I was recently asked to review a medical journal opinion piece on FDA guidance for wellness-device marketing. The piece made a careful, well-credentialed argument about the potential for patient harm from devices marketed without formal evidence review. The author was right about specific risks. Ten years ago, when I was squarely in academia, I’d have agreed with the conclusion almost reflexively.
My review commentary was different. My patients, the ones I now see in a direct-to-consumer virtual cardiology practice, value these devices. They see them filling a gap the formal system hasn’t filled. They aren’t naive about the limits; they understand the data is consumer-grade. They want it anyway, and increasingly they want an expert to help them interpret it. My reply, which I wouldn’t have written ten years ago, was that an opinion focused only on harm was missing the demand signal underneath.
But that exchange was only the visible tip of a pattern that’s now central to how my practice runs. A typical week in a virtual cardiology clinic includes patients with undiagnosed dysautonomia arriving with weeks of continuous heart-rate and blood-pressure data from a smartwatch and a home BP cuff, often the most clinically useful longitudinal record anyone has of their condition. It includes patients with palpitations sending PDF snippets of single-lead ECGs from an Apple Watch or KardiaMobile, sometimes captured during the symptoms a 30-day Holter would have spent a month trying to record. It includes patients who have ordered their own expansive lab panels from Function Health or comparable DTC services, full lipid subfractionation, ApoB, Lp(a), inflammatory markers, hormone panels, often paying more for the panel than they would have paid for a year of conventional copays. It includes patients who have had their own coronary calcium scan or CCTA done at an outpatient imaging center after their primary care declined to order it. And it increasingly includes patients arriving with their own genetic testing results, with questions about Lp(a) genetics, lipoprotein metabolism, or familial hypercholesterolemia.
Across modalities, the same pattern: patients are acquiring clinical data the formal system hasn’t ordered, paying for it themselves, increasingly running it through an LLM for an initial interpretation, and arriving at the visit wanting an expert whose judgment they trust to confirm, refine, or replace what the AI has already told them. The wellness-device conversation is one corner of this. The pattern is much broader, much more clinical, much more LLM-mediated than it was a year ago, and, for the moment, almost entirely unaddressed at scale.
Four groups in healthcare, patients, providers, payers, and the venture-funded companies built to serve them, are each rationally pursuing what they value. None of them, collectively, are building the infrastructure for what patients are now arriving with.
Providers
The providers I trained alongside, and the providers I still consult with as friends and colleagues, value something specific. They value hard mortality and hospitalization outcomes. They value evidence-graded recommendations. They value protection from acting on data that hasn’t been validated for clinical decision-making. They value reduction of medico-legal exposure from interpretation that gets ahead of guidelines.
There’s much that’s right about this. Most of what gets called “innovation” in patient-facing health tech does turn out to be noise, vendor-marketed dashboards, or outright snake oil. Academic medicine’s pace-of-adoption discipline has prevented a lot of harm. From radiation overuse to off-label drug misuse to “longevity protocols” that produce no measurable mortality benefit at substantial cost. The instinct to wait until the evidence is solid before changing practice isn’t paranoia. It’s the discipline that distinguishes evidence-based medicine from influencer cardiology.
The structural limitation is that the threshold for “validated” is set against the wrong loss function for the patient sitting in front of you with their own data. Information that a patient is willing to pay for, can correctly interpret with help, and is statistically associated with outcomes, even if not yet operationalized in a screening guideline, has positive value to that patient. Academic medicine’s training treats such information as a liability source rather than as an asset to be incorporated. The asymmetry hardens with seniority. The senior cardiologist who spent fifteen years working within a particular framework has internalized that framework as good clinical practice, and is least likely to be the person to question it.
I was trained in this frame. The training was largely correct, and I’m grateful for it. My mindset shifted in a way I didn’t expect when I began running a direct-to-consumer virtual cardiology practice. I started seeing what patients were buying, what they were doing with what they bought, and what they actually needed help with. The signal underneath their behavior wasn’t the signal I had been taught to read. (I wrote separately about what this looks like from a clinical-evolution standpoint, in The Perils of Standing Still.)
The cost of the misalignment is real. Patients who arrive with data and hypotheses get reframed as “the worried well” or “the anxious patient” rather than as data-equipped consumers. Their willingness to pay out of pocket, which in any other market is a strong signal of demand intensity, gets treated as evidence of confusion rather than as a market truth worth engaging.
Patients
The patients I see are optimizing for something specific too. They value longevity. They value self-understanding. They value control over their own bodies. They value not having ignored a knowable risk. They’re increasingly literate about lipid biology, genetics, cardiac imaging, and pharmacology in ways their parents’ generation wasn’t. They’re willing to spend on what they perceive as informational value even when the system declines to cover it.
The defining frame: each of them has one life, not a population’s worth of lives. A test that isn’t cost-effective for a screening program may still be entirely worth ordering for the individual considering whether they’re in the tail of the distribution.
The behavior isn’t consumerism. Patient-acquired data routinely reclassifies cardiovascular risk in ways the conventional markers don’t. A few patterns I see repeatedly in clinic.
A patient with an “acceptable” LDL of 105, no other major risk factors, would conventionally be told to continue lifestyle measures and check again in five years. The CAC score they paid for themselves comes back at 280, mostly in the LAD. The plaque was there. The conventional markers didn’t predict it. The patient who ordered the scan was right to.
A patient whose lipid panel reads broadly normal, LDL 115, HDL 55, triglycerides 90, wouldn’t be on a statin or any preventive therapy at age 42 by the conventional workup. The Lp(a) they ordered comes back at 240 nmol/L. By the actual risk picture, they’ve roughly twice the lifetime cardiovascular risk of someone with their other numbers and a normal Lp(a). That measurement, which most primary care visits wouldn’t have included, is the one that changes the next thirty years of their care.
A patient with an LDL of 95, under the conventional target, but ApoB of 120 mg/dL and a persistently elevated hs-CRP of 4.5 has an actual atherogenic particle count meaningfully higher than the cholesterol number suggests, and an inflammatory state independently amplifying risk. The patient who paid out of pocket for the expanded panel found a real signal the standard one missed.
These aren’t edge cases. They’re common findings. The reason they aren’t part of the standard workup isn’t that they’re wrong, and not that they’re clinically meaningless. It’s that at population scale, the additional yield doesn’t pass the cost-effectiveness threshold a screening program is judged against. That’s a defensible reason to write a screening guideline a certain way. It isn’t, however, a reason for the individual patient to ignore the information. The patient willing to absorb the cost of finding out where they actually sit in the distribution is doing exactly what rational risk-aversion at the individual level says they should do.
Their structural limitation is the inverse. They overweight the diagnostic value of any individual test result and underweight the interpretive expertise needed to translate it into action. The LLM-interpretation step that increasingly precedes the visit is sometimes excellent and sometimes confidently wrong, and the patient often lacks the expertise to know when. The frustration with mainstream medicine pushes them toward voices that are wrong-but-confident rather than measured-but-uncertain. The asymmetry favors influencers over careful clinicians for exactly the reasons influencer voice is built to win attention.
The cost of the misalignment: they’re paying for data they can’t fully use, then finding themselves alone with the data, then turning to AI assistants or influencer content for interpretation, then arriving at clinical visits that won’t engage with what they’ve assembled. They’re forming opinions about cardiovascular risk and treatment in an information ecosystem with thin quality control. And in each individual case the data they paid for may have contained a genuinely actionable signal that no one in the formal system was set up to receive.
Payers
The payers operate on a time horizon and an accountability structure that point in a different direction. They value predictable medical-loss ratios on quarterly and annual cycles. They value defensible coverage decisions that hold up to regulatory review. They value risk pooling that survives adverse selection. They value not paying for procedures whose downstream benefit accrues beyond the member’s expected tenure with the plan.
A surprising amount of expensive imaging, specialty referral, and pharmacy authorization in the U.S. is genuinely low-value or duplicative. The Choosing Wisely campaigns were correct about most of their categories. Payers who reflexively cover everything get adversely selected against and lose money, which removes them from the market entirely. The discipline that says “show me the evidence this changes outcomes” has prevented a lot of waste.
The structural limitation is that their time horizon, roughly 1-2 years per member, is misaligned with preventive cardiology’s payoff horizon of 10-30 years. They can’t capture the discounted value of avoiding a future MI that occurs after the member has switched plans, employers, or risk pools. So they systematically under-cover the highest-NPV preventive interventions, the ones whose benefit unambiguously exceeds cost over twenty years but loses money for them over five.
The cost of the misalignment: coverage decisions get made against the wrong time horizon. Members who would benefit most from preventive cardiology get told their elevated Lp(a) doesn’t warrant the workup that would clarify it. Members who pay out of pocket effectively cross-subsidize a system that won’t engage with the data they generated.
Investors
The venture capital that funds healthcare companies optimizes for capital efficiency, scalable unit economics, defensible business models, and clean return paths within a fund’s hold period. Sophisticated healthcare investors aren’t naive about the complexity of the underlying clinical problems. They know an MD on a deck doesn’t guarantee a successful business. What they require is a model the partners can underwrite and a board the operating team can run.
They’ve funded two categories exceptionally well.
The first is consumer-direct diagnostic and longevity-imaging access. Function Health and comparable expansive lab panels, AI-quantified plaque imaging (Cleerly, HeartFlow), portable MRI, full-body screening, therapeutic plasma exchange. Each of these directly answers patient demand the formal system has been slow to serve. The patients who pay for these services genuinely value them. The companies are scaling because the demand is real and the willingness to pay is verifiable. The capital here is doing exactly what it’s supposed to do: capitalizing demand that exists.
The second is direct prescription access at scale. The wave of platforms for appearance medications, GLP-1 access, hair loss treatments, ED, anxiolytics, even off-label longevity scripts. Calling these “prescription mills” is partly fair and partly dismissive of a genuine access problem they solved. These services answer a real patient demand for fast access to a specific medication, at a price and convenience point the legacy system wasn’t delivering.
Both pipelines work. Both serve real demand. Both are funded because the unit economics are demonstrable.
What capital hasn’t been able to underwrite is scalable subspecialty consultative expertise. The interpretive work that consists of a board-certified specialist sitting with a patient, the patient’s data, the patient’s history, the patient’s goals, and the field’s current best understanding, and producing judgment the patient can act on. The reason capital hasn’t flowed into this category isn’t that investors can’t see the gap. They see it. The reason is that no presentable model has been demonstrated. Every prior attempt has either failed commercially or succeeded by stripping out the expertise. A model that scales interpretive depth at unit costs that work, with patient outcomes that prove the depth survived the scaling, hasn’t arrived at the size and rigor that supports investment.
That’s the unfunded category. It’s unfunded because the underwriting case requires demonstrating what hasn’t yet been demonstrated.
What has been built
Each of the four can rationally pursue what they value when those values align. Patients get fast access to diagnostic information and to specific medications they want. Providers participate either as prescribers in the medication-access pipelines or as consumer-facing brands in the diagnostic pipelines. Payers can be circumvented entirely by cash-pay models. They aren’t the customer in either lane. Investors can underwrite both pipelines on the strength of patient willingness to pay.
Both pipelines work. Function Health and Cleerly and portable MRI and consumer-grade scans on one side. The GLP-1 platforms and the lifestyle-medication services on the other. Both serve real demand. Both are funded because the economics work.
What hasn’t
Between those two efficient pipelines sits an entire dimension of healthcare that hasn’t been built and isn’t being funded. It’s the interpretive layer. The dimension isn’t unaddressed by accident. It’s unaddressed because no single stakeholder’s rational model points to it.
Patients can’t build it themselves. They can buy the data, they can read the LLM interpretation, they can find a specific drug, but they can’t manufacture the interpretive layer they’re looking for. They’re increasingly aware of the gap and increasingly willing to pay across it; what they can’t do is supply it.
Providers can’t scale it alone. The legacy practice models, academic medical centers, employed specialists in health systems, traditional fee-for-service practices, weren’t built to expand interpretive capacity faster than population health needs grow. The volume of patients seeking specialty consultative care is growing; the capacity isn’t.
Payers can’t price it within their accountability structure. The interpretive conversation in cardiology that prevents an MI eleven years later doesn’t show up in their P&L. Their incentive structure rewards them for declining to cover services whose payoff outlasts the member relationship.
Investors can’t fund it without a model. Capital hasn’t flowed because the underwriting case for scalable subspecialty expertise hasn’t been demonstrated. Every prior attempt has either failed commercially or succeeded by stripping out the expertise. A presentable model hasn’t arrived at the size that supports a Series A check, let alone a growth round.
So the dimension stays unaddressed. Not because anyone is wrong. Because no one’s rational pursuit lands on it.
That’s the unusual discordance. Four rational stakeholders, all funded, all serving demand they can credibly meet, all collectively building approximately nothing for the dimension that consists of expert clinical judgment delivered at scale.
What it would take
The standard model for closing this gap, the one that has been pitched to investors for years, assumes the following flow: an AI system produces a clinical summary or recommendation, a physician reads it, the physician signs their name to it. Some version of this is what “AI-augmented care” usually means in deck form. It’s also why most attempts have failed.
The model fails for a reason that isn’t about AI quality. It fails for a cognitive reason that has been understood in education research for decades. Working memory is limited. Decision-making collapses when the combined cognitive load, the intrinsic complexity of the clinical problem plus the extraneous load of processing a long unfamiliar document, exceeds what working memory can hold. A board-certified specialist reading a six-page LLM-generated summary can’t meaningfully verify it. They can only rubber-stamp it. The result is a flow that looks clinically supervised but isn’t, in the sense that matters for the patient or the liability or the trust that builds a real practice. (I wrote about the underlying cognitive-load principle in Cognitive Load Theory & software-based upskilling in clinical medicine.)
What works has three components, each doing a job the other two can’t do.
The first is a transparent deterministic clinical engine that does the heavy lifting of the actual clinical reasoning. It produces interpretive atoms, risk strata, differential rankings, treatment recommendations, contraindication checks, titration logic, each derived from guideline-level rules and individually defensible. The engine is where the thinking happens. It’s also where the auditability lives: each atom traces back to a rule the specialist can verify in seconds.
The second is an LLM that does the unstructured mapping. It parses what the patient brings. The prose history, the smartwatch data, the LLM interpretation the patient pre-loaded, the ECG snippet, the lab panel. It handles intake, conversational flow, summarization, and the language work the deterministic engine can’t reach. The LLM is the system’s interface with messy, unbounded patient input. It isn’t doing the clinical thinking.
The third is the expert user. Their role is to rapidly align with the engine. They don’t construct the interpretation from scratch. They engage with the atoms the engine has produced, confirm or modify or reject any individual one based on judgment, and provide the authority that stands behind the recommendation the patient receives. The specialist’s job is validation, not generation.
The distinction matters. It’s the difference between an LLM with a medical license attached and a specialist whose authority has been amplified by software. The first produces output that looks clinical but isn’t defensible at any depth. The second produces output the specialist can defend in a deposition or in a conversation with a sophisticated patient who asks why.
This is what makes the unit economics work. The specialist isn’t being paid to do interpretive work an LLM could draft a plausible version of. They’re being paid for the judgment that decides whether the engine’s atoms are correct for this specific patient, and for the authority that stands behind the recommendation. Validation is a fundamentally more compressible task than generation. A specialist who would have spent twenty minutes constructing an interpretation can confirm or correct one in two. Without sacrificing the depth, because the underlying thinking has already been done at the engine layer in a form the specialist can verify. That’s the only path to subspecialty consultative care that scales without stripping out the expertise.
The hard problem isn’t whether AI can summarize cardiology. It’s whether the encounter architecture can decompose cardiology cleanly enough at the engine layer that a specialist can align with the atoms rapidly, without their working memory being exhausted in the process. And whether the LLM layer can map patient-brought messiness onto the engine cleanly enough that the specialist doesn’t have to re-do the mapping.
That’s the technical and operational thesis underneath any presentable model for scalable expertise. The dimension has remained unaddressed because most attempts have invested in the wrong components. Building better LLM summarization, or better physician-side UI, rather than building the deterministic interpretive substrate that the specialist’s rapid alignment actually depends on.
Returning to the wellness device
The opinion piece I started with made a reasonable argument on its own terms. The patients I see now read it differently. They aren’t naive about device risk. They’re responding to the absence of an interpretive layer between the data they’ve acquired and the clinical decisions they’re trying to make. The wellness device is one form of the data; the Function Health panel, the CCTA, the Lp(a) result, the smartwatch ECG snippet, the genetic report are others. All of them arrive at the visit with a patient who has, increasingly, already run them through an LLM and who’s looking for someone whose judgment they trust to engage with what they’ve brought.
That interpretive layer is the dimension of healthcare no one is currently addressing at scale. Patients can’t build it. Providers can’t scale it inside the existing models. Payers can’t price it within their time horizon. Venture capital can’t fund it without a model that hasn’t yet been demonstrated.
The dimension is observable, the demand for it’s verifiable, and the path to filling it’s open. Whether the next several years produce a real category there will come down to whichever team manages to present a model that’s, simultaneously, clinically real and economically scalable. And whether the capital base is paying attention when one shows up.