directly.
• 7:12
  In fact, our CMIO, doctor Topher Sharp, has
• 7:15
  done even video interviews with reporters showing them,
• 7:18
  you know, how powerful this is
• 7:20
  in terms of
• 7:22
  the physician being able to pay attention to
• 7:24
  you. Right? So in normal encounter, everything is
• 7:27
  fine. This automation describing
• 7:29
  sort of does the job. But then if
• 7:31
  there's a question,
• 7:33
  what does a physician do today?
• 7:35
  They will go to UpToDate or some other
• 7:38
  literature summary.
• 7:39
  Or
• 7:41
  if they had a question that can't be
• 7:43
  answered by literature, they will trigger this real
• 7:46
  world evidence report from from Atropos.
• 7:48
  But that's a manual step, both of them.
• 7:51
  So now you can imagine a situation where
• 7:53
  we fully have the transcript and it's electronic.
• 7:57
  Can I distill a question out of it,
• 8:00
  send it to literature evidence,
• 8:03
  send it to Atropos for data driven evidence,
• 8:06
  produce both of them back if something exists
• 8:08
  and is relevant, and put that as two
• 8:10
  links in the note that is being shown
• 8:12
  to the physician? And then they can choose
• 8:14
  to say, alright. Yeah. This looks good enough.
• 8:16
  I should click on it and read it
• 8:17
  and talk to the patient about it.
• 8:20
  Understood. No. This is incredible. And I it
• 8:23
  it really reflects what we're hearing from a
• 8:24
  lot of other health systems around the country,
• 8:27
  piloting and introducing this type of technology as
• 8:30
  well.
• 8:31
  So
• 8:32
  in that vein, Nigam, what what would you
• 8:34
  say have been some of the key challenges
• 8:36
  that systems,
• 8:37
  maybe Stanford alone, but from other systems as
• 8:40
  well that you're hearing,
• 8:41
  challenges they're facing when trying to integrate ambient
• 8:44
  AI tech into their existing workflows?
• 8:47
  Yeah. So there's sort of two broad buckets
• 8:49
  of challenges.
• 8:50
  The first one are sort of the the
• 8:53
  the seemingly daunting ones where or the easy
• 8:55
  ones, I would call them technical challenges. You
• 8:58
  know, what exact API are you gonna call?
• 9:00
  What's your latency?
• 9:01
  How fast can you get the answer? How
• 9:03
  will you maintain security, privacy, compliance? You know,
• 9:07
  all of that,
• 9:08
  excuse the word, junk that goes in with
• 9:11
  health IT in general in in in doing
• 9:13
  things with duct tape and bubble gum and
• 9:15
  string. Right?
• 9:17
  I'm sure.
• 9:19
  And then the second bucket are the challenges
• 9:21
  we haven't yet thought of, which is how
• 9:24
  are we gonna evaluate these things? Because, you
• 9:26
  know, given a Google query, for example, we
• 9:28
  always get an answer. Given an up to
• 9:30
  date search, you get an answer.
• 9:32
  Given an RW real world data search, you'll
• 9:35
  get an answer. And there's some safeguards built
• 9:36
  in that if the cord's too small or
• 9:39
  the phenotypes are not clearly defined, the system
• 9:41
  will refrain from answering.
• 9:43
  But when you put all this together,
• 9:46
  how do you evaluate?
• 9:48
  How do you know it's adding value?
• 9:51
  Those are two different kinds of evaluations. Those
• 9:54
  are the things I think in the next
• 9:56
  year, everybody is gonna be, thinking about. Yeah.
• 9:59
  No. Absolutely. It's a great point. And I
• 10:01
  that leads me to my next question for
• 10:02
  you is, how do you think that Ambient
• 10:05
  AI tech within the hospital, how's that gonna
• 10:07
  continue to evolve over the next few years?
• 10:10
  And and what's the broader impact that you
• 10:12
  foresee this having on on the industry?
• 10:16
  So, I mean, again, I'm speculating just like
• 10:18
  I'm guessing like everybody else. Sure.
• 10:21
  Ambient
• 10:22
  AI entered the scene
• 10:24
  as an automation,
• 10:27
  as in something that a human was doing
• 10:30
  typing in a suboptimal
• 10:31
  manner, or you might have had a human
• 10:33
  Skype present.
• 10:34
  And that thing is now
• 10:36
  automated,
• 10:38
  which
• 10:39
  immediate return and ROI is the physician now
• 10:42
  has the attention to spend on the patient.
• 10:44
  So that's great check.
• 10:46
  But now we're collecting a new kind of
• 10:49
  data
• 10:50
  in electronic form that we never had before,
• 10:54
  the real time conversation
• 10:56
  of what is happening.
• 10:58
  Now many things can come from that.
• 11:01
  This kind of
• 11:02
  evidence generation on demand is sort of near
• 11:06
  and dear to my heart,
• 11:07
  but you can imagine other things
• 11:11
  like documentation for billing could be completely automated.
• 11:15
  This whole idea that a biller and a
• 11:17
  coder has to now review everything and then
• 11:19
  send a query back to the physician saying,
• 11:22
  hey, doc. Do you remember? Was this the
• 11:24
  case with the patient or not? That whole
• 11:26
  thing can go away.
• 11:29
  We might be able to produce
• 11:32
  summaries
• 11:33
  that are directed to the patient. So the
• 11:35
  same note can be pushed through another kind
• 11:38
  of summarizer,
• 11:40
  AI that translate it to whatever language you
• 11:43
  need and outlines the three things the the
• 11:46
  patient needs to watch out for. So
• 11:48
  I think it's gonna enable a whole new
• 11:51
  class of care activities that we did not
• 11:53
  think were even possible.
• 11:56
  Absolutely. So it sounds like I mean, you're
• 11:58
  predicting not even just, an improvement in patient
• 12:00
  and and quality outcomes, but really massive repercussions
• 12:03
  for the compliance and reimbursement perspective within the
• 12:07
  health systems as well. So it's absolutely fascinating
• 12:09
  to think about.
• 12:11
  And even patient care. So, like, imagine if,
• 12:13
  you know, you went to you went to
• 12:14
  a physician
• 12:15
  and they said, you know, watch out for
• 12:18
  these symptoms
• 12:20
  once every week.
• 12:22
  If we recorded that conversation, we can trigger
• 12:25
  an agent that checks in with you by
• 12:28
  text, phone, or email or calling you saying,
• 12:31
  hey, Jacob. Are you having this in the
• 12:33
  past seven days?
• 12:35
  We're never gonna do that manually.
• 12:37
  Sure.
• 12:39
  No. It's it's really exciting.
• 12:41
  And and certainly,
• 12:43
  I know here in in the Chicago area,
• 12:44
  we've we've been seeing health systems, personally as
• 12:47
  well, rolling out this technology.
• 12:50
  And and it's, very clearly from the physician
• 12:52
  perspective
• 12:53
  already and in the data as well, reducing
• 12:56
  that administrative burnout that that so many of
• 12:58
  them face.
• 13:00
  So really exciting things on the horizon
• 13:02
  at Stanford as well in this atmosphere.
• 13:06
  But before we go, Nigam, what else are
• 13:08
  we missing? You know, you've got the ears
• 13:09
  of a lot of other CIOs and and
• 13:12
  health IT executives in health systems from around
• 13:14
  the country. What else do you wanna share
• 13:16
  with them about this this new pilot with
• 13:18
  Atropos or or about, this technology that we've
• 13:22
  discussed in general?
• 13:24
  So I would like to sort of plant
• 13:26
  this idea that one of our faculty members
• 13:28
  here, Eric Brynjolfsson,
• 13:30
  calls the Turing trap.
• 13:32
  I'm sure every all the listeners, the everyone's
• 13:34
  part of the Turing test where a computer
• 13:37
  passes off as being a human.
• 13:39
  The tutoring trap is the opposite of that,
• 13:41
  where we as humans
• 13:44
  only imagine doing with the computer what we
• 13:46
  already know how to do.
• 13:49
  So that's why we think about compliance and
• 13:51
  billing.
• 13:52
  But we don't follow-up with our patients
• 13:56
  routinely as much as we do. We just
• 13:57
  don't have the time.
• 13:59
  But what if technology enables that?
• 14:02
  And so the way I would encourage people
• 14:03
  to think of it is that, you know,
• 14:05
  AI writ large, or I like to think
• 14:07
  of it more as a data plus algorithm
• 14:09
  configuration,
• 14:11
  it started out doing simple things, you know,
• 14:13
  giving alerts and nudges and classifications
• 14:16
  and predictions.
• 14:17
  And then it matured into giving recommendations,
• 14:21
  giving observational studies and bedside evidence.
• 14:25
  Now it's getting
• 14:27
  further advanced into actively listening and summarizing and
• 14:31
  translating.
• 14:33
  But what's next? What other actions can it
• 14:36
  take? Can it do follow-up? Can it check-in
• 14:38
  on your meds? Can it check-in on your,
• 14:41
  glucose levels? Can the conversation
• 14:44
  trigger an agent that checks in on your
• 14:47
  CGM monitor and its values automatically.
• 14:51
  So I think what we need to be
• 14:52
  thinking about is what are the actions
• 14:55
  that we should be taking but are not
• 14:57
  yet taking
• 14:59
  that can be done for cheap and fast
• 15:02
  with AI.
• 15:05
  Fantastic.
• 15:06
  Great parting words for our listeners. So, doctor
• 15:09
  Shah, I wanna thank you for taking the
• 15:11
  time to sit down with us and for
• 15:13
  sharing about all the impactful work going on
• 15:15
  under your leadership at Stanford HealthCare. We really
• 15:18
  appreciate it. Thank you for having me.
• 15:20
  And to our listeners, if you'd like to
• 15:22
  listen to more podcasts from Becker's Health Care,
• 15:25
  you can visit beckershospitalreview.com.