Best of LinkedIn: Health Tech CW 34/ 35

00:00:00: This episode is provided by Thomas Allgeier and Franus, based on the most relevant LinkedIn posts about health tech from CW-ThirtyFour and ThirtyFive.

00:00:09: Franis equips HealthTech providers with The Market Intelligence to identify which hospitals to target... ...and how to reach decision makers for hospital digitalization as a result of the Crankin' House soup sketch sets.

00:00:21: You can find more info in the description.

00:00:23: Alright so

00:00:24: Well just imagine this for second.

00:00:26: It's uh two AM Okay.

00:00:28: A patient gets this super confusing lab result pushed right to their phone,

00:00:32: we've all been there?

00:00:33: Right?

00:00:34: and instead of you know panicking and waiting three agonizing days for their doctor to finally call them back, they just feed that raw data into an AI on their phone.

00:00:42: And boom!

00:00:43: Plain language diagnosis.

00:00:44: Instantly...and I mean this isn't some sci-fi pitch for the future.

00:00:47: This is happening right now like millions of times a week.

00:00:50: It absolutely is.

00:00:51: You know today we are extracting the most critical insights from a stack of curated posts by health tech professionals To well understand exactly how this shift is rewiring the entire industry.

00:01:03: Yeah it's huge.

00:01:05: So if you're a digital transformation leader, a CTO or just navigating the health tech space this deep dive is basically your roadmap.

00:01:13: Exactly we've got a lot to cover.

00:01:15: We really do!

00:01:16: We are going look at massive inwards shift of big electronic health record platforms The hidden economic time bomb of AI agents which is fascinating.

00:01:25: Oh

00:01:25: that one's wild

00:01:26: And then controversial push actually remove clinicians from loop plus raw reality of AI and drug discovery.

00:01:35: So let's start right at the bedrock of Hospital IT, which is The Electronic Health Record or EHR.

00:01:40: Right?

00:01:40: The foundation?

00:01:42: Because the biggest players in this space are currently redesigning how data and AI interact At very core their platforms.

00:01:49: Yana Ankudinova pointed out a major strategic pivot recently.

00:01:52: Oh yeah!

00:01:53: The epic and Oracle shift.

00:01:54: Yes They are pulling their AI ambitions inward, like instead of building open systems where outside developers just plug in third-party AI tools.

00:02:03: These giants they're embedding AI directly into their own walled

00:02:07: ecosystems.".

00:02:09: That's a crucial observation.

00:02:10: honestly I mean at Epic's user group meeting...they leaned heavily to what we call the agent factory.

00:02:17: Yeah

00:02:17: think if it is this highly secure sandbox for health systems AI agents that just act natively across clinical workflows.

00:02:25: Oh, wow!

00:02:27: And the engine running it is Cosmos which is Epic's massive aggregated patient data set and Oracle Health is... well they're taking a very similar path right

00:02:38: with their clinical AI agent

00:02:39: exactly.

00:02:40: They've integrated to handle documentation and medical coding directly inside the EHR.

00:02:47: So it's a core platform capability now not just some add-on you buy from the startup

00:02:52: and The ROI on that is already showing up in the real world, too.

00:02:55: Dr.

00:02:55: Christina Sepuentes highlighted this staggering metric.

00:02:58: Oracle health embedded AI cut physician documentation time per patient by thirty seven percent at Atlanta care.

00:03:05: Thirty

00:03:05: seven percent.

00:03:06: That is that huge.

00:03:07: It's a massive operational win.

00:03:09: But okay here's where I have to push back a little bit.

00:03:11: if the big vendors are doing all the computational thinking in house essentially creating this You know, walled garden.

00:03:18: Doesn't this leave downstream interoperability vendors totally out in the cold?

00:03:23: I mean you would naturally assume that right yeah but The reality of clinical data makes those downstream vendors more valuable than ever.

00:03:31: Because even when raw data leaves these massive systems via standardized feeds like FHIR Which is basically the API standard for health care Right it still suffers from severe duplication and something called semantic drift.

00:03:43: Okay, hold on.

00:03:44: Semantic drift?

00:03:45: For those of us who don't spend our days in health data architecture what exactly does that look like in practice?

00:03:51: Think about a highly consequential game of telephone across different hospital departments.

00:03:56: Okay!

00:03:56: Telephone I follow...

00:03:57: So the patient enters ER with specific condition let's say type two diabetes with kidney complications.

00:04:04: by time this record is passed to specialist then translated and finally dumped into research database

00:04:12: The terminology just drifts

00:04:13: Exactly.

00:04:14: It might just be coded as a generic metabolic disorder, the meaning degrades completely.

00:04:19: Wow!

00:04:19: ModestLoco actually shared a mind-boggling data point on this.

00:04:23: Hospitals generate about fifty petabytes of data annually yet only three percent is used analytically.

00:04:29: Wait...fifty petabytes and only three per cent are usable?

00:04:33: Yeah How's that even

00:04:34: possible?!

00:04:35: Because it an absolute mess.

00:04:37: Its structured codes mixed with unstructured physician notes PDF scans faxes all tangled up in strict compliance needs.

00:04:45: Right,

00:04:45: privacy stuff!

00:04:46: Exactly so.

00:04:47: even if Epic and Oracle keep their advanced reasoning native to their platforms the value of independent curation and normalization layers outside those walls goes through the roof

00:04:57: because someone still has to clean it up.

00:04:59: yes

00:04:59: someone has to reconcile that ninety seven percent of messy data before a downstream AI model can accurately trend a population's health.

00:05:06: The real opportunity for tech professionals right now is solving.

00:05:12: perfectly illustrates a quote from Meenal Shah regarding Epic's recent announcements.

00:05:16: She basically said the bottleneck for innovation is no longer technology itself.

00:05:21: Right, it is organizations capacity to change.

00:05:23: Exactly Having that AI capability natively in software Is very different than creating actual business value with it.

00:05:30: You still have to redesign your human workflows

00:05:33: And redesigning those workflows To include AI brings us to massive under discussed challenge The money.

00:05:42: If Epic and Oracle are encouraging hospitals to run constant autonomous AI agents on their internal data, who's paying for all that computing power?

00:05:50: Right.

00:05:51: Which transitions us perfectly into the economics of this.

00:05:53: we're seeing a really rapid evolution in AI consumption.

00:05:56: We really are

00:05:57: Like...we started with small chat interfaces moved models sort-of think through or prompt And now have fully Autonomous Agents that carry out multi step objectors.

00:06:08: But as you mentioned Autonomus means The meter never stops running

00:06:12: And every step up that ladder brings exponentially more computing power.

00:06:16: Lisa Mattson posted a really urgent warning for tech leaders about the economics of this shift...

00:06:21: It's all about token consumption, right?

00:06:23: Exactly!

00:06:24: AI costs in hospitals are rising rapidly but it is not because vendors hiked their software license fees…it was just agent work flows.

00:06:33: multiply token consumption.

00:06:34: Ok

00:06:35: let's clarify tokens quickly.

00:06:37: How should, say a CTO or finance director be thinking about token consumption in this context?

00:06:43: Think of token pricing like paying a taxi by the mile.

00:06:46: A simple chat query—like a doctor asking an AI to summarize a single patient file is quick trip down the block.

00:06:53: Right it costs fractions of a cent

00:06:55: Exactly.

00:06:56: But an autonomous agent That is asking the taxi to drive around the entire city all night checking every single clinic note, cross-referencing lab results and comparing them against historical data without you even in the car.

00:07:07: The meter's just spinning exponentially—spinning out of control!

00:07:11: When a health system builds its own agents that usage is billed as per million token charge.

00:07:17: That's and argues that AI computing costs now urgently belong on the executive governance agenda.

00:07:23: Honestly, if you're a hospital CTO right now... ...that should keep up at night!

00:07:28: You aren't budgeting for predictable software licenses anymore….

00:07:31: …you are budgeting for autonomous machine thought?

00:07:34: Yeah – which is completely unpredictable.

00:07:36: But that raises an even bigger question How can a hospital possibly govern something that learns, acts autonomously & burns cash by second?

00:07:46: It's a huge

00:07:46: problem because traditional software updates are predictable.

00:07:49: You QA test them, you deploy them to production and your done.

00:07:52: Governing an AI agent feels completely different.

00:07:55: it is A COMPLETELY DIFFERENT PARADYME.

00:07:57: Phil Baldacci points out that AI Agent Development Cycles Are Night & Day Compared To Package Software.

00:08:02: You Can't Just Deploy An Agent And Walk Away

00:08:04: Right.

00:08:05: These Agents Are Crossing Organizational Siloes And Health Systems Are Struggling To Figure Out How To Scale Them Safely Without Breaking Compliance Or Blowing Their Budgets.

00:08:13: So who actually owns the liability and performance of these agents?

00:08:18: That's exactly what David Abel brought up regarding radiology.

00:08:21: As AI tools enter radiology workflows to scan for anomalies, they require incredibly clear governance.

00:08:28: Meaning someone is on a hook.

00:08:29: Yes Someone specifically has to own the performance monitor usage And decide when adjust, Baws or scale the AI.

00:08:38: Because models drift over time

00:08:40: Data inputs change

00:08:41: Exactly.

00:08:42: If you don't have a designated human owner monitoring that specific agent, efficiency and clinical trust just evaporate.

00:08:49: So how do we corral this?

00:08:50: Because right now it sounds like the Wild West.

00:08:52: You've got agents running wild burning tokens drifting in accuracy.

00:08:56: It is The Wild West.

00:08:57: so How does a massive health system actually control This infrastructure?

00:09:01: well Lloyd Price predicts A major structural shift to solve this.

00:09:05: he thinks healthcare will consolidate around dedicated AI gateways.

00:09:09: He actually compares it to Stripe's seven billion dollar acquisition of Open Router.

00:09:13: Open

00:09:14: router?

00:09:14: Okay, context for that...

00:09:15: Right!

00:09:16: For Context, Open Ratter acts as a unified API gateway for general software developers to access hundreds different AI models through one single control point.

00:09:28: It just routes the prompt to most cost-effective model automatically.

00:09:32: That smart.

00:09:33: But a general-purpose gateway like that wouldn't meet strict clinical privacy requirements, right?

00:09:37: Prifisely.

00:09:38: You can send protected health information through a generic commercial router.

00:09:42: so Price sees the digital health ecosystem racing to build dedicated model control planes specifically designed for clinical governance.

00:09:50: So they throttle the usage.

00:09:51: Yes

00:09:52: these gateways will throttle token usage enforce IPA compliance and give the CTO A single dashboard To see exactly what every AI agent in the hospital is doing.

00:10:01: That makes total sense.

00:10:02: You need a bouncer at the door of your data center, but this urgent need for strict governance forces are really difficult.

00:10:08: conversation

00:10:09: it Really does.

00:10:10: if the agents are acting autonomously and we're building these massive automated gateways to manage them What happens to the humans?

00:10:17: Yeah Which brings us to a huge debate redefining the human role in clinical care.

00:10:22: This is where the industry conversation gets highly controversial.

00:10:26: Joshua Lou shared A very provocative stance recently.

00:10:30: Oh The Human In The Luke Thing.

00:10:31: Yes

00:10:32: He argues that the traditional human in-the-loop model is the wrong hill for clinicians to die on.

00:10:37: Wait, getting the human out of a loop?

00:10:40: That sounds like a massive liability lawsuit waiting to happen if an AI misses critical diagnosis!

00:10:45: I know it sounds crazy...

00:10:47: Every time we talk about medical A.I safety The emphasis has always been keeping a doctor in the loop.

00:10:53: How does he justify this

00:10:54: risk?!

00:10:55: It sounds counterintuitive but his reasoning is rooted into math and access.

00:11:00: He points out that the entire premise of introducing AI is to address this severe shortage of clinicians.

00:11:06: Right?

00:11:07: If you mandate a human must review every single AI action, You haven't solved the bottleneck!

00:11:12: You've just created faster machine waiting on attired humans.

00:11:15: Huh... That's really good point.

00:11:17: Furthermore, AI simply better at strict standardization than humans are and AI doesn't get fatigued by the end-of-a-twelve hour shift And it does not deviate from evidence based guidelines.

00:11:26: So he suggesting full autonomy

00:11:28: For specific use cases.

00:11:29: yes He predicts that for low-risk tasks like interpreting routine baseline labs or basic imaging, the goals should be safely getting clinicians out of a loop entirely.

00:11:40: Just skip them?

00:11:41: Yeah!

00:11:42: The benefit of massively improved patient access outweighs theoretical risk provided the AI is validated.

00:11:49: And honestly, from an operational standpoint that aligns perfectly with a very practical point made by Charles Hummel regarding ambient scribes.

00:11:57: Oh!

00:11:57: The note-taking AI

00:11:59: Right.

00:12:00: We've all heard about AI listening to doctor patient conversations and writing notes.

00:12:04: Hummel noted this tech only actually adopted by doctors if it's ready instantly or imperfectly.

00:12:10: That makes sense.

00:12:11: The exact moment a doctor has to manually intervene and edit the AI's work, the cost benefit of the tool just vanishes.

00:12:17: they might as well type it themselves

00:12:19: Exactly In many workflows.

00:12:21: human in loop isn't safety net.

00:12:23: They are friction that kills return on investment.

00:12:26: Wow By the way if you want stay ahead how these workflow technologies revolving Make sure to subscribe so you catch future editions of this deep dive, because the shift isn't just happening on the clinician side.

00:12:37: No it's not!

00:12:37: As humans get pulled out of the operational loop It completely shifts power dynamic for end user.

00:12:43: It really does and Jills Friedman brought up an incredible concept To describe his shift.

00:12:48: He calls it patient sovereignty.

00:12:50: Patient sovereignty?

00:12:51: I like that.

00:12:52: For last decade health care has used buzzword patient engagement Which if we are being honest basically just meant getting patients to comply with what the system told them to do.

00:13:03: Right, it was entirely top-down.

00:13:05: like please log into this clunky portal to see the message we sent you

00:13:09: Exactly.

00:13:10: but now Friedman points out that instead of waiting anxiously by the phone for a doctor to translate a test result Patients are taking their raw lab data feeding it in to AI tools at two AM and Getting plain language highly accurate explanations

00:13:25: back tour intro example

00:13:27: Exactly.

00:13:27: Yeah, nobody prescribed that AI to them.

00:13:30: no hospital authorized it.

00:13:32: But millions of people are doing in any way.

00:13:34: That is an engagement?

00:13:39: is such a vivid image.

00:13:41: It's patients taking absolute charge of their own diagnostics without asking permission, and the big health systems are finally recognizing they have to adapt to this.

00:13:51: Trevor Berceau highlighted that Sutter Health became the first organization.

00:14:01: They are acknowledging that if they don't provide a safe AI for the patient to use, the patient will just use a generic one on the internet.

00:14:08: And we're seeing the tech giants rigorously validate these patient-facing tools too.

00:14:12: Right Michael Radwin posted about how VerilyMe went out testing its AI symptom assessment tool and didn't rush it into market.

00:14:20: They tested across roughly eight hundred simulated clinical encounters

00:14:24: Eight hundred?

00:14:25: Yeah.

00:14:25: And crucially published their findings in JMIR formative research openly surfacing exactly where the AI was under-triaging patients so they could fix the model before deploying

00:14:37: it.

00:14:38: The technology is safely putting diagnostic power directly into the patient's hands.

00:14:43: It fundamentally changes the hierarchy of medicine, so we've explored how AI is changing daily hospital operations.

00:14:50: The economics have token consumption and empowering patient sovereignty.

00:14:54: right.

00:14:55: but for our final theme We really need to talk about the absolute edge of this technology

00:14:59: frontier.

00:15:00: Yeah

00:15:01: What happens when we point?

00:15:02: This massive agentic computational power?

00:15:05: not at Hospital workflows But at the fundamental building blocks of human biology drug discovery.

00:15:13: And the sheer scale of the computation here is difficult to even comprehend.

00:15:18: Jamie Church shared a recent example where hospitals successfully modeled a twelve thousand six hundred and thirty-five atom protein, but they didn't do it on traditional supercomputer.

00:15:27: They ran in on a quantum computer.

00:15:29: It's the largest biological molecule simulated in a quantum environment today.

00:15:34: A twelve thousand Adam protein on a Quantum Computer.

00:15:38: Just the physics of keeping a quantum computer stable long enough to run that calculation is mind-bending.

00:15:44: It really is,

00:15:45: but practically speaking how does modeling a massive protein translate actually making drugs at work?

00:15:51: To understand it we have look at Daphne Koller's recent insights on causal AI and in CETRO

00:15:58: Causal AI.

00:15:59: Historically, the pharmaceutical industry has a massive failure problem.

00:16:03: nine out of ten drug programs fail in clinical trials

00:16:06: Nine out of Ten

00:16:07: ouch.

00:16:08: and The main reason they fail is that the biological mechanism?

00:16:10: That drug targeted didn't actually drive the disease.

00:16:13: we've known for A decade that if it drug target Has strong human genetic evidence backing It up it Is two to four times more likely To succeed.

00:16:20: but Human Genetic data is Incredibly noisy And foggy.

00:16:24: so foggy.

00:16:24: So how does causal AI clear the fog?

00:16:27: Well, traditional AI just looks for correlations.

00:16:29: It says you know people with this disease often have this biomarker.

00:16:33: but correlation isn't causation right.

00:16:36: causal AI however tests what if scenarios.

00:16:40: it essentially runs billions of virtual biology experiments to see if flipping a specific genetic switch actually causes the disease.

00:16:47: Oh wow.

00:16:47: Caller's team at Incitro built an AI to map what they call precision phenotypes.

00:16:53: by mapping these root causal pathways, They found thirty times more genetic associations than traditional clinical staging methods.

00:17:00: Thirty times more targets is incredible but did it actually work?

00:17:03: Like the virtual experiments translate to accurate predictions...they

00:17:08: put their causal AI model which they called our Virtual Human To The Ultimate Test.

00:17:13: They asked it to predict the success of historical phase two clinical trials without giving AI actual outcomes.

00:17:20: A blind test?

00:17:21: Exactly!

00:17:22: In its top predictions, The AI's false positive rate was only ten percent in metabolic disease and twenty-one percent in cardiac disease.

00:17:29: Compare that with industry standard.

00:17:31: What is normal failure rate?

00:17:33: Historical Failure Rate for those types of trials Is fifty seven percent.

00:17:37: Wow So the AI dramatically outperformed human lead discovery.

00:17:42: That is a staggering reduction in trial failure.

00:17:45: I mean, you are shaving years and billions of dollars off the development cycle And Pooja Apatel shared findings that echoed this exact kind of breakthrough, but in oncology.

00:17:57: Oh the breast tumor study?

00:17:59: Yes!

00:18:00: An open-source AI model analyzed three hundred and thirty thousand breast tumor cells... ...and identified structural patterns human pathologists had missed for a century.

00:18:10: A

00:18:11: century?!

00:18:11: Yeah… And In that same month.. ..a randomized controlled trial of AI supported mammography showed it reduced aggressive interval cancers by twelve percent entirely without the false positive rate.

00:18:22: Patel made a crucial caveat there though.

00:18:25: in that specific oncology workflow, The AI sharpens the expert radiologist.

00:18:30: it doesn't replace them but it definitively proves that AI can see biological and molecular signals That the human eye simply cannot process.

00:18:38: okay I have to play devil's advocate here before listening to this.

00:18:41: With quantum computers folding massive proteins, causal AI predicting clinical trials with unprecedented accuracy and vision models spotting century-old blind spots in tumors it sounds like we're going to cure every known disease by next Tuesday.

00:18:57: right?

00:18:58: are we as an industry getting swept up in our own hype cycle?

00:19:02: It's a vital question to ask, and Christian Hein provided much-needed sobering reality check.

00:19:07: What do you

00:19:08: say?

00:19:09: He specifically pushed back on recent claims by Silicon Valley leaders who are predicting that AI will cure most human diseases in the next five years.

00:19:19: He points out that while AI absolutely compresses the discovery and development phases, human biology remains vastly complex.

00:19:26: And infinitely heterogeneous?

00:19:28: Exactly!

00:19:29: Right a molecule designed perfectly in a quantum simulation still has to survive contact with the messy reality of the human body.

00:19:35: Exactly discovering The Perfect Molecule is only step one.

00:19:38: A new drug Still requires Human Clinical Trials To Prove Actual Safety Across Different Demographics.

00:19:43: You have

00:19:43: to prove the dosing limits.

00:19:45: Yes You have to prove actual clinical benefit in the real world.

00:19:50: Then, it has to be physically manufactured at a massive scale.

00:19:54: Regulators have to scrutinize and approve

00:19:56: it.".

00:19:57: So...the physical world is still the bottleneck?

00:19:59: Yes!

00:20:00: AI is accelerating the intelligence layer finding targets and designing molecules.

00:20:05: but pharmaceutical companies will still carry the mass of clinical risk and executing the trials.

00:20:11: The technology's moving at light speed…but biology

00:20:16: That frames it perfectly.

00:20:17: The AI can give us the precise blueprint, but we still have to physically build the house and see if it withstands a storm

00:20:23: which really brings this entire deep dive full circle.

00:20:26: think about everything We've unpacked today.

00:20:28: It's a lot from epic an oracle shifting AI capabilities deep inside their EHR To CTO is grappling with the governance of token burning autonomous agents?

00:20:37: We've seen patients claiming sovereignty over there own data at two AM And causal AI unraveling the genome.

00:20:44: The common thread across all of this is clear.

00:20:46: The technology itself, it's no longer the rate limiting step... It's

00:20:49: a human element!

00:20:50: ...it's

00:20:50: us."

00:20:51: The true differentiator for any digital transformation professional health system or pharmaceutical company going forward isn't just having access to the best AI model-the models will commoditize.

00:21:03: The real differentiator is their human capacity to thoughtfully adapt the tools, redesign legacy workflows and manage incredible pace of change safely.

00:21:14: It really makes you wonder about the long-term trajectory, like if patients are claiming sovereignty over their diagnostics at two AM and AI agents are natively solving complex coding and triaging inside the EHR.

00:21:27: And drugs are becoming hyper personalized through causal AI?

00:21:30: How long until the traditional concept of visiting a physical hospital becomes entirely obsolete for everything except acute trauma and physical

00:21:37: surgery?".

00:21:38: We might be watching The Virtualization Of Healthcare happen in real time

00:21:41: If you enjoyed this episode, new episodes drop every two weeks.

00:21:44: Also check out our other editions on cloud insights and sovereignty digital products in services AI and agentic systems green ICT and sustainable

00:21:52: A.I.,

00:21:52: ICT & Tech Insights and DefenseTech.