Best of LinkedIn: Health Tech CW 22/ 23

00:00:00: This episode is provided by Thomas Allgaier and Frennus, based on the most relevant LinkedIn posts of health tech in CW-twenty two and twenty three.

00:00:09: Frenness equips HealthTech providers with market intelligence to identify which hospitals target and how to reach decision makers for hospital digitalization.

00:00:18: as a result You can find more info in the description.

00:00:23: So, welcome to The Deep Dive everyone!

00:00:26: We are thrilled.

00:00:27: a

00:00:30: really targeted look at the most critical health tech trends that surfaced across LinkedIn over calendar weeks, twenty-two and twenty three.

00:00:38: Right we've basically combed through all the noise to bring you the insights from folks who are actually building in deploying these systems

00:00:44: Exactly!

00:00:45: And looking closely of conversations among developers and clinical officers lately there's very clear pattern emerging.

00:00:52: What is fascinating here is defining theme right now Is about connected operational execution.

00:00:58: Yeah, I'd say the industry is officially exiting that honeymoon phase.

00:01:02: You know?

00:01:02: That era of broad innovation messaging where everyone was just hyped about the potential?

00:01:08: Oh absolutely!

00:01:08: It's not about the raw capability of a shiny new technology anymore.

00:01:12: People want implementation discipline.

00:01:14: They want validated outputs and safe operating models.

00:01:18: Let's unpack this because the industry has clearly demanding real-world utility.

00:01:22: now We are seeing a massive shift from just, you know admiring a cool algorithm to actually asking how it integrates into a high-pressure clinical workflow.

00:01:31: Right which brings us to AI.

00:01:33: as an active participant in the care team Brent Strayley and Thomas Iowansiu recently shared some really deep insights into Oracle Health's Clinical AI agent.

00:01:43: Yeah And we're finally moving way past that error where AI just sat on background and acted as passive note taker.

00:01:48: Exactly!

00:01:49: The underlying shift here is basic dictation to actual semantic reasoning.

00:01:54: This system is parsing the messy, completely non-linear way doctors and patients talk...

00:01:58: Which it's incredibly hard by the

00:02:00: way!

00:02:00: It is but finds its clinical intent maps into a specific hospital protocol so actively drafting prescriptions queuing up lab orders, imaging requests all of that.

00:02:11: The operational metrics they shared are just wild.

00:02:15: At St.

00:02:15: John's Health, after deploying this agent they saw a massive seventy-four percent reduction in physician documentation time.

00:02:21: Seventy

00:02:22: four percent?

00:02:23: I mean when you look at healthcare burnout rates right now.

00:02:26: that kind of metric fundamentally alters the health system operating margin.

00:02:31: Oh

00:02:31: for sure.

00:02:31: it is like think about upgrading from passive court stenographer to proactive paralegal.

00:02:37: That s great way to put it.

00:02:38: The stenographer just gives you a massive unreadable transcript But the AI paralegal actually anticipates what you need next and has the forms fully drafted before you even ask.

00:02:48: Yeah, And Stephen Elsie actually pointed out that this Oracle clinical agent is now available in the UK too.

00:02:54: it's proven its worth across several NHS trusts.

00:02:56: so The global rollout is moving incredibly fast.

00:02:59: That's huge.

00:03:00: but AI isn't just staying on the clinical side right?

00:03:02: It's moving to patient navigation.

00:03:05: Joshua Liu highlighted this post about Novant Health.

00:03:07: They just launched Aubrey, which is powered by Coventus.

00:03:10: It's a heap of A compliant AI assistant specifically for surgical patients.

00:03:15: Okay So what does Aubrey actually do?

00:03:18: Well instead of Just answering FAQs on a website it actively texts patients with pre-op instructions fasting reminders and Day of surgery guidance.

00:03:27: see I have to play devil's advocate first.

00:03:29: second here automated patient texting makes me a bit nervous.

00:03:33: like If a patient replies saying they feel dizzy, you really cannot have a chatbot hallucinating medical advice.

00:03:39: Oh completely and that's exactly why ovaries architecture is so critical.

00:03:43: it's a human in the loop system.

00:03:44: ah yeah.

00:03:45: So the AI handles the routine logistics but it runs constant sentiment and keyword analysis on the patients replies.

00:03:52: if It flags any clinical concerns?

00:03:54: It instantly escalates the thread to a human nurse.

00:03:57: That makes a lot more sense.

00:03:58: It's complimenting the team, not replacing them.

00:04:01: but you know going back to my devil's advocate hat if you look at your average LinkedIn feed.

00:04:04: right now it feels like literally every startup is an AI pioneer.

00:04:08: Oh tell me about it.

00:04:09: Agamon Gore put out a really sharp critique of this recently.

00:04:13: He argued that a lot of these startups are basically just well-dressed API call.

00:04:18: Yes He made such a great point.

00:04:20: Right, like are they actually running their own predictive logic or... Are they just formatting large language model outputs from like GPT or Gemini and calling it proprietary?

00:04:31: And his critique hits on a fundamental misunderstanding of what an LLM does.

00:04:36: It predicts.

00:04:36: the next statistically likely word-it does not reason through a biological pathway

00:04:41: Exactly!

00:04:43: And Jan Beger shared a systematic review with Eighty One Studies that backs up this exact skepticism.

00:04:49: Right.

00:04:49: The review on LLM limitations in health care, everyone always talks about hallucinations

00:04:54: but the reviews showed those failures usually trace back to something much earlier in the process.

00:04:59: right.

00:04:59: flawed input data and poor prompts

00:05:02: exactly algorithmic bias directly from training data.

00:05:05: plus clinicians were having these highly engineered multi-query prompts.

00:05:09: just get an actionable response

00:05:11: which is not feasible at a rushed clinic.

00:05:13: this brings us that mind blowing stat you mentioned before we started recording.

00:05:17: Oh!

00:05:17: From Jayati Pandy.

00:05:18: Yeah, she's at the overjet summit and shared something that really shifts the paradigm for diagnostic AI.

00:05:24: Let's hear it!

00:05:24: She

00:05:24: noted that licensed dentists only agree on a diagnosis about twenty-two percent of time.

00:05:30: Twenty two percent?

00:05:31: That is just... I mean think about that

00:05:33: Right And consider how you train machine learning model to read an x-ray.

00:05:37: You feed it thousands of images tagged by human experts with the correct diagnosis.

00:05:42: But

00:05:42: if humans only agreed twenty-two percent at a time...

00:05:45: Exactly!

00:05:46: The AI is just learning the aggregate majority vote, of highly subjective profession.

00:05:51: If we connect this to bigger picture Consistency isn't accuracy?

00:05:55: Wow So that model could perfectly reproduce the same diagnosis every time but It might consistently be wrong.

00:06:01: call

00:06:01: Precisely Especially on borderline lesions where human labels are weakest.

00:06:06: It feels like a win for efficiency, but it's just standardizing human flaws.

00:06:10: That is a terrifying thought and it really exposes the major roadblock because if AI outputs are only as good at input data how do we free that data from siloed hospital systems to begin with?

00:06:22: Well...that IS THE MILLION DOLLAR QUESTION, ISN'T IT?

00:06:24: The operating room has actually become huge battlegrounds in this.

00:06:28: Amadra Hotcar observed that the OR is shifting from isolated smart devices.

00:06:36: Yeah, he mentioned Johnson and Johnson and Striker.

00:06:38: Right?

00:06:38: They're building these intelligent ORR networks.

00:06:41: The value is shifting from selling standalone surgical tools to owning the digital operating layer that connects everything.

00:06:48: But if you are a hospital executive That sounds like an integration nightmare.

00:06:53: Bjorn von Siemens brought up great point about this.

00:06:56: Yes

00:06:57: He warned too many medical device companies are still trapping information in closed proprietary data cages that only link their own hardware.

00:07:05: it's totally intentional.

00:07:07: Even after a decade of pushing for connected ORs, we still don't have across manufacturer view surgical performance.

00:07:13: because the silo

00:07:14: which is why I love what Javed Ali shared he just bypassed them.

00:07:17: manufacturers entirely.

00:07:19: oh his open source prototype.

00:07:20: yeah that was brilliant right.

00:07:22: He built pipeline to convert messy legacy hospital data like Old CSBs and HL-Seven feeds into standardized FHIR bundles for India's ABDM ecosystem.

00:07:32: For

00:07:32: anyone listening who isn't familiar, FHir is basically a universal grammar for healthcare data.

00:07:38: Ali treated interoperability like a hardcore engineering challenge instead of waiting for policy changes.

00:07:44: So

00:07:44: what does this all mean for health systems?

00:07:46: bottom line?

00:07:47: Like What the actual ROI of breaking open these data cages?

00:07:51: Well, Seema Verma shared a very concrete number from St.

00:07:53: Joseph's Health.

00:07:55: They saved an estimated seven point two million dollars in transfusion costs.

00:07:58: Wait really?

00:07:59: Just for data integration.

00:08:00: how does that work mechanically?

00:08:02: So they simply surfaced lab results and near real time within their electronic health record.

00:08:07: Imagine a doctor orders the blood transfusion based on a lab from two hours ago.

00:08:11: Okay

00:08:12: The blood is prepped but the patient actually stabilized twenty minutes ago.

00:08:16: If that updated data is stuck in another department, then transfusion happens anyway.

00:08:20: Ah I see!

00:08:20: But if it's in the EHR?

00:08:22: The doctor sees this update and cancels order.

00:08:24: Exactly It prevents an unnecessary expensive potentially risky procedure.

00:08:30: That one point of integration saved them millions...

00:08:33: That makes total sense.

00:08:34: And when you scale up those stakes are enormous.

00:08:37: General Broughton highlighted the ultimate interoperability challenge the VA's sixteen billion dollar EHR modernization with Oracle Health.

00:08:47: Sixteen billion, it is

00:08:49: staggering!

00:08:50: It s the largest federal health IT initiative in history.

00:08:53: They are trying to create a single longitudinal record for veterans across their entire lifetime

00:08:58: And we're seeing this internationally too.

00:09:00: Erin O'Halloran and Pedro Ferreira noted similar massive momentum in Nova Scotia and Ireland.

00:09:05: Ireland is building its first national maternity EHR.

00:09:08: It just proves that trust and coordinated care are really driving these massive shifts, which actually transitions perfectly into imaging and diagnostics.

00:09:16: Yeah

00:09:16: because this shift toward a connected digital layer Is fundamentally changing how hospitals view their heavy capital equipment.

00:09:22: Exactly Imaging is evolving from standalone hardware Into broader software platforms

00:09:28: And that changes the whole purchasing strategy.

00:09:31: Pat Tampong-Pollhorn and Daniel Carter shared some incredible updates about GE Healthcare.

00:09:36: They're upgrading eleven in twelve year old CT and MRI scanners to their latest tech.

00:09:42: Yeah, the SIGNA Victor & Apex Edition upgrades.

00:09:45: Here's where it gets really interesting.

00:09:47: It's like downloading an over-the-air software update that somehow gives a decade old car brand new engine.

00:09:53: That

00:09:53: is the perfect analogy because the core magnet in an MRI lasts for decades.

00:09:57: so GE just leaves it and upgrades its reconstruction software, adds deep learning AI around without those massive turnkey replacement costs.

00:10:06: And results speak to themselves.

00:10:08: At a Bangkok hospital, Patampong noted this upgrade reduced MRI scan times by an average of thirty-five percent.

00:10:15: Thirty five percent is massive!

00:10:16: Yeah They boosted their daily throughput without buying new machine.

00:10:20: Your capital expenditure stays low but your revenue capacity jumps.

00:10:23: Of course sometimes you do just need net new hardware right?

00:10:26: Definitely Maria Vanderville highlighted Phillips' new titanium ultra-high gradient MRI.

00:10:33: It's built for highly precise biomarker quantification, where older Magnus just don't have the

00:10:38: power.".

00:10:39: And Dr Daniel Stromers showcased a really cool clinical case using Sipman's health in ears.

00:10:44: photon counting CT... Oh!

00:10:46: The pediatric case?

00:10:47: Yeah they use it to find ureter stones and a nine year old boy The Photon Counting Tech gives this incredibly detailed anatomical imaging, but at an ultra-low radiation dose.

00:10:57: Which is so critical for kids!

00:10:59: And speaking of pediatric patients... A tool Goopta shared a really touching update about improving the patient experience.

00:11:05: Yes…the Philips and Disney partnership.

00:11:07: I love it

00:11:08: Right.

00:11:08: they partnered to bring storytelling characters directly into Pediatric MRI suites.

00:11:13: So its not just like stickers on machine

00:11:15: No..its fully integrated ambient experience Lighting sound digital animation and it resulted in a forty-three percent reduction in pediatric stress.

00:11:24: Wow!

00:11:24: Yeah, that led directly to a sixty three percent drop In scan interruptions because if the kid is distracted and calm they don't squirm

00:11:32: which means you probably Don't have to use heavy sedation as often either.

00:11:35: That is just brilliant.

00:11:37: But okay Let me pivot us here Because this brings up A massive structural question for

00:11:41: Me.

00:11:42: Hold For It.

00:11:42: We're talking about upgrading diagnostic systems on The fly And deploying self learning AI.

00:11:49: In a highly regulated industry where lives are literally on the line, how do you regulate software that constantly changes?

00:11:56: That is the defining puzzle right now.

00:11:58: Asim Khan provided some much-needed clarity on this looking at the FDA's framework for AI and patient monitoring

00:12:05: because the old system wasn't built for AI

00:12:07: exactly.

00:12:07: The old five ten K process was for static devices.

00:12:11: You test a scalpel it doesn't change but AI as dynamic.

00:12:14: so the FDA created the predetermined Change Control Plan or PCCP.

00:12:18: The PCCP?

00:12:19: How does that work in practice?

00:12:21: Khan described the sweet spot as a hybrid model, locked plus governed retraining.

00:12:25: basically manufacturers agree up front with EFTA on boundaries for future updates

00:12:30: so they set guardrails early

00:12:31: right and pre-approved guardrails like retraining the model to fix data drift, they don't need a brand new marketing submission every time.

00:12:41: That

00:12:41: is so smart.

00:12:42: it lets the software evolve safely and we really need those guardrailes because the stakes of a biased AI are just devastating.

00:12:48: Absolutely.

00:12:48: Trina D posted a very sobering analysis about this regarding AI in pharmacovigilance.

00:12:54: Right

00:12:54: which is how we track adverse drug reactions after a drug hits the market.

00:12:58: Exactly The FDA's Fayers database has over thirty one million adverse event reports.

00:13:05: Human teams can't triage that volume manually, so AI is mandatory.

00:13:10: But Trina pointed out a massive flaw in the data — women experience adverse drug reactions at twice the rate of men... Right!

00:13:16: ...but if the AI is trained on older male-dominant reporting data it's going to have systemic blind spots.

00:13:22: It could literally cost lives because the AI isn't sensitive to signals affecting half the population.

00:13:28: And this raises an important question for developers Are you auditing your models for diversity?

00:13:34: Because this reality that digital tools have life or death consequences ties into what Isaac E. shared about Germany's Digi two point.

00:13:42: oh program,

00:13:43: right?

00:13:43: Digital health is losing its protected status as just a neat innovation category

00:13:48: exactly.

00:13:49: Health systems are treating it has hardcore healthcare infrastructure.

00:13:52: now.

00:13:52: Reimbursement doesn't guarantee adoption anymore.

00:13:54: No, the tools have to actually integrate operationally and reduce workforce pressure if they want to survive.

00:14:00: They can't just be a shiny app on a tablet any more

00:14:03: Spot-on!

00:14:04: The bar has officially been raised.

00:14:05: It really is Well.

00:14:07: If you enjoyed this episode new episodes drop every two weeks.

00:14:10: Also check out our other editions on ICT & Tech Insights DefenseTech Cloud Digital Products & Services Artificial Intelligence And Sustainability in Green ICT.

00:14:19: Yeah, thank you so much for joining us on this deep dive.

00:14:22: Don't forget to subscribe So you never miss an update.

00:14:25: absolutely.

00:14:26: Do have a final thought everyone before we sign off?

00:14:28: I do Think about what we covered today.

00:14:31: if AI consistency merely scales human inconsistency like We saw with that twenty two percent dentist agreement rate and If digital health is now judged as critical infrastructure Yeah.

00:14:42: Will the next decade of health tech dominance be decided not by who writes best algorithm, but by who controls the most unbiased diverse ground truth data?

00:14:52: Something for you to explore on your own!

00:14:54: That is a fascinating question.

00:14:55: end-on thanks for listening everyone.

00:14:56: we'll catch ya.