Latest / Becker’s Healthcare Podcast / Scaling AI in Healthcare Through Better Data, Workflow and Trust with Nirav Shah

Transcript

Scaling AI in Healthcare Through Better Data, Workflow and Trust with Nirav Shah · Becker’s Healthcare Podcast

• 0:00
  - What we wanna do with AI is help
• 0:01
  - clinicians go from being reactive
• 0:04
  - to proactive.
• 0:05
  - At Philips, we can recognize patterns in monitoring
• 0:08
  - data across populations
• 0:10
  - and apply those insights to specific patients.
• 0:14
  - Safe and responsible AI can support sharper insights,
• 0:18
  - pave the way for faster diagnosis,
• 0:21
  - and help streamline care.
• 0:23
  - That can benefit clinicians, it can benefit health
• 0:26
  - systems, and most importantly, it can benefit patients
• 0:29
  - and their families.
• 0:33
  - Phillips.
• 0:34
  - This is Gracelyn Keller with the Becker's Healthcare
• 0:36
  - podcast.
• 0:37
  - Thank you so much for tuning in to
• 0:39
  - the podcast series.
• 0:41
  - Joining me for today's discussion is Nirav Shah,
• 0:44
  - who is the associate chief medical informatics officer
• 0:47
  - AI and innovation at Endeavor Health. So, Nirav,
• 0:50
  - thanks for being here, and let's jump right
• 0:52
  - into our conversation.
• 0:54
  - Health systems are receiving a growing stream of
• 0:56
  - patient data from bedside monitoring, ambulatory devices, and
• 1:00
  - care delivered at home. Where is that data
• 1:02
  - creating actionable insights for clinicians, and where does
• 1:06
  - it still create more work than insights?
• 1:09
  - Well, thank you for having me. Really appreciate,
• 1:12
  - highlighting the work that we're doing.
• 1:14
  - So the toll of this data stream is
• 1:16
  - incredibly real. And as an infectious disease physician,
• 1:19
  - I see it firsthand, and I see it
• 1:20
  - with my colleagues.
• 1:21
  - On one side, we have closed loop continuous
• 1:24
  - monitoring with clear clinical protocols.
• 1:27
  - For example, continuous glucose monitoring for our PCPs
• 1:30
  - and for our endocrinologists,
• 1:32
  - or thinking about remote heart failure telemetry,
• 1:35
  - for our cardiologists.
• 1:36
  - And these are highly actionable because the data
• 1:39
  - feeds directly into structured dashboards and trigger a
• 1:42
  - clear clinical action.
• 1:44
  - Where it becomes a massive burden is when
• 1:46
  - we're hit with uncontextualized
• 1:48
  - consumer grade data dumps. If a patient hands
• 1:51
  - me three months of Apple Watch data,
• 1:54
  - I don't really know what to do with
• 1:55
  - it. It creates massive cognitive,
• 1:58
  - fatigue, and it transfers all of the analytical
• 2:01
  - work and the sorting and the data cleaning
• 2:03
  - and all of that onto the physician, turning
• 2:05
  - us into human data cleansers and analysis machines.
• 2:09
  - The key takeaway is that data without a
• 2:11
  - workflow
• 2:12
  - is just a distraction.
• 2:14
  - We don't need more raw data points. We
• 2:16
  - need AI to act as a signal to
• 2:18
  - noise filter that surfaces only the true anomalies
• 2:22
  - requiring human clinical judgment because that's where there'll
• 2:25
  - be a change to the workflow and where
• 2:26
  - we'll be able to make sense of that
• 2:28
  - data.
• 2:29
  - When clinicians are trying to reach a diagnosis,
• 2:32
  - they often must piece together information from radiology,
• 2:35
  - cardiology, pathology, other systems.
• 2:38
  - And so where does that fragment create the
• 2:40
  - most friction for your organization, and how do
• 2:43
  - you bring together data sources and address that,
• 2:46
  - or close the gaps?
• 2:47
  - A very complex problem and something I faced
• 2:50
  - this past week when I was on clinical
• 2:52
  - service. So, infectious disease
• 2:54
  - is clinical detective work.
• 2:56
  - To treat a patient, I have to read
• 2:58
  - radiology reports, look at pathology, check microbiology cultures,
• 3:02
  - review surgical
• 3:03
  - notes, look at the electronic health record for
• 3:06
  - just notes in general. It's in multiple
• 3:07
  - disjointed systems.
• 3:09
  - So the friction isn't that the data doesn't
• 3:11
  - exist.
• 3:12
  - It's the steep cognitive burden of constant context
• 3:15
  - switching between different software interfaces or different tabs
• 3:19
  - within the same software
• 3:21
  - just to build a cohesive mental time line
• 3:24
  - of the patient.
• 3:25
  - But you cannot fix this fragmentation
• 3:28
  - with another shiny bolt on AI app. You
• 3:31
  - fix it by doing the hard, unglamorous work
• 3:33
  - of transforming your foundation.
• 3:35
  - At Endeavor Health, we made the massive decision
• 3:38
  - to unify our EHR into a single instance
• 3:41
  - and streamline our semantic data layer. So once
• 3:45
  - you have that unified,
• 3:46
  - clean data foundation, you can point advanced tools
• 3:49
  - like enterprise LLMs,
• 3:51
  - the data to generate a cohesive patient narrative
• 3:53
  - rather than forcing the physician to act as
• 3:55
  - that human data aggregator.
• 3:58
  - And patients increasingly move among the hospital outpatient
• 4:01
  - settings and home.
• 4:03
  - So where does clinical context most often get
• 4:06
  - lost during those transitions, and what would it
• 4:08
  - take to give care teams a more continuous
• 4:10
  - view of the patient?
• 4:12
  - This is a great question.
• 4:14
  - Clinical
• 4:16
  - context is most often in the why.
• 4:18
  - That's where it's lost. When a patient moves
• 4:20
  - from the hospital to an outpatient clinic or
• 4:22
  - back home, the receiving team gets a medication
• 4:25
  - list or a standardized discharge PDF,
• 4:28
  - but what gets completely lost is the clinical
• 4:30
  - nuance.
• 4:31
  - Why did we choose to stop a certain
• 4:33
  - antibiotic? What were the family's hesitations
• 4:35
  - or this decision or that decision? What were
• 4:38
  - we actively watching out for?
• 4:40
  - To solve this, we have to move past
• 4:42
  - static discharge summaries and look toward dynamic agentic
• 4:46
  - workflows. We need AI agents that don't just
• 4:48
  - pass along a flat document,
• 4:50
  - but actively track the patient's care trajectory
• 4:53
  - and highlight unresolved
• 4:54
  - clinical questions or things that were under discussion
• 4:58
  - for the next team. Giving care teams a
• 5:00
  - continuous view means ensuring
• 5:03
  - the narrative threat of the patient's care isn't
• 5:05
  - severed the moment they roll through a new
• 5:06
  - set of physical or virtual doors.
• 5:09
  - And with so many AI use cases competing
• 5:12
  - for attention, how do you distinguish a competing
• 5:14
  - pilot from something ready to become part of
• 5:16
  - everyday care, and what must be in place
• 5:19
  - to scale it responsibly?
• 5:21
  - To cut through the noise of the AI
• 5:23
  - hype cycle, we can't rely on gut feelings
• 5:25
  - or vendor promises.
• 5:27
  - In our system, which has over a 100
• 5:29
  - active AI tools, we evaluate the readiness of
• 5:32
  - any pilot through a con... Conceptual compass or
• 5:35
  - framework
• 5:36
  - built around three fundamental
• 5:38
  - features or lenses.
• 5:39
  - First, we look at model performance.
• 5:41
  - Does the model actually perform reliably in our
• 5:44
  - specific clinical environment or administrative environment, delivering a
• 5:47
  - high quality signal without drowning our teams in
• 5:50
  - alert fatigue, false positives, or hallucinations?
• 5:53
  - Second, we look at technical integration and workflow.
• 5:56
  - Can this tool integrate into our existing technical
• 5:59
  - stack, and does it slide naturally into the
• 6:01
  - daily workflow of our end users without requiring
• 6:05
  - disruptive bespoke workarounds?
• 6:07
  - Third, we look at value. Is it solving
• 6:09
  - a valuable problem that needs to be solved?
• 6:11
  - I I think of this as potential value.
• 6:14
  - And then we need to understand,
• 6:16
  - do we realize that value? We must see
• 6:18
  - a clear, measurable reduction in cognitive burden, a
• 6:21
  - drop in administrative costs for a tangible clinical
• 6:24
  - outcome. I like to think of it in
• 6:25
  - terms of the quintuple aim. So our guiding
• 6:27
  - philosophy is that we pilot to learn, but
• 6:29
  - we only scale at the speed of trust,
• 6:31
  - and that trust is earned in solving real
• 6:33
  - problems, accurate models, seamless workflows, and value creation.
• 6:37
  - If a pilot doesn't meet these criteria, we
• 6:39
  - optimize, pause it, or kill it.
• 6:42
  - And as technology relationships move beyond onetime purchases,
• 6:46
  - what should a partner remain accountable for after
• 6:48
  - go live to make the relationship generally value...
• 6:51
  - Genuinely valuable over time?
• 6:53
  - This is a great question, and it requires
• 6:55
  - a fundamental rethink about new technology solutions and
• 6:59
  - existing ones. So the days of a technology
• 7:02
  - vendor dropping software on a health system, cutting
• 7:04
  - a ribbon, and walking away are completely over.
• 7:08
  - AI is a living organism.
• 7:10
  - Models drift continuously
• 7:12
  - as clinical population shifts, and models and workflows
• 7:15
  - constantly evolve.
• 7:17
  - Post go live, a partner must remain accountable
• 7:19
  - for continuous model validation
• 7:22
  - and workflow adoption.
• 7:23
  - If a model's accuracy degrades over time, or
• 7:26
  - if our clinical or administrative teams start building
• 7:28
  - manual workarounds to bypass the tool, the vendor
• 7:31
  - must be in the trenches with us fixing
• 7:33
  - it.
• 7:34
  - If we share concrete workflow issues or performance
• 7:37
  - degradation
• 7:38
  - and we see zero road map velocity or
• 7:40
  - improvement
• 7:41
  - within a fixed amount of time, typically a
• 7:44
  - few months, which is the amount of time
• 7:45
  - it takes to validate,
• 7:47
  - a model, we pivot.
• 7:48
  - Health systems shouldn't fund a vendor's r and
• 7:51
  - d on our clinical time,
• 7:52
  - on our patients' clinical time. The relationship is
• 7:55
  - only valuable if the tool continuously earns its
• 7:58
  - place in our toolkit and in our workflow.
• 8:02
  - And finally, as care increasingly moves beyond hospital
• 8:05
  - walls, how are you using connected technologies and
• 8:08
  - AI to identify patients who need earlier follow
• 8:10
  - ups, support care coordination teams,
• 8:13
  - and close gaps in post discharge care?
• 8:16
  - So, again, this gets at the concept of
• 8:18
  - an agent that is able to constantly monitor
• 8:22
  - and potentially take autonomous decisions, obviously, with very
• 8:25
  - close human in the loop,
• 8:27
  - validation and assessment.
• 8:29
  - And so
• 8:30
  - these things are already part of our vendor
• 8:33
  - platforms road map, and these are things that
• 8:35
  - I think are going to help us with
• 8:37
  - that care coordination moving between physical walls and
• 8:40
  - virtual walls to help us really identify those
• 8:43
  - patients in advance who need that support and
• 8:46
  - care.
• 8:47
  - Traditionally, we use a lot of people who
• 8:50
  - triage the highest risk patients. But with autonomous
• 8:53
  - agents, you can imagine a time of abundance
• 8:55
  - where you're able to do this on every
• 8:57
  - patient and to monitor every patient to understand
• 8:59
  - when someone needs something continuously
• 9:02
  - so that you can provide the best care
• 9:04
  - possible and catch things before they become issues.
• 9:07
  - Well, thank you so much for joining me
• 9:09
  - today on the Becker's Healthcare podcast
• 9:12
  - and sharing your thoughts and insights. And we
• 9:14
  - also wanna thank our podcast sponsor, Philips. You
• 9:17
  - can tune in to more podcasts from the
• 9:19
  - Becker's Healthcare
• 9:20
  - podcast by visiting our page at beckershospitalreview.com.
• 9:24
  - Thank you so much.