Physician Entrepreneurship Is Rising. Medical School Still Isn't Teaching Business.

We are watching physicians lose power in the system at the exact moment we need physicians to gain it.

Corporations are consolidating at a pace that's swallowing private practices whole. Meanwhile, healthcare is more broken than ever. It’s inefficient, expensive, and desperately in need of the kind of front-line innovation only physicians can drive. The issue is most doctors were never taught how to do any of this.

In this article, I’ll discuss what medical training is missing, what the data says about physician entrepreneurship, and why one major organization is finally trying to do something about it.

Medicine Doesn't Teach the Business of Medicine

I went to medical school intending to learn pathophysiology and clinical medicine. But I came in with an entrepreneurship background, and I was surprised by how little innovation and business training was offered. As of 2016 (almost 10 years ago, but it’s the most recent data I could find), only 13 of 158 (8.2%) U.S. allopathic medical schools had a formal innovation and entrepreneurship program. Given the pace of innovation in healthcare over the past two decades, from artificial intelligence to new reimbursement models—this is concerning. Are medical schools keeping up? A 2024 survey of over 500 medical students found 90% reported no, minimal, or basic knowledge of the business of medicine.

Parambath and colleagues (including fellow Huddler, Sherman Leung) put it plainly:

The irony shows up even in residency selection. 61% of program directors said startup experience was a positive in applications. But only 25% thought taking time away from training to actually do it was worthwhile. The message to trainees is entrepreneurship looks good on paper. Just don't let it get in the way of medicine (story of my life).

Clinical expertise alone isn't enough anymore. The problems worth solving in healthcare require physicians who can both diagnose the dysfunction and build something to fix it. That combination (clinical insight plus business fluency) is exactly what's been systematically left out of our training. And the data on what physicians are actually doing out there suggests the instinct is already there but the training just hasn't caught up.

Current Trends in Business of Medicine

Healthcare is becoming more corporatized. Nearly 80% of all physicians are now employed by a health system or corporation—up from 60% in 2019. Four out of five physicians in the U.S. right now work for Optum, CVS Health, or HCA Healthcare. As I wrote recently:

The economics show horizontal and vertical integration increase negotiating power and help the corporate entity own the entire care journey. Physicians, systematically undertrained in business, are losing ground to the people who understand that game.

The formal response has been underwhelming. MD/MBA programs have more than doubled since 2002, from 26.4% to 60.9% of medical schools. But the numbers reveal a subpar response, with only 0.8% of medical students actually graduating with an MBA—roughly 160 students per year out of ~22,000 annual graduates. And the degree itself has been diluted, since a vast majority of dual programs require fewer business credits than a standalone MBA. The system responded to the demand by offering a credential but it didn't change what physicians actually learn.

Here's what the data elucidates, though: physicians aren't averse to business. The instinct is already there. What's missing is the pipeline.

A study of Massachusetts-licensed physicians found that the proportion who had ever founded a company—clinical practice, biotech, or otherwise—was higher for physicians who graduated in 1970 than for those graduating in 2010.

Source: Greenblatt (2021)

This shows that older physicians, trained before the corporatization era, were more likely to build something. And when physicians do break into business today, it happens late—on average, about 20 years post-graduation, when they're in their mid-to-late 40s.

Source: Greenblatt (2021)

The question worth asking is could that timeline be shortened with earlier exposure to business training? Could some of that consolidation have played out differently if physicians had understood the game while they were still being trained?

The outcomes, when physicians do engage, are significant. Parambath and colleagues found that more than 25% of the 105 healthcare companies that reached unicorn status ($1B+ valuation) between 2015–2024 had at least one physician founder—collectively valued at $68.9 billion. Among nearly 3,000 U.S.-based venture-backed unicorn founders, ~4% had medical degrees. More than 25 health systems had launched their own venture divisions as of 2023.

Informally, the gap is being filled. Communities like Scrub Capital—a VC firm backed by hundreds of physician investors and entrepreneurs—and MDplus, a network for business-minded medical students and trainees, have emerged to provide the experiential learning that training programs never offered.

What's more significant is who's paying attention now: the largest physician organization in the country.

Dashevsky’s Dissection

I’ll be the first to tell you how important any business knowledge is for physicians. My entrepreneurship exposure before medical school, and during training with Healthcare Huddle, has given me a different perspective on the healthcare system and why so much dysfunction exists.

The American Medical Association thinks so, too.

Dr. John Whyte—the AMA’s CEO and Executive Vice President—is about one year into the role, and is already pushing physician entrepreneurship and career growth beyond clinical medicine. His background is noteworthy, coming through clinical medicine, federal health policy, and corporate roles including CMO of WebMD. He's not an organizer but he is someone who's lived at the intersection.

Dr. Whyte and the AMA team are hosting their first AMA Physician Entrepreneur Forum in Chicago August 7th through August 8th. It’s meant to be intimate—a small, application-based cohort of 100 attendees—designed for physicians, fellows, residents, and med students who want to explore business leadership pathways, whether founding a company, joining a venture, or simply understanding the business dimensions of healthcare.

The agenda warrants a closer look because it's covering exactly what medical school never did:

• How to evaluate and pursue entrepreneurial opportunities: from a practical framework for assessing whether a business idea is worth building, to a candid panel on venture-backed entrepreneurship that addresses what founders usually get wrong about raising capital.

• The full spectrum of ownership and financing models: VC and PE, bootstrapped/capital-light businesses, and practice ownership are all on the table, so whether you want to start a company or just own your own practice, there's a track for it.

• Regulatory and risk realities: a dedicated session on the legal, operational, and compliance risks physicians need to anticipate as they grow, which is the stuff no one talks about until it's too late.

• Pitch feedback labs: pre-selected participants get real-time feedback on their ideas from experienced physician entrepreneurs and investors across multiple sessions focused on clarity, scalability, and investor readiness.

• Where physician entrepreneurship is headed: the forum closes with a moderated roundtable on the major trends likely to shape physician ownership and healthcare innovation over the next five years.

I wish something like this had existed when I was figuring out what "medicine plus something else" could even look like. The forum won't close the systemic gap overnight but it's exactly the kind of signal that matters. The largest physician organization in the country is finally saying out loud what a lot of us have been saying in our group chats for years: clinical expertise alone doesn't cut it anymore, and the system has a responsibility to do something about that.

In summary, physicians are already shaping billion-dollar companies and founding businesses at remarkable rates—yet fewer than 1 in 12 medical schools had a formal program just a decade ago, and most students today still graduate without a single business course. The demand exists, the interest exists, the need exists. The system just hasn't caught up. That’s why the AMA Physician Entrepreneur Forum is a significant step in the right direction.

Space is capped at 100. If this is the gap you've been feeling, save your spot.

Wearable Tech in 2026: What Every Physician Should Know

We’re living in the wild world of wearables (WWW). I’m writing this article with my Garmin on my left wrist, WHOOP on my right wrist, an iPhone in my pocket, and a Dexcom Stelo en route to my apartment—which I’ll place behind my left arm. I’m basically getting hospital-level monitoring at my desk. (Joking, but it feels like it.)

From Oura Ring to Apple Watch, Dexcom Stelo to Abbott Lingo, the market is saturated with wearable tech. The wearable space is at an inflection point, and it's worth explaining why now: three in five U.S. adults own a wearable (up 33 percentage points since 2015), AI is being layered on top of sensor data to turn raw biometrics into actionable insights, CMS is actively piloting reimbursement for wearable-supported care via the ACCESS Model, and peer-reviewed trials are validating what these devices can do clinically. When technology, policy, clinical evidence, and financial incentives all move in the same direction at the same time—that's what an inflection point looks like.

In this healthcare trends report, I'll break down the wearable tech space—key players, the consumer vs. medical-grade distinction, and where I think this is all headed.

Wearable Tech: What Is It?

Wearable tech spans more ground than most people realize: smartwatches tracking heart rate and blood pressure, rings monitoring sleep architecture, patches reading interstitial glucose in real time, even earrings measuring body temperature. The form factors keep expanding, but the underlying goal is the same—continuous, passive data collection from your body.

Consumers (emphasis on consumers, not patients) use them across a wide spectrum—from passive tracking to active optimization. Some barely glance at their data; others structure their entire day around recovery scores, strain metrics, and HRV trends (me). That spectrum matters clinically, because it shapes how consumers interpret—and act on—the data they bring into your exam room.

Who’s Wearing Wearable Technology?

Rock Health just published their 2025 consumer adoption survey, finding that three out of five U.S. adults own at least one wearable or connected device. That’s up 33 percentage points compared to 2015. Smartwatches continue to dominate the space (looking at you, Apple Watch).

Source: Rock Health

According to the survey, wearable users tend to be younger, wealthier, and healthier. The mix is roughly even between men and women. They are also more likely to be commercially insured. Smart rings, specifically, are gaining significant traction among millennials (we account for 59% of owners!).

One stat we should all pay attention to is how often people are actually wearing these devices. Eighty-three percent of wearable owners wear their device five or more days per week, and nearly 60% of surveyed consumers are always or nearly always wearing it. WHOOP's motto is "always on," and the product design reinforces that (charging is portable).

The equity gap in that adoption data deserves a closer look. Among millennials, ownership sits at 71%. Among the Silent Generation, it's 37%. High-income households (>$200k) own wearables at 67%. Low-income households (<$25k) own them at 41%. Commercially insured patients: 69%. Uninsured: 38%. Urban consumers own wearables at 64% vs. 48% in rural areas. We can't talk about wearables improving population health without acknowledging that the people most likely to benefit from continuous monitoring are the least likely to own these devices.

One more data point that I can corroborate from both being a patient and a clinician: 59% of wearable owners have discussed their data with a healthcare provider. Thirty percent do it regularly. Another 20% want to but haven't. Only 17% have no interest at all. As physicians, we're going to see more wearable data in our exam rooms whether we're prepared for it or not.

Target Markets

As a loyal wearable tech user since I bought my Garmin Forerunner 110 in 2010, and now a WHOOP owner since 2021, and a physician who thinks about wearables and data every day, I group the consumer market into three buckets. This framework helps me think about the impact of wearable tech.

1. Performance Optimization: these are your athletes and extremely health-conscious consumers who want to optimize their health. They focus on sleep, diet, and vigorous activity. They care about “longevity”. WHOOP, Garmin, and Oura target these markets.

2. Lifestyle: these are your average consumers who care about their health but aren’t as focused on optimizing every aspect of their data (HR, HRV, sleep, “strain”, “recovery.”). Step count and resting heart rate are enough.

3. Detection: these are your consumers and patients who use wearables to essentially “watch” them (no pun intended). They’re monitoring heart rate to detect abnormal rhythms, gait, sound levels, and falls. I’ll talk more about this below, but this is like your “guardian angel” technology. Apple Watch falls into this category, although it also fits into the other two categories above.

Key Players

The wearable market isn’t equally distributed. A handful of players are driving most of the innovation and most of the clinical implications. Mapping them onto the three-bucket framework above—Performance Optimization, Lifestyle, and Detection—makes it easier to see who matters to us as physicians, and why.

• Apple Watch: Most of your patients already own one. It’s the only consumer wearable that meaningfully spans all three buckets (fitness tracking for the optimizer, step counts for the lifestyle user, and AFib or fall detection for the patient you’re concerned about). That breadth, combined with more FDA clearances than any other consumer wearable on the market, is what separates it clinically from the rest of the pack.

• Oura Ring: They’ve raised $875 million at an $11 billion valuation, acquired a CGM startup (Veri), brought in Galen AI for EHR integration, launched Oura Ring 5, and are heading toward an IPO. Started squarely in the Performance Optimization bucket; now pushing hard toward clinical integration.

• WHOOP: Built for the Performance Optimization user. There’s no screen, no display, just continuous biometric data piped into a clean app, worn on the wrist, bicep, or in undergarments. Last raise: $575 million at a $10.1 billion valuation. They recently hired physicians to offer in-app video consults, a meaningful step toward the Detection bucket and something closer to care delivery.

• Fitbit (Google): Fitbit Air recently dropped, no subscription, $99, and a direct challenge to WHOOP’s model. Google wants to own the health data aggregation layer across Android’s global user base.

• Garmin: The OG endurance wearable that most people sleep on. Q1 2026: $1.75B in revenue, fitness segment up 42% year-over-year. Flat hardware fee, no subscription, one of the most loyal user bases in the space.

• Dexcom Stelo / Abbott Lingo: CGMs marketed directly to non-diabetics. We've ordered CGMs for patients with diabetes for years. Stelo and Lingo are asking a different question: what does continuous glucose data tell us about everyone else? The clinical jury is still out, but these two sit at the clearest intersection of consumer wearable and medical-grade device in the entire space.

• Samsung Galaxy Watch: The most underrated player here. Blood pressure monitoring (cleared in select markets), ECG, sleep apnea detection, body composition analysis—all on one of the most widely distributed hardware platforms in the world. Android's market share gives Samsung a quiet scale advantage that rarely gets the credit it deserves.

Consumer vs Medical Grade

I emphasize “consumers” because the line between consumer wearable tech for consumers and medical wearable tech for patients is getting increasingly blurry. As a consumer reading this, you may not care about that distinction. As a physician reading this, you have to.

We love data, especially when we’re making clinical decisions, but the data is only useful if it’s trustworthy. The most practical way to judge that is whether it’s been reviewed by the FDA under three different pathways:

• 510(k) Premarket Clearance: This is the most common pathway, accounting for roughly 85% of FDA-authorized medical devices. The 510(k) process requires companies to demonstrate their device is "substantially equivalent" to an already-approved predicate device.

• Premarket Approval (PMA): PMA is the most rigorous pathway, reserved for high-risk devices where no substantially equivalent device exists or safety concerns require comprehensive clinical data.

• De Novo Classification: This pathway exists for novel devices that don't fit existing classifications and lack appropriate predicate devices for 510(k) clearance.

The best way to understand the distinction is to read my prior content, Apple Watch vs. WHOOP: The Battle for Blood Pressure Tracking.

• The TL;DR: Apple Watch and WHOOP both launched blood pressure features.

• Apple Watch notifies you if you have high blood pressure, but it won’t give you an actual reading. Apple received 510(k) premarket clearance for this feature.

• WHOOP estimates your blood pressure using a ML algorithm. WHOOP did not seek any FDA clearance or approval for this feature.

• In the summer of 2025, the FDA sent WHOOP a strongly worded letter stating that because the blood pressure feature provides actual blood pressure readings, it’s considered a medical device that requires FDA clearance or approval.

• WHOOP responded by saying this feature is a “wellness” or “healthy lifestyle” tool rather than diagnostic. (Like tracking steps.)

• The FDA said, “wrong,” arguing that providing blood pressure estimates with color-coded indicators (green, yellow, orange) clearly serves a diagnostic function.

• The debate continues, but it seems like WHOOP is in the clear (see below “Policy”)

All of this to say, a device that is cleared or approved by the FDA is worth a physician’s attention if a patient brings in Apple Watch data suggesting hypertension, but that may not be true for devices without FDA clearance or approval.

Wearable Tech Tailwinds

Conditions are ripe for wearable tech right now.

1. Policy

CMS is testing the ACCESS (Advancing Chronic Care with Effective Scalable Solutions) Model to bring more wearables to the Medicare population. It’s a new payment option that allows clinicians to offer innovative, technology-supported care to improve patients’ health and complement traditional care. WHOOP was one of the companies chosen. Additionally, the FDA has loosened, or more accurately “revised,” what a wearable intended for wellness is allowed to say and claim. This works well for WHOOP because it allows WHOOP’s blood pressure insight feature to remain, as long as it does not claim it can diagnose hypertension or treat it.

2. Artificial Intelligence

Every major wearable platform has layered AI on top of its sensor data in the last two years. WHOOP has an AI Coach. Oura has an AI Advisor. Hims & Hers Labs is running an AI care agent that reasons across 130+ biomarkers. The shift is from raw data to interpreted insights. Instead of seeing your HRV drop, you’re told what it means and what to do about it. Apple remains one of the few holdouts without a native AI coaching layer, which says as much about their broader AI strategy as it does about their wearables roadmap.

3. Longevity

Longevity has become a cultural phenomenon, and wearables are central to it. Devices in the “Performance Optimization” bucket (WHOOP, Oura, Garmin) were already the preferred tools for people who track every biomarker with intention. The next evolution is combining biometrics with biomarkers: direct-to-consumer lab testing is now integrating with wearable data. WHOOP offers lab panels. Oura offers lab panels. The data picture is expanding well beyond heart rate and sleep.

Wearable companies are also pushing into care delivery. Oura just signed a partnership with Counsel Health:

4. Clinical Evidence Is Building

The Apple Watch asthma study—a randomized controlled trial backed by Apple and Elevance—showed that app-based wearable management led to meaningfully improved asthma control scores, especially among Medicaid patients. WHOOP published data from 30,000 members showing sustained reductions in alcohol consumption over 72 weeks. Studies like these are important because they give physicians clinical permission to engage with wearable data rather than dismiss it. The evidence base is growing, and as it does, the gap between “consumer device” and “clinically relevant tool” continues to narrow.

My Vision: Guardian Angel Tech

Guardian Angel Tech is my vision for where the wearable space is headed. Everyone will own some wearable that is always analyzing vitals and movement—a “guardian angel watching you”—and can intervene early to prevent bad outcomes.

Quick examples off the top of my head:

• You have heart failure with reduced ejection fraction. Your Withings smart scale trends a 5 lb increase in weight in two days. Apple Watch detects a drop in baseline O2 saturation. Guardian Angel Tech tells you to seek medical attention for an impending heart failure exacerbation.

• Your iPhone and Apple Watch detect a trend toward fewer stairs climbed, a rising resting heart rate, decreasing heart rate variability, and a slower walking pace. It suggests seeing a cardiologist for an echocardiogram. Pulmonary hypertension is detected earlier rather than later.

• Your Apple Watch detects AFib. It then detects an abnormal gait. It tells you to go to the emergency department to rule out embolic stroke.

• Your Apple Watch detects an abnormal rhythm. You have Kardia and obtain a 6-lead EKG using just your fingers. It shows atrial fibrillation with rapid ventricular response. You call your cardiologist. They tell you to go to the emergency department ASAP.

It's like the light that turns on in your car mirror when you want to change lanes, but there's a car in your blind spot. The wearable is catching things before things go south!

Dashevsky Dissection

Wearables are exciting, and I say that as someone with four of them on or near his body right now. But excitement doesn't mean we should turn our brains off.

Here's my clinical take: when a patient brings me their wearable data, the first question I ask myself isn't "what does this show?" but instead "how was this measured, and has the FDA looked at it?" This distinction matters a bit more than most of us are taught to think about. I do NOT think wearable data should integrate into the EHR unless it's been FDA cleared or approved. WHOOP's blood pressure insights? Not in my chart. My patient's Omron readings—FDA-cleared—absolutely. That line is important since it's the difference between data I can act on and data I have to take with a grain of salt.

The other thing: the people who would benefit most from this technology are the least likely to have it. I covered the equity gaps above and noted that lower-income, uninsured, rural, and older patients own wearables at significantly lower rates. If wearables become a meaningful part of how we monitor chronic disease, that gap has consequences. We can't close that loop without talking about access.

That said, I'm optimistic. The guardian angel vision—a wearable that's always on, always analyzing, and actually intervenes when something's wrong—is closer than it's ever been. AI makes it more plausible every year. I just want us to get there with the same rigor we'd apply to anything else we use to make clinical decisions.

I've lost count of how many times I've been at the bedside of a critically ill patient with no clear surrogate, or a healthcare proxy form that was never filled out. These situations force physicians into ethically complicated decisions under pressure, and they're almost entirely preventable.

The problem traces back to a system that treats advance care planning as a checkbox, not a clinical priority. By the time a patient lands in the ICU, the window has already closed.

In this week's Inefficiency Insights, I walk through why this keeps happening and the workflow I've built to get ahead of it.

Yesterday on rounds, I watched someone use a clinical AI tool in the middle of a discussion and I felt proud. Part of me thought, look how far this has come. Then I caught myself: are we sure we should be doing this?

Clinical AI is everywhere now, and the default behavior is to treat the output like a consult. The problem is that these tools can sound certain even when they’re guessing, omitting something important, or applying the right guideline to the wrong patient—like reflexively recommending antibiotics for asymptomatic bacteriuria—and we’re the ones responsible for what happens next.

In this Deep Dive, I’ll lay out what “trust” should mean at the point of care, break down how UpToDate stress-tests Expert AI, and show how I’m using it in practice.

What physicians actually mean by “trust”

Trust, for me as a clinician, means two things when it comes to clinical AI tools:

1. I can verify what I’m reading quickly.

2. The system fails safely, or at least transparently, when it’s uncertain.

Trust doesn’t necessarily mean the tool is always correct. It means I can audit it fast enough to use it responsibly at the point of care.

When I try a clinical AI tool, I’m looking for basic behaviors: does it surface the assumptions it’s making, does it flag where the evidence is thin, and does it point me back to the source quickly enough that I can sanity-check it before I act?

Here’s a quick example. I’m taking care of a kidney transplant patient on immunosuppression who has asymptomatic bacteriuria on a routine urinalysis. I ask: Do I treat this? What I’m looking for is whether the tool surfaces the key assumptions (time since transplant, symptoms, upcoming urologic procedure), links me back to the source, and flags uncertainty instead of sounding falsely definitive. I’m not looking for a yes/no.

Tools that sound confident when they’re guessing train clinicians into complacency. That’s why I care less about benchmark scores and more about how a tool behaves in real clinical context. UpToDate’s latest whitepaper is an attempt to measure that directly.

How Wolters Kluwer Stress-Tests and Validates UpToDate Expert AI

Wolter Kluwers generative AI tool, UpToDate Expert AI, lets physicians interact with UpToDate’s curated clinical content through natural-language queries. As I wrote in my last piece on UpToDate Expert AI:

It’s fast. But can I trust it?

UpToDate is well aware that the actual output UpToDate Expert AI produces—the words—can greatly influence clinical decisions. So, they want to ensure UpToDate Expert AI is reliable. And to do that, they stress-test the heck out of it.

In their latest white paper A Measured Approach to Evaluating AI at the Point of Care, instead of asking, “Can Expert AI pass a test like USMLE?”, they ask, “Is the answer UpToDate Expert AI produces clinically useful, grounded in trusted knowledge, and safe when things get messy?”⁠ To evaluate Expert AI’s answers, they group them into three different buckets:

1. Clinical intent: is the answer faithful to their point-of-care standards?

2. Knowledge integrity: is the answer grounded in trusted clinical knowledge?

3. Potential risks: how does the system behave under stress, uncertainty, and adversarial use?

This framework is what I wish more companies would lead with. It’s certainly closer to how we actually practice. I break it down further in the next section.

1) Clinical intent: rubrics, not vibes

UpToDate Expert AI answers were evaluated and validated against physician-authored rubrics. These rubrics, built by UpToDate physician editors across 25 specialties, rate what an “ideal” answer should contain for a given question. In other words, what a “good” answer looks like. Take the example question “when to anticoagulate for atrial fibrillation?” Below is the physician-authored rubric:

Source: A Measured Approach to Evaluating AI at the Point of Care

The rubric breaks a free-text answer into independently scored elements: what’s present, what’s missing, what’s incorrect, and what details matter at the point of care. The more interesting part is omissions. Rubric testing forces you to look at what didn’t show up in the answer. And omissions are where clinical tools quietly harm people. A model can look polished and still miss the one sentence that changes the plan. Omissions are what get you paged at 2 AM. Check out my video below to see what this actually looks like:

Three results stand out in the white paper:

1. UpToDate Expert AI provided clinically aligned information for 99.9% of assessed criteria (1,669 queries; 15,000+ criteria).

2. It met 13–15% more Essential criteria than two general-purpose LLM comparators.

3. Those general-purpose comparators had omission rates about 15% higher, translating to one additional omission error for every seven queries.

This is also how you monitor drift. If a model starts getting worse as it iterates, rubrics catch it.

2) Knowledge integrity: “model knowledge” leakage (this is the x-factor)

UpToDate Expert AI is designed to derive answers from UpToDate content. The system prompt specifies that only UpToDate content can be used, it uses a multi-step retrieval process, and it’s designed not to respond if relevant UpToDate content can’t be found. If content in the answer is not derived from UpToDate’s trusted content, this is called model knowledge leakage. Sometimes that leakage can be correct (like interpreting a weird abbreviation), which is exactly why it’s dangerous. Plausible isn’t the same as grounded.

In UpToDate’s analysis, they estimated model knowledge use to be similar to their test’s background noise of 1–4%. This shows UpToDate Expert AI is focused on already trusted, curated content within UpToDate.

The through line is verifiability. UpToDate Expert AI is designed to derive answers from UpToDate content, and the system is built so you can trace an output back to the underlying source material. That’s what makes the tool usable at the point of care: I can click, confirm, and move. This is the ‘verify quickly’ part of trust.

3) Potential risks: red teaming the failure modes we worry about

A dedicated team of clinical AI specialists and domain experts—the Red Team—was assigned the job to stress test UpToDate Expert AI by throwing it:

• Complex clinical scenarios

• Purposely jumbling how a question is structured

• Multi-turn conversations

• Adversarial attempts to provoke an undesirable answer

This team has put in 200+ hours and tried 1000+ times to get UpToDate Expert AI to produce undesirable answers. And when it does, they codify known risks into rubrics for ongoing surveillance testing.

They concluded that harm is best assessed by experts, and they’re developing an expert audit process to catch meaningful errors that automated scoring (e.g. testing against multiple choice questions) can miss.

Benchmarks and ratings, though, aren’t worthless. They’re just incomplete. Standardized exams, vignettes, and user ratings can give signals, but they weren’t designed to determine whether AI-generated content is appropriate at the point of care.⁠ Third-party validation, too, is essential. No one should be grading their own homework, and UpToDate agrees.

With that framework in mind, here’s how I think about UpToDate Expert AI on the wards—and what earns trust fast enough to use at the point of care.

Dashevsky Dissection

UpToDate Expert AI stands out as a system built around a model:

• Curated knowledge base

• Outputs you can trace back to source material

• Evaluation that matches how clinicians use answers in the moment

At the bedside, that combination matters more than raw cleverness. I’m making decisions under time pressure, and I need an answer I can audit fast enough to safely act on.

I treat clinical AI the way I treat peers on my team in the hospital. The most trusted team members move quickly and surface uncertainty early. They show their work. They name what they’re assuming. They make it easy for you to check them before you sign the order. That’s the bar I want these tools held to, and it’s essentially what UpToDate is operationalizing with rubrics, knowledge-integrity checks, red teaming, and a feedback loop back to the underlying content.

Here’s my simple rounds test:

• Does it broaden the differential instead of anchoring me?

• Does it surface uncertainty instead of smoothing it over?

• Can I click straight to the source and verify in seconds?

When a tool consistently passes those checks, it earns a place in my workflow.

In summary, clinical AI earns trust at the bedside when we can verify the reasoning in seconds, the tool shows its sources and assumptions, and it communicates uncertainty instead of smoothing it over, because at the point of care, “good enough” answers that sound definitive are exactly how omissions and misapplied guidelines slip into real decisions.

15‑second Take (TL;DR)

Zocdoc is a middleman for healthcare scheduling because it bundles four broken layers we still haven’t fixed: provider discovery, directory and insurance normalization, appointment inventory access, and conversion tooling.

It makes booking feel consumer-grade by aggregating high-intent patient demand and integrating (or syncing) into messy EHR scheduling so availability is actually bookable.

The business model follows the incentives: per-new-patient booking fees, sponsored placements, infrastructure/API products, and enterprise partnerships, which pushes the platform to optimize completed bookings more than clinical appropriateness.

Net: patients get speed and convenience, while physicians and health systems gain acquisition and backfill, but everyone risks more fragmentation and growing dependency on the marketplace toll.

Why Zocdoc Exists (and Why It’s So Hard to Replace)

I had to schedule a primary care appointment. Where did I go? Zocdoc.

I could have dug through a health system website for recommendations, but those interfaces feel archaic. A lot of systems now embed Zocdoc into their own sites to make scheduling easier, which made me ask the obvious question: is Zocdoc a middleman?

Zocdoc and similar companies are scheduling marketplaces. They exist because healthcare still behaves like a pre-internet consumer market: fragmented, full of unreliable directories, high-friction booking, and expensive demand generation for clinicians and health systems. Middlemen like Zocdoc showed up to bundle the inefficiencies around scheduling.

This includes:

• Provider discovery and reputation

• Directory + insurance normalization

• Appointment inventory access

• Conversion tooling (reminders, re-booking, waitlists, etc.)

But should a scheduling click require a venture-backed company to translate insurance networks, directory data, and EHR availability into something a patient can use?

Those four layers are what Zocdoc bundles and controls to make scheduling seamless for patients. Zocdoc is where patients book doctors. But it’s also where demand aggregation happens. Patients search by specialty, location, insurance, availability, and ratings. The platform captures high-intent traffic and routes it to bookable supply. When I searched for a PCP in my area, it surfaced a list of local physicians who accept my insurance, which is exactly what patients want.

But determining who is actually in-network is its own problem. I wrote about Ghost Networks several months ago. Insurers do not reliably update provider directories. A physician in Ohio can stay listed even after closing shop and moving to Texas. Patients end up trying to schedule with a ghost (spooky). Scheduling platforms can become a better source of truth for who is really in network. They answer the questions:

• Do you take my insurance?

• Are you accepting new patients?

• Where are you located?

• Can I get an appointment soon?

Then there’s appointment inventory access, which is the hard part. A directory is easy. A list of open slots is not. These marketplaces only feel magical when you can actually book—right now—without picking up the phone.

To pull that off, they need a bridge into the scheduling system. Either they integrate directly with an EHR’s calendar, or they build a sync layer that can read availability and write back confirmed appointments across a messy mix of scheduling tools. Zocdoc’s Sync framework is their bet on being that interoperability layer.

Once booking works, everything downstream becomes conversion: reminders, rescheduling, cancellation handling, and backfilling last-minute gaps. Waitlists fit here too. It’s the same hidden inventory problem—just packaged into a consumer-grade flow.

Where They Sit in the Food Chain

These platforms don’t fit a neat middleman sandwich. They sit on top of payers, directories, and EHR scheduling and translate the whole stack into a booking flow:

• Payers: insurance networks and benefit design define affordability and “in-network” status.

• Provider orgs: own the appointment supply and clinical protocols.

• EHR scheduling modules: store the inventory and rules.

• Directories/search: are scattered across payers, Google, health system sites, and review sites.

These scheduling marketplaces basically say, “We will translate this fragmented mess into a consumer-grade booking experience and charge providers for the privilege of access to our demand.” It’s a clean consumer experience. That’s the point. And that’s why it’s powerful. Once a platform owns discovery, it also controls ranking—and ranking becomes an ad product. At that moment, Zocdoc is selling demand with attribution—new patients you can measure. Scheduling turns into performance marketing, and the price drifts toward whatever a specialty can afford to pay per acquired patient.

Origin Story

These marketplaces show up when:

1. inventory is fragmented

2. the booking mechanism is locked behind phone calls and proprietary EHR scheduling.

Zocdoc was a first mover in the space. They were founded in 2007 with the pitch that finding an in-network doctor with availability is painful. It still is, but it’s a little less painful now. The hardest parts for these marketplaces have been integration, distribution, and unit economics.

Once you solve booking at scale, the business becomes a toll on new-patient demand.

How the Money Flows

There are four ways scheduling marketplaces make money.

1. Per-new-patient booking fee (charged at booking; incentives)

2. Sponsored placement (bidding on top of booking fee; ranking becomes pay-to-play)

3. Infrastructure/API (sell the plumbing; defend against disintermediation)

4. Enterprise partnerships (distribution + credibility + sometimes direct revenue)

Impact Analysis

• Patients: The obvious upside for patients is much less friction in finding an in-network physician and scheduling an appointment. You book faster. You can often see someone in days, not weeks, compared to going through a health system website, calling around, or relying on word of mouth. There is also clearer transparency into availability.

  The downside is that the marketplace optimizes for booking completion, not for steering patients to the right level of care at the right time. Sponsored placements and advertising can also make patients confuse a top result with clinical quality, when it is often just paid placement. Zooming out, if a patient relies on Zocdoc to book across clinicians who all use different electronic health records, care can become even more fragmented than staying within a single health system. This is probably why Zocdoc partners with health systems.

• Physicians and Health Systems: Scheduling marketplaces give us a strong top-of-funnel for new patient acquisition, help fill cancellations, and smooth demand variability. Over time, they can increase revenue. The risk is dependency: practices can end up paying a toll to access demand.

  Another dynamic is that these platforms list you directly alongside competitors. If you pay for ads, you might show up higher, but the alternatives are still one click away, and patients may choose someone else.

Could We Live Without Them?

We can live without Zocdoc the company. We can’t live without what it does until payers, health systems, and EHRs make scheduling inventory truly interoperable and directories reliably accurate. This means addressing ghost networks and inadequate physician supply.

In summary, scheduling marketplaces exist because the system hasn’t earned the right to be self-serve. Until directories are accurate and appointment inventory is interoperable, this middleman stays valuable—and expensive.

Read more from The Middlemen Series here.

How My Use of AI Will Change as I Become an Attending

I’ve been thinking about how my relationship with AI will change over the next decade.

Not in the abstract “AI is the future” way, but in the way that matters in the hospital, like when do I reach for it, what do I trust it with, and what happens when I’m no longer a trainee who can say, “I’m still learning”?

Right now, I use AI for two things: learning medicine faster and getting the administrative stuff out of my way. This mix feels good for training, but I don’t think it holds as I become an attending.

The Workflow Part

There are tasks I do today that are technically “doctor work,” but functionally clerical.

• Discharge summaries

• Transfer-of-care notes

• Inbox messages

• Form letters

• Prior authorization letters

These tasks require accuracy and attention, but they don’t require my best thinking.

When I use AI well, it doesn’t replace me, but it gives me a first draft so I can spend my attention on things that matter like patients at the bedside, my team, and the fun, nuanced decisions that aren’t obvious.

The Clinical Knowledge Part

The other way I use AI right now is the part people love to debate: clinical knowledge and reasoning.

As a resident, I’m still building my internal library. AI helps me search faster, synthesize faster, and test myself in ways I can’t always get in real time.

I’ll take a case and ask the model to function like an attending and poke holes in my assessment and plan. I’ll have it turn patient presentations into board-style questions. I’ll use it to walk me through PFT interpretation and force me to explain each step.

The question is what happens when I’ve seen enough patients that I’m not asking “what is this” nearly as often? Because the goal of training is that the fundamentals become automatic.

I think the kinds of questions I ask will shift, not disappear. As I move into pulmonary and critical care medicine on July 1st, and get more specialized, the bread-and-butter gets boring. The edge cases get interesting.

My future AI use is probably less “teach me pneumonia” and more “help me think through the weird drug interaction I haven’t seen in two years,” or “what’s the nuance in the guideline footnote that matters for this exact patient.” Fewer questions but higher-level questions.

Trusting Myself vs The Model

There’s another factor that changes as you become an attending: responsibility.

As a resident, I’m the frontline clinician. I see the patient more than anyone else on the team (including the attending). As an attending, though, the buck stops with me—no matter what. I’m responsible for what the residents do, what the interns do, what the notes say, what the plan is, and how it all connects. That responsibility changes how you think about tools.

I don’t want AI to be the thing that does my thinking for me. I want it to be the thing that makes my thinking sharper.

And there’s a kind of clinical intuition that comes from repetition that AI can’t replicate.

I’m thinking about the patient with severe right ventricular dysfunction who’s now hypotensive. The room has a feel to it. You’re weighing physiology, trajectory, what’s about to happen next, and what you’re willing to try in the next five minutes (actually, probably seconds).

You can ask an AI model for a differential and management options. That can help. It can also distract you if it’s noisy, overconfident, or generic. The seasoned physician’s job is pattern recognition plus judgment.

AI has the pattern library. It doesn’t have the judgment that comes from owning outcomes. (Feel free to quote this).

This Is Important: Signal-to-noise

Here’s what I suspect will separate “useful AI” from “annoying AI” as I get more experienced: how quickly it gets to the point.

Early in training, more context feels helpful. I want the explainer and I want the background. Later, I’ll want the answer, the caveat, and the one thing that would change my plan.

Some tools will keep producing long, fluffy, “nice-to-know” responses. Seasoned clinicians won’t have patience for that (let me know if I’m wrong, seasoned physicians).

In summary, what I want from AI is simple. I want AI to take the low-value work off my plate and make my high-value work better. I want it to help me learn faster now, and help me practice cleaner later. And I want it to free up enough attention that I can become the kind of doctor I’m trying to become: decisive, and human at the bedside.

We order too many labs. I'll include myself in that. Routine panels drawn without a specific clinical question behind them, and suddenly the inbox is flooded with borderline values that are almost certainly insignificant — but still require a response. Every single one. And somewhere buried in all of that noise is the result that actually matters. A biopsy. A high-risk HPV. A final pathology report. Text-based results that don't auto-flag, don't trigger alerts, and just quietly land in a pile.

That's the problem I dig into in this week's Inefficiency Insights — why physician inboxes are set up to miss the things that matter most, and how I'm using AI right now to close the loop faster without the cognitive overhead.

Paperwork — I hate it. A couple of weeks ago in clinic, a patient brought in routing forms for home health services that they'd had filled out before. Because of state requirements, the form needs to be renewed every year. Typically, the admin staff handles renewals. That keeps the visit focused and the chart clean.

But the admin team was so short-staffed that these forms landed on my desk. And when physicians are filling out paperwork, it's a signal that the system has run out of bandwidth to protect clinical time.

There are a lot of things to complain about in medicine, but the bureaucracy is what I'll complain about until I die. Spending 20 minutes on chart review, gathering patient information, and tracking down diagnostic codes is time that could be spent on actual clinical care.

I filled out the form, signed it, and handed it over… only to be told I'd done it incorrectly. I don't have high blood pressure, but it was certainly elevated after that. So I filled it out again, this time making sure my handwriting was actually legible. That time, it cleared.

This is the paperwork problem in miniature: redundant processes, fragmented accountability, and physicians absorbing the overflow. We haven't fixed it (yet). We've just decided doctors are the catch-all.

Root Cause Analysis: 5 Whys

The problem: A physician was pulled away from clinical care to complete routine administrative paperwork.

1. Why?: The admin team was too short-staffed to manage routine paperwork tasks that typically fall outside physician scope.

2. Why?: Healthcare organizations chronically underinvest in administrative staffing relative to the volume of regulatory and documentation requirements they face.

3. Why?: Administrative labor is treated as overhead to be minimized rather than as infrastructure that protects clinical capacity.

4. Why?: Reimbursement models reward clinical volume, not operational efficiency—so there's no financial signal that admin under-resourcing is costing anything.

5. Why (root cause)?: The fee-for-service model creates no incentive to protect physician time from non-clinical tasks, so when administrative systems break down, the physician becomes the default catch-all, absorbing the overflow at the highest cost per hour in the building.

Impact Analysis

• Patient: Patients are burning a visit just to hand in forms, which can mean time off work, transportation, and a copay, even though the forms could have been handled remotely (e.g., short telehealth visit) or routed through a portal. When an in-person visit is not clinically necessary, requiring one for paperwork is a failure of process design, not a care decision.

• Clinician or Provider: Clinical time gets swallowed by tasks that have nothing to do with clinical judgment. Every minute spent hunting down diagnostic codes or re-filling a form incorrectly is a minute not spent at the bedside. It accelerates burnout, erodes morale, and turns physicians into the most expensive administrative assistants in the building.

• System: We're already paying for admin staff but chronic understaffing means the work doesn't disappear, it just shifts to physicians, which is both inefficient and expensive. The infrastructure exists to do this better, and AI is increasingly being deployed to do exactly that (see below). The will to implement them is the variable.

Solution

There are two potential solutions.

1. Add more human resources to handle administrative work, so physicians can focus on patient care.

2. If that staffing is unattainable, whether because of cost or limited supply, use generative AI.

I’ll focus on the generative AI path.

First, many parts of the form can be streamlined before the visit. Every form has a patient information section (name, DOB, SSN, family history, etc.). In my experience, patients leave these sections blank in about 99% of cases, which pushes the work onto admin staff and physicians. It seems minor, but across dozens or hundreds of forms per week, it adds up.

We can get ahead of this with automating patient intake or form builders. These tools can be patient-facing and physician/admin-facing. On the patient side, the patient enters the basic information required for the form. On the physician/admin side, the team adds the diagnostic codes and medical decision making needed. Once both inputs are complete, the form can be auto-filled. This beats the current workflow of using PDFs, adding text boxes, typing in fields, and copy-pasting boxes for future forms.

A lot of prior auth AI products already work this way, pulling data from the EHR to complete prior auth forms. The core capabilities are auto-populating forms from existing chart data, flagging incomplete documentation before submission, and routing routine renewals without a physician ever touching them. These tools should be more widely available. Epic should ship a native feature that does this.

But they don’t have one quite yet, or at least I’m not aware of it.

In the meantime, I’ve been bootstrapping with clinical AI tools like Doximity. For the form scenario I described above, I uploaded the form to Doximity, provided a prompt, and then uploaded patient information (e.g., prior visits).

You can watch the video below to see how I do it in real time.

As I mentioned in the video, if you have any cool use cases with generative AI tools, let me know and maybe we can work on something together.

This method, of course, is not perfect. But the cognitive load (chart review, diagnostic code lookup, clinical summarization) is largely handled. At that point, the remaining work is clerical, not clinical.

The broader point is this is a workaround, not a solution. We shouldn't need to bootstrap this. These capabilities should be native to Epic or any major EHR. Until they are, tools like Doximity give us a way to claw back some of that time.

In the same quarter that the women's health menopause space achieved its first unicorn (≥$1B valuation), the NIH started scanning new grant applications for the word "gender." Private capital is hitting record highs while federal research funding and reproductive policy are moving the other way.

The women’s health market is maturing past fertility apps into menopause, primary care, oncology, and biopharma. At the same time, abortion bans are bending OB/GYN, primary care, and internal medicine training pipelines and worsening maternal outcomes in states that already had the worst numbers.

A lot moving in opposite directions at once.

As I do every May, here's the update on the women’s health space, touching on VC funding, policy, research, and what comes next.

Women’s Health Venture Capital Funding

Historically, women’s health funding has been concentrated in reproductive and maternal health, since these are conditions that uniquely affect women.

However, there are major areas of unmet need where women carry a higher burden and prevalence, including:

• Cardiovascular disease

• Osteoporosis

• Menopause

• Alzheimer’s disease

Boston Consulting Group estimates that closing the gap on these four areas alone could unlock a $100B+ market by 2030. And yet, despite women being half the population, women's health still receives only 4–6% of total healthcare VC dollars.

Silicon Valley Bank's Innovation in Women's Health 2025 report found women's health funding more than tripled from 2019 to 2024, outpacing healthcare overall. However, funding fell—or recalibrated—in 2025, to $1.5B, around where it was in 2023.

Here's how the past three years stack up:

Hottest Sub-markets

Five sub-markets are pulling the most attention from investors and acquirers right now:

1. Maternal and women's primary care platforms. Value-based, virtual-first, payer-contracted models expanding past maternity into the full female lifecycle. Pomelo Care, for example, raised a $92M Series C at a $1.7B valuation in January 2026. The expansion plan moves from maternity into pediatrics, hormonal health, perimenopause, and preventive care. Pomelo now supports about 7% of U.S. births. The thesis is that a value-based, virtual-first, payer-contracted model can both improve outcomes and prove ROI — and right now it's the most-funded thesis in the category.

2. Fertility and reproductive health. Moving past consumer cycle tracking into therapeutics, IVF lab automation, and ecosystem benefit deals. Carrot Fertility, for example, has raised $116M and signed three big partnerships in late 2025: BCBS Global Solutions plus Oura and Dexcom. Those deals are how Carrot is building a moat — fertility benefit + ring + CGM as one stack. Underneath, the category itself is shifting from consumer apps into therapeutics, lab automation, and diagnostics (ReproNovo, Gameto, Conceivable Life Sciences, Inito).

3. Women's oncology. PwC names oncology as the second-fastest-growing core women's health segment after menopause through 2030.

4. Cardio-metabolic health. Heart disease is the leading cause of death in women, and the category is finally drawing female-specific dollars.

5. Autoimmune health. Women carry roughly 80% of the U.S. autoimmune disease burden, and biopharma is leaning in.

Hottest Companies

Below I highlight some of the hottest companies in the press, whether it be for general news, expansion, partnerships, or fundraising.

fundraising.

The common thread: companies getting the most coverage are the ones with either a category-defining unicorn moment (Midi, Pomelo), a category-resetting M&A (Hologic), or a credible AI/automation angle (Gameto, Overture, Midi).

U.S. Policy on Women's Health and Research Funding

The private market is being asked to absorb work that the federal research apparatus used to do, and there's a limit to how far that can go. NIH-funded longitudinal cohort studies—the Women's Health Initiative being the canonical example—are not the kind of science venture capital steps in to fund.

The 2024 Biden buildup put real money on the table:

• An executive order created the Fund for Women's Health Research at NIH, with more than $200M proposed for FY2025 and a long-term $12B ask to Congress.

• ARPA-H launched the Sprint for Women's Health in February 2024 with $100M committed— the first White House initiative dedicated to women's health R&D.

• The Gates Foundation, Novo Nordisk Foundation, and Wellcome pledged $300M over three years for global health equity, including women's health.

• Bipartisan menopause bills advanced, including the Menopause Research and Equity Act and the Advancing Menopause Care and Mid-Life Women's Health Act.

The second Trump administration has reversed most of that posture. The NIH canceled hundreds of grants covering reproductive health, fibroids, vaccine safety in pregnancy, and autoimmune disease. The CDC was instructed to remove terms like "pregnant person" from internal reports. In April 2025, HHS announced it would terminate the Women's Health Initiative Regional Center contracts and the Clinical Coordinating Center. After significant public backlash, HHS reversed course 24 hours later. The reversal stuck, but the researchers I follow describe a lasting chill on what they will actually propose.

Per April 2026 New York Times reporting, NIH is no longer canceling grants en masse. They’re now using a computational text analysis tool that flags applications mentioning terms like "gender" before approval. NIH grant spending is roughly $1B behind historical pace, per a Washington Post analysis, and women, cancer, and mental health have absorbed the steepest drop in new awards. In February 2026, reports surfaced that the administration was winding down ARPA-H's Investor Catalyst Hub in Cambridge—home of the Sprint for Women's Health infrastructure.

Net effect: federal funding for women's health research is flat to down in real dollars, narrower in scope, and politically volatile. Private capital can supplement that work, but it cannot replace it.

IVF serves as a great case study for what strong funding and scientific innovation can do for women’s health, by driving demand and building policy support. As highlighted in the World Economic Forum, IVF—a once-stigmatized experiment—is now a multibillion-dollar industry. The graph below tracks IVF investment and deal count from 1978 through 2025 alongside the major scientific breakthroughs (sorry, you may have to enlarge the image, I tried my best!).

Four Years Post Dobbs

We're nearly four years out from Dobbs. Thirteen states enforce total abortion bans, and another 28 have gestational-age limits of varying severity.

Total U.S. abortions have actually risen post-Dobbs. In 2023, the country saw an 11% increase compared to 2020, driven largely by telehealth medication abortion. Medication abortion accounted for roughly 63% of all abortions in the U.S., with many delivered via telehealth and supported by shield laws in 22 states and D.C.

That mail-based access is now in active jeopardy:

• October 2025: Louisiana sued the FDA to roll back the 2023 REMS revision that allowed mifepristone to be mailed and dispensed at retail pharmacies.

• May 1, 2026: the Fifth Circuit sided with Louisiana and reinstated the in-person dispensing requirement nationwide, effective immediately.

• May 4, 2026: Justice Alito issued a one-week administrative stay through May 11 while Danco and GenBioPro's emergency appeal sits with the Supreme Court.

If SCOTUS lets the Fifth Circuit ruling stand, mail-based mifepristone collapses overnight in every state. Providers will likely fall back to misoprostol-only regimens, which are relatively less effective effective compared to ~99% for the combination protocol, with more pain, bleeding, and GI side effects.

The FDA's posture complicates things. Under pressure from state attorneys general and anti-abortion groups, the Trump-era FDA announced in September 2025 that it was conducting a comprehensive review of the 2023 REMS, and the Fifth Circuit cited that announcement as the agency conceding its earlier decision was procedurally flawed.

States aren't waiting on SCOTUS. Mississippi's HB1613, effective July 1, 2026, makes manufacturing, distributing, dispensing, or prescribing abortion medication a felony punishable by up to 10 years. Louisiana has reclassified mifepristone and misoprostol as controlled substances. Texas's HB7 lets private citizens sue providers and mailers. Whichever way SCOTUS rules this month, the federal scaffolding around medication abortion looks a lot less stable than it did six months ago.

Impact on Trainees

The OB/GYN training pipeline is bending around the abortion ban map. In the first post-Dobbs match cycle (2022 to 2023), OB/GYN applications to programs in ban states dropped about 10.5%, double the decline in states without bans. The 2023–2024 cycle widened the gap to 6.7% in ban states versus +0.4% in legal states, and internal medicine applications in ban states fell more than five times faster than in legal states.

Training quality has eroded alongside the application drop. ACOG’s 2025 position statement found that about 1 in 6 OB/GYN residency programs has lost local abortion training, even though ACGME still requires it.

Idaho already lost roughly 22% of its practicing obstetricians between August 2022 and November 2023. Since 57% of physicians practice in the state where they trained, today's application gap is tomorrow's maternity-care desert. Ban-state programs are still filling their slots because the U.S. match has more applicants than seats overall, which masks a meaningful decline in program competitiveness.

For internists, hospitalists, and EM physicians in ban states, this is your future consult bench shrinking. The same states with the worst maternal mortality numbers are losing the OB/GYN, MFM, and family medicine OB colleagues you'd call from the wards or the ED.

Source: McEachern JE, Traylor TA, Roman D. Change in Number of OB/GYN Physicians Practicing Obstetrics After the Dobbs Decision. JAMA Netw Open. 2025;8(7):e2524893. doi:10.1001/jamanetworkopen.2025.24893

Dashevsky’s Dissection

As a resident, I read this story two ways at the same time. The VC numbers tell me women’s health is finally a category investors take seriously, and the products/services my future patients will use are getting funded faster than ever. The policy story tells me the federal infrastructure that generates the evidence base behind that care is contracting at the same time. Both are true, and they’re happening to the same patients.

VC dollars cannot fund a Women’s Health Initiative. Thirty-year cohort studies, registries, long-tail safety data on pregnancy and autoimmune disease—none of that has a five-year exit. As NIH narrows what it will support and ARPA-H’s Investor Catalyst Hub winds down, the evidence pipeline behind future women’s health products gets thinner while the products themselves get richer.

From the IM side of the wards (and my future critical care ICU), the trainee data is what I keep coming back to. I’ll be admitting OB/GYN patients, co-managing peripartum cardiomyopathy, and calling MFM for consults for my whole career. A 22% drop in Idaho’s practicing obstetricians shows up on my pager as a future consult that goes unanswered. Residents who choose programs outside ban states are also choosing to avoid the legal ambiguity around miscarriage management, ectopic care, and a malpractice posture none of us were trained to navigate.

Three things I’m watching over the next 90 days:

• SCOTUS on Louisiana v. FDA. Whichever way the Court rules, the FDA will look like an unstable regulator on this drug class for the rest of the year.

• FY2027 NIH appropriations. That’s the cycle where the gender-flag text-screening posture either becomes permanent or gets walked back.

• Late-stage women’s health rounds. If Series B and late-C deals don’t unfreeze in 2026 H2, the Q1 unicorn pop was a top, not a turn.

In summary, women’s health is getting its biggest private-capital tailwind to date, while the federal research and regulatory infrastructure behind women’s care is contracting and getting more politically volatile. For us clinicians, that combination shows up as faster product adoption with a thinner evidence base, and a training pipeline that is already bending around abortion policy and will reshape who can practice where over the next decade.

Most of us treat precharting as a quality-of-care chore—a quick scan of labs, specialist notes, and the last visit before walking into the exam room. The CMS 2027 Medicare Advantage rule changes the stakes entirely. Starting in 2027, MA plans can no longer rely on retrospective chart reviews to capture diagnoses missed during the encounter. Diagnostic capture has to happen prospectively, in the room, in real time.

That shifts a lot of weight onto the few minutes before each visit. Every chronic condition, lab trend, and specialist update we fail to surface in the prep doesn't just affect the visit—it affects the risk-adjusted payment.

So how do you preround a full panel without spending an hour on chart synthesis, and still catch what now drives both clinical outcomes and revenue?

You've seen the patient. Written the H&P. Worked through the assessment and plan. And yet — there's still that nagging feeling. Did I miss something? Is there a diagnosis on the differential I didn't think to consider? A piece of the workup that could change the picture?

Most of us have been using generative AI for the easy stuff — drafting prior auth letters, writing discharge summaries, cleaning up patient notes. Useful, but indirect. I've been doing something different over the past couple of months, and it's the first time I've felt like AI is actually strengthening my clinical reasoning in real time, at the bedside, on complex patients.

Full breakdown in my latest Huddle+ article here.

A few weeks ago, I watched a colleague pull up ChatGPT mid-rounds to look something up. My instinct was immediate: why are you using that? Use Open Evidence. Use Doximity. Use UpToDate. Those tools were built for us! ChatGPT was not.

Now OpenAI has launched ChatGPT for Clinicians, which is a free, verified platform aimed squarely at physicians, NPs, PAs, and pharmacists.

I’m not surprised by the launch of ChatGPT for Clinicians. AI medical search is becoming a commodity. The same way ambient scribes went from novelty to baseline expectation over the course of my residency, AI-powered clinical knowledge tools are heading the same direction. If you don't offer it, you're already behind.

So, in this article, I'll recap what ChatGPT for Clinicians actually is, map where it fits in an increasingly crowded market, and explain what it takes to win in a space that's quickly turning into table stakes.

What is ChatGPT for Clinicians?

ChatGPT for Clinicians is free for verified U.S. clinicians (physicians, NPs, PAs, and pharmacists) and takes about five minutes to set up via NPI verification. It offers citation-backed clinical search, a deep research mode across the medical literature, repeatable documentation workflows (prior auth letters, patient instructions), and CME credit tied to evidence review. HIPAA support is available, but only through a business associate agreement—it's not on by default (big FYI!).

This is OpenAI's third healthcare product in rapid succession. ChatGPT Health launched in early 2026 for consumers—patients linking medical records and getting personalized answers. ChatGPT for Healthcare is the enterprise B2B version, with governed security controls. ChatGPT for Clinicians fills the individual clinician layer. They're building a full healthcare stack, one user segment at a time.

genAI depiction of OpenAI health stack.

AI Medical Knowledge Tools Are Becoming a Commodity

When a feature becomes so widely available that its presence is expected and its absence is a liability, it has commoditized. That's where AI medical search is right now.

Two years ago (nearly the entirety of my residency) platforms like Open Evidence and DoxGPT were novel. A lot of us were experimenting with them, and the concept still felt new. Today, nearly every major clinical platform has layered in some version of AI medical search. When Heidi launched Heidi Evidence earlier this year, it barely registered as a surprise because of course they did. And now OpenAI. The market has saturated, and the message is quite clear that if you don't offer this, you're already behind.

For completeness, I’m going to copy and paste my brief market analysis of the key players below. I also recorded a quick video where I take ChatGPT for Clinicians, Doximity, OpenEvidence, and UpToDate’s Expert AI, then copy and paste the same clinical question into each to see how they present and structure the evidence.

Market

• UpToDate Expert AI (Wolters Kluwer): The longstanding gold standard in clinical decision support, now layering generative AI on top of its physician-reviewed content. Its moat is trust, EHR integration, and the fact that most of us already have it open during rounds. I wrote a deep dive on Expert AI if you want the full breakdown.

• Doximity + Pathway Medical: Doximity acquired Pathway Medical to build out DoxGPT, which bundles clinical Q&A, scribe, and documentation tools into one physician-facing platform. Its advantage is distribution—Doximity reaches over 80% of U.S. physicians. Check out my deep dive here.

• Open Evidence: The fastest-growing platform in this space. Open Evidence went from a $1 billion valuation in February 2025 to $12 billion by January 2026, raising roughly $700 million across four rounds in under a year. It has content deals with NEJM and JAMA, and its focus on medicine-only answers with primary research citations has made it a daily driver for many physicians, myself included.

• Heidi Evidence: Heidi’s scribe platform now includes a free, citation-backed medical search tool.

• ClinicalKey AI (Elsevier): Pairs Elsevier's premium medical content with generative AI, often bundled into enterprise deals with health systems.

• DynaMed / Dyna AI (EBSCO): An evidence-curated database with AI-enhanced natural language queries and EHR integrations. Competes directly with UpToDate in the institutional market.

• AMBOSS: Expanding beyond medical education into clinical decision support with an AI "medical co-pilot." Strong editorial credibility and a loyal user base among trainees and early-career physicians. I’ve been experimenting with this platform some more!

The broader AI in healthcare market was estimated at $36.7 billion in 2025 and is projected to top $500 billion by 2033 (39% CAGR), with healthcare AI spending nearly tripling in 2025 to $1.4 billion. Within that wave, AI clinical decision support (a decent proxy for AI medical search) reached roughly $2.8 billion in 2025 and is projected to hit $15.3 billion by 2033.

Dashevsky’s Dissection

When I first saw OpenAI’s announcement, my knee-jerk reaction was, "Great, another player crowding the medical AI search space." That reaction is part of what makes AI medical search feel like a commodity now.

OpenAI knows the market is saturated, so this move feels strategic. They can collect clinician identity data, see what we search, and learn how we search. That can translate into targeted ads, product strategy, or both. Making it free also turns it into a front door: get clinicians using the tool, anchor habits, and then lean on those clinicians to push their organizations toward the enterprise version of ChatGPT for Healthcare. And ChatGPT already has massive reach among Americans, including physician-users, so the distribution advantage is obvious.

As I’ve stated before in my course on AI and Healthcare Utilization (enroll here), in a saturated market, platforms win on three things: distribution, trust, and frictionless workflows.

On distribution, OpenAI is hard to beat. It was the first generative AI platform I used. I no longer use it regularly, but the point stands: it’s everywhere, on everyone’s phone and computer.

Trust is a different question. Up until now, if I saw a colleague using ChatGPT to look up medical information, I’d give them a slap on the back of the head, telling them to use Open Evidence, Doximity, or UpToDate instead (see my opening sentence). Those products are built for clinicians, rely on primary sources, and (in UpToDate’s case) sit on top of content written and reviewed by experts. I do not trust OpenAI the same way. I’m especially skeptical after recent reporting in The New York Times and The New Yorker on the company’s history. And when I say “trust,” I mean trust with health information. I’m not convinced, even with promises that queries will not be used to train models. If OpenAI eventually goes public, the pressure to maximize shareholder value will shape what happens next.

Then there's workflow—and this is where I think people underestimate the switching cost.

"Frictionless" doesn't require EHR integration per se. It just requires that a tool already lives where physicians are. I open Doximity multiple times a day: to dial a patient, send a message, look something up, draft a prior auth letter. The clinical search is already there, already trusted, already one tap away. This is a habit, now, and habits in clinical medicine are sticky by necessity. We are trained to build them because consistency under pressure matters.

For ChatGPT to displace that, it would need to offer something meaningfully better at a moment when I already have a tool open. Not marginally better… truly meaningfully better. The bar to interrupt an established workflow is high, and OpenAI hasn't cleared it for me—at least not yet. What could change that? If it proved more accurate on specific clinical question types, if it integrated into a platform I already use, or if an institution mandated it. Short of that, most physicians will keep doing what they already do. Inertia is underrated as a market force in healthcare.

In summary, ChatGPT for Clinicians matters less because it is a breakthrough product and more because it signals where the market is headed: AI medical search is becoming table stakes, and the winners will be the platforms that earn clinician trust, minimize workflow friction, and control distribution as this capability turns into a baseline expectation.

15‑second Take (TL;DR)

Healthcare has thousands of AI tools competing for attention, and no one has the time (or expertise) to sort through them. A new kind of middleman is coming: the AI broker. Just like insurance brokers navigate plans for beneficiaries and PBMs navigate formularies for payers, AI brokers will navigate the vendor landscape for healthcare purchasers. We don't fully have them yet. But we will.

Where Americans Get Care in 2026—And Why It Matters

I’m a pretty healthy person, and I try to see my primary care provider once a year. What’s wild is that I have to schedule next year’s annual visit as I’m walking out of this year’s appointment. I have access to great healthcare, and still have to book that far out. If I need a sooner visit for something urgent, good luck to me.

Over the past week, a few things have stood out to me:

1. A new CDC study showing where U.S. adults are getting their healthcare.

2. A LinkedIn post by Howard Wilson, MD, showing that the number of urgent care centers in the U.S. has surpassed the per-capita density of primary care physicians.

3. At the same time, I’ve been working with Lumeris, which is using AI (via their agentic AI tool called, Tom) to help expand primary care panels.

It’s all coalescing at once, and I think it’s worth talking about.

In this article, I highlight the latest CDC healthcare data, map primary care trends, and explain why we need to increase both physician supply and panel capacity at the same time.

The Deets: Where Do People Get Their Healthcare?

Researchers at the CDC released a report analyzing sources of usual health care among adults age 18 and older, by sex and age group, for 2024. I encourage you to read it yourself, but I’ll highlight the results I found most interesting.

First, 90% of adults reported having a usual source of care, meaning a place they can turn to when they get sick or need care. That’s reassuring. Still, 10% didn’t have a “usual place” to go if they became ill, so I’m assuming their default is the emergency department (though they may not have reported it).

Second, there was a clear age disparity in who uses a doctor’s office or health center for usual care. Among 18–34 year olds, only 68% reported a doctor’s office or health center as their usual source of care, compared with 89% of adults 65 and older. This increased with age.

Source: CDC

On the flip side, this pattern was reversed for adults who reported an urgent care center or a clinic in a drug store (e.g. CVS Health) or grocery store (e.g. ACME) as their usual source of care, starting at 12.2% for 18–34 year olds and decreasing to 4% for adults 65 and older.

Source: CDC

As an extension to the above, Dr. Howard Wilson, posted the below, showing how urgent care centers have nearly tripled since 2010 while PCP density dropped 22%.

Broader Primary Care Trends

There are several key trends in primary care that I want to quickly mention, which may help explain the above:

1. Persistent and Worsening Workforce Shortages: The U.S. faces a projected shortfall of up to 70,610 primary care physicians by 2038, compounded by record burnout rates (62.8% of physicians in 2021) and an aging physician workforce nearing retirement.

2. Maldistribution: Urban-Rural Divide: Rural areas average just 5 primary care physicians per 10,000 people compared to 8 primary care physicians 10,000 people in urban areas.

3. Decline in Per-Capita Primary Care Visits: Primary care visits among commercially insured patients fell 7% from 2018 to 2024, and for the first time, behavioral health visits surpassed primary care visits in 2024 (66.4M vs. 62.8M).

4. Shrinking Panel Sizes: Traditional PCP panels range from 1,500–2,500 patients, but there's a growing movement toward smaller panels in direct primary care and concierge models to combat burnout, despite the access trade-offs.

5. Rapid Expansion of NPs and PAs: Ambulatory care visits (not primary care visits, specifically) to NPs, PAs, and RNs nearly doubled (+98%) from 2010 to 2021, driven by expanded scope-of-practice laws, lower training costs, and rising physician shortages.

6. Shift Toward Employment and Consolidation: Nearly 80% of U.S. physicians are now employed by or affiliated with hospital systems or other corporate entity (e.g. retailer, private equity), largely driven by reimbursement pressures and administrative burden. See my latest article on this here, and my latest class on physician consolidation here.

7. Evolving Care Models (e.g., Direct Primary Care): The number of concierge and direct primary care practices grew 83% between 2018 and 2023. Read my latest, The Future of Healthcare: Why Subscription Medicine Is Taking Over.

So, primary care access is shifting fast, and whether we shore up the workforce or scale new models, the next few years will define how—and where—patients get care. On to my next section….

Dashevsky's Dissection

I'm biased, of course, but as both a physician who practices primary care and a patient who needs primary care, I see firsthand how imperative it is. The population is aging, we're living longer, and we'll be living with more chronic disease, which means someone needs to help manage it all—and that someone is the primary care physician. Urgent care, at the end of the day, cannot do that and shouldn't try—unless it is staffed by physicians trained in primary care or internal medicine. It shouldn't be another specialty (e.g. emergency medicine) trying to fill that longitudinal role.

If a lot of young people are using urgent care for their "urgent" needs, that could actually be a smart top-of-funnel opportunity to connect them with a primary care clinician. That's likely one reason health systems have launched their own urgent cares: to funnel patients into their primary care and specialty practices. But if you do a real root-cause analysis of the primary care problem, it comes down to supply. There simply aren't enough primary care physicians. We can try to incentivize more medical students to go into primary care, but that isn't happening. Opening up more slots for IMGs is another potential solution. Neither will close the gap fast enough—and we need it to happen now (more like yesterday).

So if we can't quickly scale the number of PCPs, the only realistic lever is helping existing PCPs do more with the same hours (which is what Tom, by Lumeris, is built for). Tom is an agentic AI. It continuously monitors a physician's entire patient panel, identifies the highest-impact next action for each patient, and then does proactive outreach, care gap closure, post-discharge follow-up, chronic disease management. It’s all autonomous and it can certainly help scale PCP panels.

This is all to say we'll need both solutions—increasing PCP supply and scaling agentic AI to expand panel capacity—to happen in parallel. But supply won't move fast enough on its own. The near-term answer is making what we have go further.

The VA has been trying to replace its 1990s EHR for nearly a decade. Four attempts. $16 billion spent. 96% of its hospitals still on the old system.

In this article, as I rotate through a VA hospital for the last time, I'm breaking down why the most powerful healthcare system in the country can't seem to modernize the one tool its clinicians use every single day—and what it means now that AI is changing everything.

How Ambient AI Scribes Are Changing Medical Coding Intensity

On the surface, AI scribes are meant to reduce the time we physicians spend typing in the EHR so we can focus more on patients. More or less, they are doing that.

But a second benefit is starting to get more airtime. Depending on who you ask, it is either the most exciting thing about these tools or the most troubling:

• Ambient scribes are great at coding accuracy.

In this article, I look at how AI scribes are changing coding accuracy and why that matters for physicians.

The Deets: Coding Accuracy

When we see patients, we document what we are treating. We do not naturally think in ICD-10 codes in the exam room (although, maybe some of us do). If a patient walks in with a chronic cough, mild hyponatremia, and a recently changed diuretic, we manage all three, but we might only formally code for one.

Clinical Documentation Integrity (CDI) teams exist because of this gap. Their job is to chase down every billable condition after the fact and make sure it gets captured. It is labor-intensive, inconsistent, and structurally slow. AI can close that gap in real time with ambient scribes.

Ambient scribes listen to the full encounter, capturing every diagnosis mentioned and every condition discussed. Some platforms go further by layering on retrospective coding AI that scans the full chart and surfaces additional billable diagnoses a physician referenced but never formally listed. The ROI numbers from early adopters are significant. One ambient vendor markets roughly $13,000 per clinician annually in recovered revenue.

From a pure documentation standpoint, this is an improvement in coding accuracy, since we were likely undercoding before and AI is correcting that.

Trilliant Health’s Data on Coding After AI Scribe Adoption

A recent analysis from Trilliant Health examined outpatient E/M billing patterns at six large health systems that have publicly adopted ambient AI scribing, using national all-payer claims data from 2018 to 2024. Across every system, visits shifted toward higher-intensity codes for both new and established patients.

For new patient visits, the share billed at the highest acuity levels (CPT 99204–99205) rose by 12 to 20 percentage points across all six systems (orange line). One health system saw 80% of new patient visits billed at high intensity by 2024. For established patients, the increase ranged from 7 to 12 percentage points. These changes were consistent across geographically and organizationally distinct institutions, all of which adopted ambient AI during the study period.

Trilliant Health

The increase appeared across nearly every diagnosis category. For factors influencing health status, respiratory disease, and mental and behavioral disorders, coding intensity rose across the board.

Trilliant's interpretation is measured. They argue the increase likely reflects better rules-based documentation rather than fraud. Ambient AI records every word spoken. A tool that captures everything clinically relevant will naturally generate more complete, higher-acuity documentation than a physician typing notes in Epic at 11 PM. And with the 2021 E/M coding revisions that shifted emphasis toward medical decision-making and total time, more encounters now legitimately qualify for higher codes.

Trilliant is also candid about the limits of what its analysis can show:

• No control group of health systems that did not adopt ambient AI, which means we can't isolate the technology's contribution from broader secular trends in coding intensity.

• It can't distinguish between documentation that accurately reflects clinical complexity and documentation that overstates it. That distinction between accurate coding and upcoding is precisely what's in dispute, and the data alone can't answer it.

A single outpatient visit coded one level higher might mean $10–40 more in reimbursement. Across millions of visits, that compounds fast.

Blue Cross Blue Shield’s Study

In my initial article, How AI Documentation Tools Are Making Upcoding Worse, I covered what Blue Cross Blue Shield's research arm found when it examined the same question across 62 million commercial members.

Their analysis flagged postpartum anemia as a case study: coding for that diagnosis tripled at the highest-growth hospitals between 2022 and 2025, a surrogate for AI adoption, while transfusion rates, the standard treatment for clinically significant postpartum anemia, stayed essentially flat. One BCBS plan audited a major outlier hospital system and found that fewer than 20% of cases coded with postpartum anemia actually met established clinical criteria.

BCBS

BCBS estimated that coding intensity shifts in maternity admissions alone added roughly $22 million in spending over the study period.

The AI tools are not making clinical decisions, but they are doing exactly what they were designed to do: capture everything (Deming’s Principle!). In a fee-for-service system that rewards documentation volume, "capturing everything" and "maximizing revenue" can end up being the same thing.

Dashevsky's Dissection

We're likely to see a cat-and-mouse game, as Bryan Vartabedian, MD mentioned in a comment on my LinkedIn post. These AI scribes and CDI platforms will capture more diagnostic codes, which will increase revenue cycle management activity. Insurers, recognizing that they will have to reimburse more, will probably use their own AI tools to downcode or deny care through prior authorizations. In effect, it becomes AI versus AI. Here is how this could affect patients, physicians, and health insurers.

Patients may not feel this directly at first. But as coding intensity rises and insurers pay out more, those costs will be passed along through higher premiums and, for patients on high-deductible plans, higher out-of-pocket costs tied to higher-coded visits. If payers respond by tightening prior authorization, patients will feel that too through delays, denials, and care fragmentation. The premium increase is the visible part. The prior auth response is where care gets disrupted.

For physicians, provider groups, and health systems, more accurate coding should mean better reimbursement for the services we actually provide. I want to be clear that this is meaningfully different from what we see in Medicare Advantage, where upcoding has been used deliberately to inflate risk scores and capture government dollars without a corresponding increase in care delivered. With ambient AI, the argument is that we're correcting a longstanding undercoding problem and finally getting paid for conditions we were already treating. Up until now, this has been labor-intensive and inconsistently applied. Coding that happens prospectively, during the encounter, is more accurate and more defensible than a CDI team reviewing charts two weeks later.

But here is where it gets complicated for us as individual physicians. Most of us are not seeing the final coded output. The ambient scribe captures the encounter, the coding AI scans the chart, and a claim goes out, often without the physician reviewing what was actually billed. While we signed off on the note, we didn't necessarily sign off on the code. The False Claims Act does not distinguish between intentional fraud and negligent overcoding. If an audit finds that your name is attached to claims that don't meet documentation criteria, the liability question becomes very real very fast. This is something physician groups and hospital legal teams need to get ahead of.

Even when coding is clinically accurate, payers may still deny or downcode claims. A payer's definition of "clinically justified" does not automatically align with what an ambient AI tool documented. So physicians and health systems could find themselves in a position where they coded more completely and correctly than ever before and still face aggressive pushback from insurers who are running their own AI against those claims.

When it comes to payers and CMS, if costs rise, they will offset them through higher premiums, tighter prior authorizations, and more aggressive claim audits. The BCBS analysis I covered earlier is already evidence of this in motion, although that report isn't neutral analysis.

One thing Trilliant noted in its conclusion is that ambient AI creates a time-stamped record of every encounter, which now makes it technically possible to audit whether a code was justified in a way that wasn't feasible before. So it is a tool for accountability and transparency, but it also means there is a paper trail that payers can weaponize in audits that simply didn't exist when we were typing notes after the visit.

In every prior tech arms race in healthcare, from Medicare Advantage risk adjustment to prior authorization automation, the arms race itself became a cost center, and none of it improved care. We're watching the same dynamic start to play out here, just faster and at a larger scale.

Can Apple Watch Can Detect a Fatal Lung Disease Early?

Idiopathic pulmonary arterial hypertension (IPAH) is a mouthful to say, but it’s worth talking about because we diagnose it late.

Briefly, here’s idiopathic pulmonary arterial hypertension in plain terms:

• The lungs have their own blood supply.

• Blood from the heart travels through the pulmonary arteries to pick up oxygen, then returns to the rest of the body.

• In IPAH, those arteries progressively stiffen and narrow for no clear reason.

• The heart has to work harder and harder to push blood through them.

• Over time, it fails.

There is no cure. Treatments can slow the disease, but they work best when started early.

As hinted at above, most patients do not get diagnosed until they are already sick. The average time from first symptoms to formal diagnosis is about three years. By then, the disease has often progressed significantly. Around 75% or more are in advanced functional decline (NYHA functional class III or IV) when first diagnosed, meaning they struggle to walk across a room.

We are not diagnosing IPAH slowly because it’s invisible. The initial symptoms of shortness of breath, fatigue, and slowing down overlap with a hundred other explanations, like deconditioning, heart failure, COPD/asthma. So the disease hides in plain sight while we chase other diagnoses.

That inefficiency costs lives. Below I do a root cause and impact analysis, then highlight how Apple Watch, and the broader “guardian angel tech” can help detect this disease earlier.

Root Cause Analysis: 5 Whys

The problem: IPAH is routinely diagnosed 2–3 years after symptom onset, by which point many patients have advanced disease and worse outcomes.

1. Why? Symptoms (shortness of breath, fatigue, reduced exercise tolerance) are vague and overlap with dozens of other conditions.

2. Why? Current screening relies on patients presenting to a clinician, describing symptoms, and triggering a specific diagnostic workup. There's no passive early warning system.

3. Why? Formal diagnosis requires an invasive procedure called a right heart catheterization where a catheter is threaded into the heart to measure pressures directly. Physicians are reluctant to order this until there's strong suspicion.

4. Why? The early functional decline (the gradual slowing down) happens outside the clinic, in daily life, where no one is watching.

5. Why (root cause)? We've built a diagnostic system that depends on patients and clinicians recognizing a pattern that unfolds over years, with no continuous data stream to catch it early.

• Patient: By the time most IPAH patients are diagnosed, they've lost years of optimal treatment window. Earlier diagnosis means earlier therapy, better functional capacity, and longer survival. Delayed diagnosis means sicker patients with higher likelihood of mortality.

• Clinician: We're working with snapshots of a patient. That is, a clinic visit every few months. Between those visits, we have no visibility into how a patient is actually functioning day to day. This is a significant blind spot for a disease that progresses continuously.

• System: IPAH's prevalence has more than doubled over the past 15 years. It's rare but expensive, including hospitalizations, advanced therapies, transplant evaluations. Earlier detection changes the trajectory and the cost.

Solution

A research team from Imperial College London, Stanford, and several UK specialty centers asked a simple question:

• Can passively collected data from a smartphone or Apple Watch identify patients with IPAH before they're even diagnosed?

This was a pilot study of 109 participants across the UK, including 33 patients with confirmed IPAH, a disease control group (people with other serious conditions), and healthy volunteers. They used an app called My Heart Counts, originally developed at Stanford, paired with an Apple Watch Series 4.

What makes this study unusual is the timeline. Some participants had up to eight years of retrospective HealthKit data (step counts, walking pace, flights of stairs climbed, heart rate, heart rate variability, sleep) collected passively through their iPhones and watches before they were ever formally diagnosed.

Results

Here were the key findings of the study:

• IPAH patients walked slower.

• They climbed fewer stairs.

• Their heart rates were higher at rest and during activity.

• Their heart rate variability, which is a marker of how well the nervous system regulates the heart, was lower.

These differences were present before diagnosis.

A machine learning model trained on just the pre-diagnosis watch data distinguished IPAH patients from controls with an AUC of 0.87. This means the algorithm could correctly classify about 87% of cases using nothing but passive activity data collected before anyone knew the diagnosis.

Add in a simple lifestyle questionnaire, and that number climbs to 0.94.

Even using just an iPhone (no watch required) the model hit 0.91 with lifestyle questionnaire data. Phones are more accessible, more universally adopted, and still captured enough signal to do meaningful classification.

AUC

Interestingly enough, once patients were diagnosed and started treatment, their activity metrics improved. Step counts went up, heart rates came down toward normal.

Dashevsky’s Dissection

IPAH is rare, and the framework this study demonstrates is broadly applicable.

What we're talking about is passive digital phenotyping: using data you're already generating every day to detect early signals of serious disease. Your phone is already in your pocket, counting steps and logging walking pace, building a longitudinal record of physical function.

I've written before about what I call guardian angel technology. This is the idea that wearables, powered by machine learning, will eventually act as a silent monitor running in the background of your life. Always watching. Catching things that would otherwise go unnoticed. Not requiring you to do anything differently. Just generating signal while you go about your day.

This study is a proof of concept for that vision. The Apple Watch wasn't doing anything exotic. It was just counting steps, measuring heart rate, and tracking how fast someone climbed a flight of stairs. Ordinary data, analyzed longitudinally, drew a picture of disease developing in someone who did not yet have a diagnosis.

Guardian angel technology, applied to rare disease, could build a baseline, detect drift, and flag a pattern that no clinician could catch from a quarterly visit.

The three-year diagnostic delay in IPAH is not unique. We see it in pulmonary fibrosis, in heart failure, and in early-stage COPD. The common thread is slow, progressive functional decline that happens between clinic visits, invisible to the healthcare system until it's advanced. Guardian angel technology can close that gap, if we build the tools to use it.

There are limitations to the study, of course. This was a small pilot. The model trained on UK patients performed poorly when applied directly to US patients, because activity patterns and lifestyle differ across populations. The researchers had to retrain with some US data to get acceptable performance. That is a legitimate challenge for real-world deployment.

The signal also isn't perfectly specific. Slower walking pace and lower heart rate variability can show up in atrial fibrillation, heart failure, and musculoskeletal disease too. A clinically deployable screening tool would require bigger, more diverse datasets, plus integration with electronic health records.

Still, the direction is right. The foundational question—can a smartwatch catch a fatal disease before your doctor does?—has a preliminary answer: yes, it looks like it can.

The technology exists and the data is already being collected. The remaining work is building the infrastructure to use what's already in people's pockets.

In summary, IPAH is a progressive, fatal disease that we consistently diagnose too late. This pilot study shows that passive data from iPhones and Apple Watches (step counts, walking pace, heart rate patterns) can identify patients before formal diagnosis with impressive accuracy. This is a proof of concept, not a clinical tool yet. But it's a meaningful signal that continuous, passive monitoring could fundamentally change how we catch diseases that hide in the slow erosion of daily function.

Every physician has a prior authorization story. Mine involves a patient who waited two weeks for a medication I had already prescribed but they never got it. The drug was perfect for her—but insurance hadn't gotten around to approving it yet.

Prior authorization costs the U.S. healthcare system an estimated $35 billion a year in administrative burden alone, and the average physician spends 14 hours a week on it. That is a full workday (for a resident, at least). Every week.

And now it is changing fast.

CMS launched the WiSER Model—a six-year AI pilot that deploys automated PA reviews inside Traditional Medicare, and pays vendors a percentage of the savings from denied claims. Early data from Texas showed only 62% of requests approved. The Electronic Frontier Foundation has already sued CMS over it.

Meanwhile, both sides of the fight are arming up with AI. Insurers are using it to deny faster. Providers are using it to appeal faster. The algorithms argue with each other while patients wait. A $5.8 billion market has grown up around the whole mess.

I break down who hates prior auth, who loves it, why it exists in the first place, what WiSER actually changes, and what the AI arms race means for how we practice—all in my latest Huddle #Trends report.