Best of LinkedIn: Health Tech CW 26/ 27

00:00:00: This episode is provided by Thomas Allgeier and Frennis, based on the most relevant LinkedIn posts of health tech in CW-twenty six and twenty seven.

00:00:08: Frenis equips HealthTech providers with The Market Intelligence to identify which hospitals target and how to reach decision makers for hospital digitalization as a result.

00:00:22: So welcome everyone to today's deep dive.

00:00:24: We got a lot of ground to cover.

00:00:26: Yeah, we are taking a huge stack of curated insights from across the health tech space on LinkedIn and well We're looking at the operational realities in the hospital for right now.

00:00:35: Right because we really want to move past Theoretical hype you know like what happens when artificial intelligence?

00:00:42: And digital Health start actually colliding with methi.

00:00:45: clinical workflows

00:00:46: and legal liabilities yes

00:00:48: exactly legal liabilities, and ultimately actual patient care.

00:00:52: Because the industry has officially entered its proof-it phase.

00:00:56: I mean we aren't sitting around asking if a language model can pass a medical exam anymore.

00:01:00: No nobody cares about that any more

00:01:02: Right!

00:01:02: The only metric that matters now is whether these tools actually function.

00:01:06: for a clinician who's just you know trying to survive a chaotic Tuesday shift

00:01:11: That is the perfect way to frame it And the data emerging from the field right now paints a picture of two real extremes.

00:01:18: Take The Insights shared by Shespartovi and Brian Scott regarding the twenty-twenty six Phillips Future Health Index.

00:01:25: on the surface, the adoption signal is massive.

00:01:28: like

00:01:28: how massive are we talking?

00:01:30: Well, seventy one percent of clinicians report that AI has already improving their workflow efficiency

00:01:35: Wow

00:01:36: Yeah, and half of them explicitly say it's increasing their capacity to see more patients.

00:01:40: So the technology is genuinely delivering on its core promise of saving time.

00:01:45: And I know there's a but here.

00:01:46: There was a massive operational roadblock Because in that exact same report A staggering seventy percent Of those clinicians Say the training they receive for these AI tools Is either completely inadequate Or just wildly inconsistent

00:01:59: Which is terrifying if you think about It

00:02:01: really is.

00:02:01: I mean, think about what that actually means.

00:02:03: on the ground it's like tossing a surgeon-the keys to a Formula One car right?

00:02:07: Right But you're refusing to give them driving lessons.

00:02:10: The car is incredibly fast and has massive potential but the driver is just terrified of crashing

00:02:16: it.

00:02:16: And in healthcare, crashing means patient harm.

00:02:18: Exactly

00:02:19: That analogy hits the core of this issue.

00:02:22: If you don't get the clinician training to understand how the model weighs variables and arrives at conclusion You just can't expect them to confidently integrate it into high-stakes care.

00:02:34: Yeah, you can't

00:02:35: which is why Implementation science is suddenly the most critical field in health tech right now.

00:02:40: It's the whole study of how to actually get humans to use the tools effectively

00:02:44: which seems to be something the industry has largely ignored until very recently.

00:02:49: Unfortunately,

00:02:50: yes.

00:02:50: Like Jan Beger shared this highly revealing literature review on AI in healthcare supply chains.

00:02:56: they looked at thirteen qualifying studies on the topic just to see how AI was optimizing inventory and resources

00:03:03: And... The findings were a huge wake-up call.

00:03:06: out of those thirteen studies zero were deployed In A Real Functioning Hospital.

00:03:11: Wait!

00:03:12: Zero.

00:03:12: Not a single one.

00:03:13: That's

00:03:14: crazy!

00:03:14: Every single study was a simulation running on sanitized data.

00:03:19: The authors of that review basically concluded the bottleneck isn't algorithmic math, it is the data infrastructure.

00:03:26: Right because when you move from clean lab simulation to real hospital You hit all these data silos and messy.

00:03:33: legacy software Interoperability

00:03:34: walls?

00:03:35: All of them.

00:03:35: Yeah You have to pilot those systems under real conditions with real dirt before you can trust them.

00:03:41: But we are finally seeing people do that hard work, right?

00:03:44: The real world implementation.

00:03:46: We are.

00:03:46: Bilal Amitin shared results... No they didn't.

00:03:53: They embedded an LLM you know a large language model basically a massive text prediction engine directly into the electronic health record system at Penda Health

00:04:02: Right in to the EHR.

00:04:03: Right!

00:04:04: The digital filing cabinet of the hospital put AI right in the middle And this provided continuous, real-time feedback for over a hundred clinicians covering more than seventeen thousand patients.

00:04:17: The friction they encountered there proved the exact point Beeger was making —the software engineering to build the LLM—was just the baseline

00:04:25: requirement.".

00:04:26: The true hurdle is human computer interaction...

00:04:29: Like how do you design an interface that a tired, overt human will actually trust?

00:04:34: Exactly and utilize it in real time without becoming you know, another annoying pop-up on their screen that they just click away from.

00:04:41: Yeah!

00:04:42: Click fatigue is real and that friction is forcing developers to rethink how they measure success.

00:04:47: Asim Khan introduced this concept called the safety aware ROC framework

00:04:51: which is super interesting.

00:04:52: let's pause an ROC for a second because it's usually just A highly technical math term

00:04:56: right?

00:04:56: It stands for receiver operating characteristic.

00:04:59: Usually an ROc curve Is Just a graph showing How well a model balances false positives against False negatives.

00:05:05: basically it tells You how accurate The tool is.

00:05:07: Right but Khan argues that accuracy isn't the same as safety.

00:05:12: Instead of just looking at standard validation metrics like area under that curve, developers need to design for operational reality.

00:05:19: Okay so how do you do that?

00:05:20: Well, Khan says an AI must be designed for three distinct modes.

00:05:25: first there is rule and safe where you trust the AI make a positive call got it.

00:05:30: second there was rule out safe or you trusted to safely dismiss something.

00:05:34: But third mode what really matters most?

00:05:37: The gray

00:05:38: zone.

00:05:39: That's where the AI recognizes its own statistical uncertainty, right?

00:05:43: Yes It explicitly knows it's limitations and defers that decision back to human clinician.

00:05:49: So if you're building medtech now Your user experience, your monitoring dashboards, service level agreements All of this has to account for grey zones.

00:05:57: Absolutely Because If You don't engineer a graceful handoff Back To Human You Aren't Building A Safe Product.

00:06:02: And that gray zone limitation is exactly what is driving a massive structural shift in how we build these systems right now.

00:06:11: Because single AI model inherently struggles with complex real-world grey areas, the industry moving away from single models toward what people are calling agentic AI.

00:06:22: Oh, right.

00:06:23: Nishant Mishra pointed this shift out and it's a really crucial distinction because the traditional AI model analyzes data and gives you an output like summarizing a medical chart.

00:06:33: It is just one action

00:06:34: But in AI agent can understand context reason across multiple different systems And actually execute a multi-step workflow.

00:06:41: Yeah So doesn't just read the chart?

00:06:43: It reads the chart checks the pharmacy inventory and then drafts a prescription order.

00:06:48: But giving these autonomous agents the power to execute workflows creates a whole new layer of risk, particularly in the development phase.

00:06:55: Oh for sure!

00:06:56: Dr Ankit Singh shared a sobering case study from his experience building GenieCare.ai which is an oncology platform moving toward a regulated medical device classification.

00:07:06: He used Claude Code and AI agent To help build this software architecture.

00:07:11: Then he brought into completely different agent OpenAI Codex to review Claude's work.

00:07:17: And codecs immediately found three major real-world issues that Claude had completely missed, right?

00:07:23: Yes

00:07:24: a navigation bug... A database migration gap and missing regression tests.

00:07:30: If we look at the mechanics of why that happened, it reveals a fundamental flaw in relying on a single AI

00:07:36: because It was too close to its own work.

00:07:38: exactly

00:07:39: Claude missed those bugs Because Its process Was reactive And more importantly it was context blind To its own mistakes.

00:07:45: It had just built The architecture.

00:07:46: so I did Too much historical Context.

00:07:48: Codex however came In as A second agent with Fright eyes and Caught the logic gaps immediately.

00:07:53: So the takeaway Dr.

00:07:54: Singh arrived at is that having an AI grade its own homework in regulated bed tech, it's just fundamentally unsafe.

00:08:00: you cannot rely on one monolithic agent right?

00:08:03: You need multi-agent workflows

00:08:05: exactly where different specialized AI models proactively peer review each other Exactly The way human engineering teams do One agent writes the code a second agent tries to break It and A third Agent verifies the

00:08:18: fix.

00:08:19: And That concept Building an ecosystem of specialized agents rather than one monolithic system is echoing across the entire infrastructure layer of health care right now.

00:08:30: Alan Pitt highlighted a recent partnership between NVIDIA and ABRIDGE.

00:08:33: Oh

00:08:33: yeah, they're building a healthcare-specific foundation model entirely focused on clinical conversations!

00:08:38: Yeah...and Pitt referred to this as Healthcare's App Store Moment.

00:08:42: I love that phrase, app store moment because the underlying logic there is fascinating.

00:08:47: for decades hospitals have relied on these massive monolithic electronic health records

00:08:53: The big legacy systems?

00:08:54: Right!

00:08:54: Systems from companies like Epic or Cerner That try to do absolutely everything Usually at a price point.

00:09:00: most hospitals just struggle to afford.

00:09:02: But Pitt is arguing the market moving toward open composable infrastructure.

00:09:07: You build a secure foundation and then let a thousand specialized AI agents plug into it, just like apps on smartphone.

00:09:15: It sounds incredibly efficient on paper –it really does!

00:09:18: But Josh Hulu offered an necessary reality check…on this entire multi-agent ecosystem.

00:09:24: What's his take?

00:09:26: Well you can have the most advanced AI infrastructure in world capable of building & peer reviewing your digital health tools at light speed.

00:09:33: But AI cannot navigate hospital politics.

00:09:36: Oh, right?

00:09:37: It could not sit in a boardroom and convince the health system executive to fund department pilot let alone sustain it into an enterprise-wide rollout.

00:09:44: Yeah because you can automate code but you CANNOT automate trust.

00:09:48: Exactly Executives need look at human being before they sign off on systems that potentially alter patient care

00:09:55: Precisely!

00:09:56: AI can accelerate building of digital health But you still need human-to-human dynamics to actually sell the transformation and manage the change management within a hospital system.

00:10:06: And speaking of human trust, that brings us back to liability... The big one!

00:10:10: If we are empowering these multiagent ecosystems to write code execute workflows and make clinical suggestions We run into a massive legal blind spot like if the machine gets it wrong who actually get sued?

00:10:23: It really is the elephant in room for health tech.

00:10:26: Geoddy Pandey laid out exactly how this is playing in dentistry, and it's a complete legal gray area right now.

00:10:33: How so?

00:10:34: So imagine that Dentist uses an AI diagnostic tool follows its advice And the AI turns to be wrong.

00:10:41: The dentist is fully liable.

00:10:42: Right because they didn't go to dental school

00:10:45: Exactly!

00:10:45: And the dentist signed final chart.

00:10:47: But...and this is crazy part.

00:10:49: If they override the

00:10:50: A.I.,

00:10:51: and the AI was actually right They are still liable Because they ignored a logged algorithmic finding that was sitting right in front of them and the medical record.

00:11:00: So logically, if you're first-year clinician The safest legal move is to just agree with the software.

00:11:04: it becomes a defensive medicine tactic.

00:11:07: Yeah A defensive tactic.

00:11:08: But If They do That Don't.

00:11:10: They Slowly Lose Their Own Diagnostic Instincts Like They Just Calibrate To The Machine's Blindspots Without Ever Developing Human Intuition To Catch What It Misses.

00:11:19: You've just described exactly what Benjamin W. Nelson identified as one of the most insidious risks in clinical AI.

00:11:26: Nelson researches mental health, a safety and he divides the risk into two categories type I and type to harms.

00:11:33: break that for us.

00:11:33: so type i harms are acute immediate failures For example an ai chatbot completely missing a textual cue.

00:11:40: That patient is suicidal?

00:11:46: Right.

00:11:47: But type two harms are much harder to spot, right?

00:11:49: Yeah If Type I is an acute failure, Type II must be a systemic degradation.

00:11:54: Exactly We're talking about the slow de-skilling of the clinician over time De-skillin',

00:11:59: wow!

00:11:59: It's this creeping harm that emerges over long multi turn interactions.

00:12:04: No single message or suggestion from AI looks dangerous in isolation but overtime The human relies on machines reasoning so heavily they fostered dependence

00:12:13: They lose their own analytical capabilities.

00:12:16: Right, but Eddie Hernandez is actually trying to engineer a solution against this with the platform called Stillpoint.

00:12:23: Okay so how does a platform design against human dependence?

00:12:27: Stillpoint is an AI designed strictly to support the therapist never to replace them or make the final call Interesting.

00:12:34: The mechanism is fascinating, the AI tracks the continuity of therapy sessions.

00:12:39: it surfaces historical insights from past conversations but forces a human therapist to make clinical synthesis.

00:12:48: As Fernandez phrased that architecture specifically designed.

00:12:52: keep clinician outsourcing inside their own head.

00:12:55: Outsourcing the inside of their own head, that perfectly captures the psychological risk over reliance.

00:13:00: It really does!

00:13:01: And

00:13:01: it's not just a clinical risk for individual doctor either... ...it is massive enterprise risks to hospitals themselves.

00:13:07: Definitely.

00:13:07: Seth Jeremy Katz highlighted recent incident where an AI vendor breach exposed one point four million patients across seven different health systems.

00:13:17: And the underlying dynamic there is that AI is accelerating the velocity of data processing, which inherently accelerates the risk of a breach.

00:13:25: Right

00:13:26: and health systems cannot outsource the accountability.

00:13:29: when an autonomous agent starts making decisions triggering external workflows and moving data around then vendor gets hacked.

00:13:37: The hospital still left holding bag in eyes public law.

00:13:42: So if the hospitals are holding all this liability, how do their regulators ensure that software is actually safe before it even reaches the hospital's server?

00:13:50: It

00:13:50: was a huge challenge.

00:13:51: Rudolph Wagner brought up really pragmatic... bottleneck strategy regarding the EU Medical Device Regulation or EUMDR.

00:13:59: The EMDR is this strict European rulebook dictating what software counts as a safe medical device, right?

00:14:05: And Right now Wagner points out there are nearly one hundred thousand health apps on the market.

00:14:10: regulators just do not have the manpower to audit a hundred thousand individual software developers.

00:14:14: No it's impossible.

00:14:15: but under the regulations app stores like Apple and Google legally act as distributors.

00:14:21: Oh, I see where this is going.

00:14:22: So instead of trying to police a hundred thousand different developers you just police the two toll booths they all have to drive through?

00:14:29: Exactly!

00:14:30: Wagner's arguing.

00:14:31: regulators should audit the massive platforms hosting the apps forcing Apple and Google to enforce the EU-MDR compliance before an app can even be listed.

00:14:40: It's really the only scalable way to enforce governance.

00:14:43: but...you know While regulators, vendors and hospital executives sit around debating liability app store audits in compliance we kind of have to look at what the patients are doing.

00:14:53: Yeah because patients are tired of waiting for the industry to sort out its governance issues.

00:14:58: they are bypassing the traditional health care system entirely

00:15:01: And The numbers on this patient by pass or just staggering secret.

00:15:05: Burge van Roigen and Krista Calpe shared data showing that two hundred million people are asking chat GPT health questions every single week.

00:15:15: Every Week!

00:15:15: Two hundred million.

00:15:17: That's incredible and the context of when those queries are happening is most telling part.

00:15:22: Seventy percent of these conversations taking place outside standard clinic hours.

00:15:26: And Birchman-Rosian noted that in Europe, The average wait time to see a specialist is seventy days.

00:15:32: SEVENTY DAYS

00:15:33: Which is wild...and the healthcare industry often looks at those two hundred million queries & panics.

00:15:39: They treat AI Chatbots as this dangerous misinformation virus, you know worrying about AI hallucinations and self-diagnosis.

00:15:48: People aren't choosing AI over doctors because they think a chatbot is smarter than the specialist?

00:15:52: Exactly!

00:15:52: They are choosing AI because the clinic has closed at two AM or they simply cannot wait seventy days for human to answer basic question.

00:16:00: This is symptom of massive access drought.

00:16:03: Not a symptom patient ignorance.

00:16:05: It's structural design problem in health care.

00:16:08: But there is a physical limit to how much of that access problem we can solve just by throwing more digital tools at patients.

00:16:14: True

00:16:14: Wolfgang

00:16:15: Schleifer brought up a vital counterpoint in this, he calls it the tolerability crisis.

00:16:20: What happens

00:16:20: when you push too much tech onto patient?

00:16:22: Like what does that look

00:16:23: like?

00:16:24: So think about post-operative patient recovering home.

00:16:27: We're giving them digital trackers wearable monitors virtual portals automated check and apps.

00:16:33: It's a lot

00:16:34: Right.

00:16:34: Schleifer warns that this bombardment is actually causing cognitive fatigue.

00:16:39: If an app pings a recovering patient with an automated alert about their heart rate but provides zero clinical context or human reassurance, That digital interface suddenly becomes the source of anxiety...

00:16:51: Becomes a burden not a benefit.

00:16:53: Exactly we are hitting the limits of human tolerability for digital monitoring

00:16:58: which means The innovation has to evolve.

00:17:01: Philip Radcliffe argued this point perfectly.

00:17:03: We have to shift away from these disconnected individual points solutions, you know an app for those so wearable for that and start designing holistic care pathways.

00:17:12: Yeah You cannot solve fragmented care with another fragmented out?

00:17:15: You just can't!

00:17:16: You HAVE TO design the technology around the entire patient journeyer From initial screening all through treatment & follow

00:17:23: up And we're actually seeing that shift toward Holistic Pathways.

00:17:26: change the economics of MedTech right now.

00:17:28: Pankaj Chutrala highlighted an amazing use case in India with the IntelliJoint KNE system.

00:17:35: It completely rethinks how you map anatomy for a knee replacement.

00:17:39: How does it work without relying on, you know massive traditional infrastructure?

00:17:44: Well traditionally You need a preoperative CT scan extra incisions for tracking pins and A massive expensive robotic arm taking up half of operating room.

00:17:53: Right The IntelJoin system replaces all of that with an AI-powered tablet and a micro sensor used during the actual incision.

00:18:00: Oh

00:18:00: wow!

00:18:01: Yeah, the software maps the joint dynamically in real time.

00:18:04: it cuts out the imaging costs reduces the hardware footprint improves at high precision.

00:18:09: AI doesn't just belong to massive heavily capitalized metro hospitals.

00:18:14: It really democratizes the precision.

00:18:16: It's a perfect example of evaluating the economics of the entire workflow, not just the hardware itself.

00:18:23: Dr.

00:18:23: Cyril Vithalani proved this exact point with his data on robotic surgery economics.

00:18:29: Fascinating!

00:18:29: Because hospital administrators always push back asking isn't robotic surgery just vastly more expensive?

00:18:35: And Vithulani's answer is well it depends entirely on what column of the spreadsheet you're looking at.

00:18:40: Right because if you only look The cost of the operating room itself, then yes.

00:18:44: The robot costs more.

00:18:46: he actually ran the numbers on a sigmoid colectomy?

00:18:48: The robotic procedure cost about sixty thousand rupees more in the operating Room compared to open surgery.

00:18:54: It was roughly two point seven nine lakh versus two point seventy-two lack.

00:18:58: But that theater bill isn't the total episode of care.

00:19:01: when you zoom out and look at the holistic pathway the math completely flips.

00:19:06: because Of the robotic precision the patient spent fewer days occupying a hospital bed.

00:19:10: Yes.

00:19:11: They avoided wound infections that cause expensive readmissions, and they returned to work twenty-six days faster than with open surgery.

00:19:18: That's a huge difference

00:19:19: It is!

00:19:20: When you calculate the total episode cost The robotic surgery was actually significantly cheaper.

00:19:24: overall.

00:19:25: it Was three point three seven lakh compared to four point seventy nine lakh for the open surgery.

00:19:30: It compressed recovery timeline And reduced downstream complications.

00:19:33: But if hospital administrator siloed only looking at the upfront invoice from that one operating room, they missed this systemic savings entirely.

00:19:41: They miss The Big Picture.

00:19:42: Right, it all comes back to seeing the interconnected systems whether that's multi-agent AI workflows legal liability chains or holistic patient pathways.

00:19:52: Yeah

00:19:52: we've covered a massive amount of ground today from the friction of clinical implementation and the rise of agentic ecosystems To the complex legal gray zones And the access droughts driving patients to digital front doors.

00:20:04: We really have...and as you digest All this I want to leave You with a final thought experiment based on everything.

00:20:12: So, we talked about the clinician's fear of trusting the AI and legal trap overriding it.

00:20:18: As these AI models become deeply seamlessly embedded into every single step with a care pathway will eventually reach point.

00:20:26: mutual-legal dependency?

00:20:29: Yeah

00:20:30: think about future where doctor isn't legally permitted to diagnose patient without an AI statistical confirmation but in AI is not legally permitted prescribe treatment Both entities entirely dependent on the other to move forward.

00:20:45: It is a balance of power that industry, regulators and hospitals are going.

00:20:52: That is a profound dynamic to end on and one that will undoubtedly shape the next decade of health tech.

00:20:58: If you enjoyed this episode, new episodes drop every two weeks!

00:21:01: Also check out our other editions on ICT and Tech Insights, DefenseTech Cloud digital products & services Artificial Intelligence and Sustainability in Green ICT.

00:21:10: Thank You so much for giving us your time today And exploring these vital shifts with us.

00:21:15: Don't forget to subscribe So never miss an insight.