Best of LinkedIn: Health Tech CW 24/ 25

00:00:00: This episode is provided by Thomas Allgeier and Frennus, based on the most relevant LinkedIn posts on health tech in CW-Twenty Four and Twenty Five.

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

00:00:18: as a result of Welcome to the deep dive everyone.

00:00:26: We are really glad you're here with us.

00:00:28: over the next few minutes we're going to explore The top health tech trends that have just been dominating the conversation on LinkedIn over the past two weeks.

00:00:36: Yeah, it's been a really busy couple of weeks on the timeline.

00:00:38: lots of debate

00:00:39: Oh for sure?

00:00:40: We've curated the most vital insights from leaders across the space.

00:00:43: So will be looking at everything from AI integration To the really gritty unglamorous realities Of medical device adoption.

00:00:51: so Okay, let's unpack this starting with what seems to be the biggest wake-up call in the industry right now which is that AI adoption and healthcare has officially shifted from those Shiny experimental tech demos too.

00:01:04: The actual messy reality of hospital operations.

00:01:08: Yeah And we were actually seeing that shift reflected at the very earliest stages of company building.

00:01:15: Look at the data.

00:01:16: so Garcia Hortes did a really fascinating analysis the latest Y-combinator batch.

00:01:21: Oh, health care startups?

00:01:22: Right!

00:01:22: The healthcare startup and given the whole hype cycle around generative AI you'd probably assume every founder is trying to build like...the ultimate AI doctor

00:01:30: right?

00:01:30: Like a neural network that just replaces human clinical reasoning.

00:01:33: entirely

00:01:34: Exactly.

00:01:35: but this signal was pointing in exact opposite direction.

00:01:37: The most crowded cluster of start ups isn't focusing on patient facing diagnostics at all.

00:01:42: No they're building BtoB autonomous agents for least glamorous parts

00:01:47: Operations.

00:01:48: We're talking referrals, facts and takes scheduling prior authorizations revenue cycle management

00:01:55: which is incredibly smart when you think about the architecture of modern medicine.

00:02:00: I mean health care's essentially this ecosystem Of fragmented systems relying on manual handoffs And you know unstructured documents

00:02:08: right somebody literally printing a PDF in faxing it.

00:02:10: exactly so.

00:02:12: agentic systems can create massive measurable ROI just by automating that administrative burden.

00:02:20: They're capturing value in the operational layer long before they even touch.

00:02:29: That makes perfect sense, especially when you look at the track record of clinical tools that fail.

00:02:33: Like Dr Vinyan Masangi shared a pretty stark reality check on this recently.

00:02:38: Yeah I saw that

00:02:39: He pointed out that AI in hospitals is basically doomed to fail unless clinicians are helping map workflows from day one.

00:02:45: he sees these brilliant engineers building tools with solid algorithms clean UIs and then nobody uses them

00:02:51: Because it doesn't fit the actual job.

00:02:53: Exactly!

00:02:54: Not because the tech is flawed, but because nobody designed a ward round into product workflow.

00:02:59: And so it's like... So many of these startups have brilliant engineers.

00:03:03: But as Dr Masangi asks Who in your team has done a Ward Round?

00:03:07: Such a good question

00:03:09: Right If we're building cutting-edge Tech that a nurse literally doesn't have time to open during a twelve hour shift Aren't we just building sports car for dirt road?

00:03:17: Wow

00:03:18: Yeah, that supports car analogy spot-on.

00:03:21: What's fascinating here is we consistently confuse a technology problem with an implementation problem.

00:03:27: Yes.

00:03:27: Print and Paul articulated this translation gap brilliantly.

00:03:30: he noted that AI tools perform perfectly.

00:03:33: in you know controlled testing environment.

00:03:36: they often just fall apart the second they meet different patient population or some unfamiliar EHR workflow The

00:03:41: dreaded EHR integration?

00:03:43: The

00:03:43: last mile of integration was everything without it.

00:03:46: instead of the AI managing workload the clinician ends up managing the AI.

00:03:50: Which is the worst case scenario?

00:03:52: Exactly, it generates alert fatigue.

00:03:54: It creates more manual data entry.

00:03:56: I've always

00:03:56: compared this to the history of electronic medical records.

00:03:59: like Vivek Aurora pointed out that workflow mismatch Is the exact same reason EMR go lives have historically failed.

00:04:05: Oh

00:04:05: absolutely

00:04:06: It's almost never that the underlying database doesn't work.

00:04:09: Its change fatigue, I mean if users are already overloaded with parallel changes even a technically perfect system gets rejected because The clinical reality just doesn't align With the assumptions baked into the software.

00:04:22: But you know we also have to look at what happens when developers actually get That alignment right true?

00:04:27: Because the upside is undeniable.

00:04:29: Jeff DeLello shared insights from the twenty-twenty six Phillips future health index and they survey'd Over twenty two thousand healthcare professionals globally.

00:04:39: That's a huge sample size

00:04:41: massive and the data shows that when AI is successfully integrated into the workflow it genuinely works.

00:04:46: Thirty six percent of US health care professionals reported The AI has increased their capacity to see an average of five additional patients per week.

00:04:54: Five extra patients a week per provider.

00:04:58: Wow, I mean multiply that across an entire health system And that is a massive metric.

00:05:03: It fundamentally changes hospital economics And more importantly, it gives clinicians time back to fulfill the purpose that brought them to medicine in their first place.

00:05:13: Sitting with patients listening caring.

00:05:16: When the tech is integrated properly, it absorbs friction rather than creating it.

00:05:20: Exactly!

00:05:21: Okay so integrating AI into the workflow is clearly the first massive hurdle.

00:05:26: but let's say you actually get that right... The nurse uses a tool and the hospital adopts it.

00:05:30: You're still not done?

00:05:31: No-you are NOT at finish line yet because keeping it safe, compliant & unbiased over time seems to be second major hurdle.

00:05:38: And I want throw a theory.

00:05:41: I always just assumed Hypo-compliance was essentially a matter of keeping data encrypted at rest and in transit like, uh...a security checkbox you hit right before launch.

00:05:51: Oh man!

00:05:52: That is the exact trap most founders fall into Really?

00:05:55: Yeah And it's why Navineeth recent posts struck such accord.

00:06:00: He pointed out that hypercompliance isn't a feature you just tack on at the end of development.

00:06:05: It requires foundational architectural decisions made on day one.

00:06:08: interesting because it's not Just about data storage anymore, it's about how The neural network actually processes that data

00:06:15: and the regulatory framework around.

00:06:17: That is surprisingly robust even if it feels old right.

00:06:21: Larry Trotter added a brilliant perspective to this.

00:06:23: yeah

00:06:24: He pointed out that everyone is running around looking for some official, shiny new hyper-compliant AI certification.

00:06:31: That just simply doesn't exist

00:06:32: right?

00:06:32: They want a stamp of approval

00:06:33: exactly.

00:06:35: but the reality Is that the Hypo security rule is twenty two years old But it already contains The exact controls needed to make AI compliant today the core principles, data privacy access control audit logs.

00:06:48: they don't change just because the processing mechanism is a neural network instead of you know relational database.

00:06:53: The foundational security principles might remain the same sure but the behavior or the software definitely does not

00:06:59: right.

00:06:59: yeah

00:07:00: unlike at traditional database an AI model ages the moment it goes live in a dynamic hospital environment and this brings us to a highly technical reality that Asim Khan broke down regarding AI drift.

00:07:13: This is such a critical concept because AI models, especially in patient monitoring.

00:07:18: They're trained on specific static slice of reality.

00:07:22: but hospitals are living breathing ecosystems.

00:07:25: Exactly think about it like calibrating highly sensitive weather barometer But then physically moving the house and sitting into totally different altitude.

00:07:33: Oh

00:07:33: that's great analogy.

00:07:35: The instrument isn't broken but baseline reality was tuned for no longer exists.

00:07:40: Kahn outlines four distinct types of drift.

00:07:42: we have to watch.

00:07:43: First, there's covariate or data drift.

00:07:46: Okay so what does that look like in practice?

00:07:47: Well maybe a hospital upgrades its physical sensors Or alters a triage workflow which fundamentally alters the data inputs feeding the algorithm.

00:07:55: Oh I see!

00:07:55: So AI is suddenly trying to read French when it was trained on Spanish.

00:07:59: Yes

00:07:59: exactly.

00:08:00: Then there's concept drift Which is highly urgent.

00:08:03: This is where fundamental relationship between inputs and outcomes changes.

00:08:06: Give me an example of this.

00:08:07: Imagine step down unit taking higher acuity patients.

00:08:12: The AI's threshold for what constitutes an emergency is now dangerously misaligned.

00:08:17: Right, because the baseline of normal has shifted.

00:08:20: Exactly!

00:08:21: Then you have Label Drift where the baseline rate of event changes...the classic example here a deterioration model trained pre-pandemic.

00:08:29: when COVID-IX hit overall case mix fundamentally shifted and those models struggled to adapt.

00:08:36: new baseline of respiratory failure.

00:08:38: That makes total sense.

00:08:40: And finally, there's bias drift which honestly might be the most insidious of all them.

00:08:45: Definitely

00:08:46: This is where the aggregate metrics look perfectly fine on a dashboard right?

00:08:49: But the model is quietly degrading for specific subgroups whether that's based on age ethnicity or care setting.

00:08:55: Which is why you cannot just deploy a model and walk away con argues.

00:08:59: You need to really robust three layer monitoring stack looking at input distributions unsupervised model behavior an outcome-based performance tracking.

00:09:08: Okay, but let me push back on that for a second.

00:09:10: But what happens when the AI is perfectly compliant?

00:09:14: It's learning from inherently flawed or opinionated data like even if you monitor it perfectly.

00:09:21: That blind spot is exactly what Sara Fuentes Molina highlighted.

00:09:25: We spend millions governing the enterprise AI tools inside hospitals... ...but almost none on consumer AI side.

00:09:34: Consumer-generative AI models inherit institutional bias because they treat institutional authority and clinical medical authority as the exact

00:09:43: same thing.

00:09:43: Because the algorithm literally doesn't understand contact?

00:09:46: Exactly!

00:09:46: She's a great example, so The Cleveland Clinic obviously a highly trusted brand They published an article on the pop culture concept of the ICC framing it loosely as a cognitive bias.

00:09:55: Which is, you know...a subjective cultural interpretation

00:09:58: Exactly!

00:09:59: But because its published under highly credible institutional banner an AI will scrape that and present to consumer clinical medical fact.

00:10:07: It literally doesn't the difference between a peer reviewed double blind study And a subjective blog post if they share same URL domain.

00:10:14: So we didn't remove the bias We just gave more authoritative distribution engine

00:10:18: Spot.

00:10:18: on Speaking of trust There is a massive blind spot in how we are bringing patients along for this ride.

00:10:25: If we look at recent analyses of clinical trials involving AI, the transparency is shockingly low.

00:10:31: Oh yeah!

00:10:32: The stats on these are pretty grim.

00:10:33: Olinoshvets shared a staggering statistic.

00:10:36: Researchers analyzed one hundred and fourteen cynical trials with an AI component... ...and fifty-five percent of the consent forms failed to disclose to participants how AI was actually using their data.

00:10:47: Wow Over half.

00:10:49: If we want patients to trust these systems, hiding the ball in the consent form is just a terrible way to start.

00:10:55: It really is.

00:10:56: Transparency is the foundation of clinical trust.

00:10:59: But you know if governance transparency and workflow integration are the main hurdles We have to look at where the industry is actually clearing them successfully.

00:11:06: Yeah Where's it working?

00:11:08: Because despite all these friction points there was one clinical domain undeniably conquering the market right now.

00:11:15: And that is imaging and diagnostics.

00:11:18: Oh, absolutely.

00:11:18: The numbers on this are wild!

00:11:21: Joel Kent analyzed the FDA's Q-One.

00:11:23: twenty-twenty six data for AI and machine learning enabled medical devices... ...the submissions are massively concentrated.

00:11:30: Like incredibly skewed.

00:11:32: Radiology had sixty nine submissions in just one quarter.

00:11:35: Cardiovascular was a distant second with ten.

00:11:38: everything else was just in the single digits.

00:11:41: Radiology is completely dominating the AI space

00:11:44: And if we connect this to the bigger picture, We have to ask why radiology has become The beachhead for AI adoption.

00:11:51: Dr.

00:11:52: Filippo Catamartiri offered a really provocative thought on This.

00:11:55: What did he say?

00:11:56: He argues that we fundamentally misunderstand what A radiologist actually does.

00:12:00: Radiology is not an imaging specialty.

00:12:03: It's a communication specialty That uses Imaging as its medium.

00:12:06: Oh...that's a fascinating distinction.

00:12:08: Explain the mechanism

00:12:09: there.

00:12:10: So think about the clinical handoff.

00:12:12: Patients and referring clinicians don't just need the images.

00:12:14: And they don't Just need a list of findings, They need decisions

00:12:17: right?

00:12:17: They need to know what to do next

00:12:19: exactly?

00:12:20: AI will inevitably read The image is faster it can detect findings measure lesions segment anatomy quantify volume.

00:12:27: But those are just mathematical computations.

00:12:29: the human radiologist Is the one who takes that mathematical uncertainty and translates It into an action

00:12:35: plan.

00:12:35: there you're the guide.

00:12:36: yes

00:12:37: The future of the specialty isn't about how many images you can read, it's about how effectively you transform AI computation into clinical action.

00:12:45: Here is where this gets really interesting because that traditional careful clinical approach is colliding head on with consumer tech ambition.

00:12:53: He definitely.

00:12:54: Joshua Liu posted a mid-journey famous for generative AI images and they announced their building an ultrasonic CT for whole body preventive screening

00:13:04: which was quite a pivot

00:13:05: huge pivot.

00:13:06: Their goal is to put a water-emergent scanner using hundreds of thousands of ultrasound transducers into wellness spas, completing a full body scan in sixty seconds.

00:13:15: And they want fifty thousand of these globally by twenty thirty one!

00:13:19: And notice how they are intentionally starting with the term wellness and focusing on body composition?

00:13:23: To bypass the FDA

00:13:25: Exactly...to bypass heavy FDA diagnostic regulation, before eventually moving into diagnostics.

00:13:31: Yes and the medical community is pushing back hard on this.

00:13:34: I mean from a pure physics standpoint ultrasound can't see past bone or air so it's effectively useless for imaging the brain or lungs.

00:13:43: but more importantly there are decades of clinical evidence showing the cascading harm of over screening asymptomatic individuals.

00:13:51: if you scan healthy people you will inevitably find tiny benign anomalies

00:13:56: Which leads to unnecessary biopsies.

00:13:58: Exactly!

00:13:59: Severe psychological anxiety for the patient, massive cost of health care system... it's a huge issue.

00:14:05: It is the ultimate friction point because The Tech World views this pushback as medical establishment stifling innovation.

00:14:12: They just want to build Star Trek Tri-Quarter and iterate their way into massive consumer scale.

00:14:15: Move fast and break things

00:14:17: Right but the Medical Field operates on mandate first.

00:14:20: do no harm.

00:14:22: Real medical innovation, the kind that actually shifts the standard of care.

00:14:26: It requires rigorous evidence-led validation

00:14:29: which brings us to what actual rigorous med tech looks like right now?

00:14:34: Angie Hulich shared data on a fascinating development called TETCH-TU-Twenty.

00:14:38: Oh,

00:14:39: I read about this!

00:14:39: It's a deep learning enhanced protocol that delivers comprehensive whole brain MRI in under two hundred and twenty seconds on standard one point five Tesla machines.

00:14:49: And let's break down why it matters.

00:14:51: By using Deep Learning to reconstruct the images faster and clearer they are dramatically cutting down time of patient spends in scanner

00:14:59: which is critical for acute stroke patients.

00:15:01: Exactly, where the medical mantra is literally time as brain shaving minutes off a scan preserves tissue and saves lives.

00:15:08: that has real clinically implementable innovation that fits perfectly into existing high stakes hospital workflows.

00:15:16: Right, it's not sitting in a wellness spa.

00:15:17: It is in critical care setting saving cognitive function Exactly.

00:15:21: And we're seeing similar rigorous innovation on the physical device side too.

00:15:24: Ryan Sanford posted about Boston Scientific real world adoption of the ELU Pro Bioemvelope at VA and Gainesville.

00:15:31: Oh!

00:15:31: The infection control device.

00:15:33: Yeah This isn't just a physical pouch for a cardiac implant.

00:15:37: It actively elutes antibiotics into the surrounding tissue over a specific time frame.

00:15:41: Wow, it fundamentally changes the infection risk profile For patients receiving these implants while promoting healing.

00:15:50: That kind of breakthrough doesn't come from a consumer marketing playbook.

00:15:53: No absolutely not!

00:15:54: It

00:15:55: comes from robust clinical trials and hard evidence that directly translates in to patient outcomes.

00:16:01: It just proves that while the flashy consumer applications get all the headlines, The real foundational shifts in health care are happening through meticulous peer-reviewed engineering.

00:16:12: So what does this mean for people actually trying to build and sell these technologies?

00:16:16: Because we've talked extensively today about the absolute necessity of workflow integration... ...the technical complexities monitoring AI drift.... And of course the non-negotiable requirement for hard clinical evidence.

00:16:28: Yeah it's a massive checklist

00:16:29: But I want to leave you with one final provocative thought to mull over, courtesy of Adam Tarinas.

00:16:35: You can have the most elegant engineering in the world!

00:16:38: You could have peer-reviewed data proving your device saves lives... ...you can even have a chief surgeon who absolutely loves your product and champions it at the hospital… And your product will still stall out completely.

00:16:51: why?

00:16:52: Because clinical buyin' is not the same thing as commercial buyin'.

00:16:55: Oh

00:16:55: that's harsh reality—the doctor isn't always the buyer.

00:16:58: No, we are shifting from an era where clinical efficacy drove adoption to an era when IT security and workflow integration possess veto power over clinical efficacy.

00:17:09: That's a great way of putting veto power!

00:17:11: You aren't just selling medicine anymore.

00:17:12: you were selling in enterprise IT integration.

00:17:14: that happens to do medicine.

00:17:15: if you don't map the real buying committee...you know?

00:17:18: The administrators, the IT security teams looking at day one IPA architecture..the procurement officers analyzing budget....you will fail.

00:17:27: Solving the algorithm is really only half-the battle.

00:17:29: If you're building a sports car, You better make sure that hospital actually has a paved road to drive it on and that The CFO was willing to pay the toll.

00:17:36: Exactly!

00:17:37: It always comes back to operational reality...you have to solve for the ecosystem not just the symptoms.

00:17:41: Russell if you enjoyed this episode.

00:17:44: new episodes drop every two weeks.

00:17:46: also check out our other editions on ICT in tech insights defense Tech cloud digital products & services artificial intelligence and sustainability and green ict.

00:17:57: Thank you so much for joining us on this deep dive into the realities of health tech.

00:18:01: Keep questioning algorithms and don't forget to subscribe!