Paid in Full, Evidence Pending: The Coronary CT Plaque Volume Story

The pattern repeats. Once seen, it can’t be unseen. A new diagnostic technology emerges with genuine scientific underpinning, gets embraced by academic societies, earns regulatory clearance, and lands Medicare reimbursement — all before anyone has run a randomized trial showing it actually improves patient outcomes. By the time the evidence base is examined critically, the technology is embedded in practice, billing codes are established, and the people best positioned to raise questions have consulting agreements with the company.

We’ve been here before with coronary calcium scoring, and CT FFR. We’re here again with AI-powered plaque quantification.

Before we get to the new technology, a brief reckoning with the old one.

Coronary artery calcium scoring does have some value in helping guide clinical decision making in certain patients. But the logic of calcium scoring, seductive in its simplicity, was extended far beyond what the evidence supports. A high calcium score, the story goes, means you have significant plaque burden and need aggressive treatment. A low calcium score is supposed to be reassuring, but as I’ve outlined in detail in a prior post, low/zero calcium score does not mean zero cardiac events. Numerically, most patients that present with a cardiac event have a low CAC score, because the plaques most likely to rupture and cause heart attacks are often non-calcified.

This is the fundamental limitation the industry identified as an opportunity. If calcium scoring can’t detect non-calcified plaque, then we need something that can. Enter quantitative CCTA plaque analysis.

Before we evaluate the add-on plaque technology, we need to reckon honestly with the foundational trial used to justify CCTA’s dominant position in chest pain evaluation — because the HeartFlow reimbursement story rests on a superstructure that deserves closer examination.

The SCOT-HEART trial is the central pillar for proponents of using coronary CTA more widely. In its five-year results published in the New England Journal of Medicine in 2018 — and reinforced in ten-year Lancet data — the investigators reported a 41% relative reduction in the rate of coronary heart disease death or nonfatal myocardial infarction in the CCTA-guided arm compared to standard care. That headline number has done enormous work in reshaping clinical practice guidelines, earning CCTA Class I recommendations, and — critically — establishing the intellectual framework that makes HeartFlow’s plaque analysis seem like a natural extension of proven CCTA benefit.

The problem is that SCOT-HEART has serious methodological vulnerabilities that the guidelines have largely ignored.

The primary endpoint was changed. The published trial protocol, registered before any patient was enrolled, specified a primary endpoint of “diagnostic certainty of angina pectoris secondary to coronary heart disease at 6 weeks.” That was the question the trial was powered and designed to answer. The 2015 Lancet publication reported exactly that endpoint — CCTA improved diagnostic certainty. Cardiovascular death or nonfatal MI appeared in that paper as a secondary outcome among several others. When the five-year follow-up paper was published in the NEJM in 2018, the composite of CHD death and nonfatal MI had become the headline primary finding of the trial — a different question from the one the trial was designed to answer. Sanjay Kaul, a biostatistician at Cedars-Sinai and one of cardiology’s most rigorous trial appraisers, identified this in JAMA Cardiology, noting that the MI/CV death composite was one of 22 secondary outcomes in the original 2015 publication. Had the authors applied a Bonferroni correction for multiple comparisons, the p-value — 0.004 in the published paper — would inflate to a non-significant 0.08. A finding that anchored a generation of practice guidelines would not have survived standard statistical discipline.

The result is entirely MI-driven, with no mortality signal. Every cardiologist understands the hierarchy of outcome evidence. Reductions in all-cause mortality are the most robust endpoint, followed by cardiovascular mortality, then hard nonfatal events. SCOT-HEART showed no difference in mortality whatsoever — not cardiovascular, not all-cause. The entire observed benefit consisted of 29 fewer nonfatal MIs across the two arms. Twenty-nine events. That is a fragile result on which to build national testing strategy.

Ascertainment bias is a credible alternative explanation. SCOT-HEART was open-label — physicians knew which patients had received CCTA and what the result showed. Kaul emphasized that foreknowledge of a normal or non-obstructive coronary anatomy creates a systematic bias toward not pursuing MI workup in patients who later present with chest symptoms. If a physician knows a patient had a clean CCTA two years ago, they are less likely to order troponins, less likely to admit for observation, and less likely to code an enzyme-only elevation as myocardial infarction. Events that would have been captured in the standard care arm may have been missed or coded differently in the CCTA arm — not because fewer occurred, but because the physician’s prior anatomical knowledge discouraged their recognition. Critically, this bias is inert with respect to mortality — you cannot code around a death. The fact that the entire result lives entirely in the nonfatal MI category and not at all in mortality is precisely what ascertainment bias predicts.

The medication mechanism is implausible. The authors’ proposed explanation for the MI reduction was that CCTA identified non-obstructive coronary disease, prompting preventive therapy intensification that reduced events. The differential use of standard medical preventative therapy between arms was approximately 5 percentage points — roughly 100 additional patients per arm receiving aspirin + statin therapy. Statins and aspirin, at their best, have a number needed to treat of approximately 50 to prevent one major event over five years. That implies roughly 6 prevented MIs attributable to the medication differential — not 29. Mandrola has noted the arithmetic explicitly: the proposed mechanism requires an NNT for statins and aspirin of approximately 3, which is not compatible with any existing evidence base for those therapies. The effect size is implausibly large given the mechanism offered to explain it.

The medication story also fails the PROMISE test. PROMISE randomized more than 10,000 patients to CCTA versus functional testing and found no difference in outcomes — despite a nearly twofold greater uptake of preventive medications and lifestyle interventions in the CCTA arm. If seeing coronary atherosclerosis on a scan reliably translates into outcome-improving medication changes, PROMISE — with ten times the sample size, conducted across 193 North American sites, funded by the NIH rather than a device company — should have demonstrated it. It did not. The composite primary endpoint of all-cause mortality, nonfatal MI, hospitalization for unstable angina, and major procedural complication occurred in 3.3% of the CTA arm versus 3.0% in the functional testing arm (HR 1.04, p=0.75). Not significant. Not even a trend.

The 13-trial meta-analysis initially excluded SCOT-HEART for good reason. Cardiologists Andrew Foy and John Mandrola co-authored a 2018 meta-analysis published in JAMA Internal Medicine that pooled 13 trials comparing anatomical to functional testing strategies. The finding: no difference in mortality, no difference in cardiac hospitalizations. There was a small statistically significant reduction in nonfatal MI favoring CCTA (0.7% versus 1.1%), but Mandrola has been candid that this finding was an artifact of peer review pressure — the investigators originally excluded SCOT-HEART from the analysis on the principled methodological ground that it did not directly compare the two testing strategies (it added CCTA on top of standard care rather than replacing it). Reviewers forced its inclusion. When SCOT-HEART is removed in sensitivity analysis, the MI signal disappears. What remains across 12 trials of 20,000 patients is no mortality benefit, no hospitalization benefit, and higher rates of invasive angiography and revascularization in the CCTA arm — meaning more procedures, not better outcomes.

The ten-year data did not rescue the story. The 2025 Lancet ten-year follow-up reported continued benefit, but the absolute event rate differential had narrowed, and the prescribing advantage between arms attenuated substantially over time — which is the opposite of what you’d expect if the mechanism were really durable medication intensification rather than short-term ascertainment effects. The same pattern of MI-only benefit, no mortality difference, and open-label outcome tallying through administrative health records (without blinded adjudication) was carried forward from the original report.

None of this means CCTA is useless. The diagnostic clarity it provides is genuine — SCOT-HEART itself documented that it reclassified the diagnosis in nearly a quarter of patients who received it. The question is whether that diagnostic clarity translates into improved hard outcomes. PROMISE, in a larger and better-controlled trial, found that it does not, at least not compared to functional testing. The SCOT-HEART findings remain biologically interesting, methodologically vulnerable, and inconsistent with the broader trial landscape.

This matters for the plaque analysis question in a specific way: if the evidentiary foundation for CCTA itself in stable chest pain is more contested than the guidelines suggest, then adding a $950 AI layer (the current Medicare reimbursement for this test) on top of that modality to further characterize plaque burden compounds uncertainty rather than resolving it. We are building on ground that is softer than it appears.

HeartFlow’s Plaque Analysis product is one of a few vendors with a product that at its core is an AI-assisted segmentation system applied to coronary CT angiography. The algorithm analyzes the CCTA images, traces the boundaries of the coronary vessel lumen and outer wall, and quantifies the tissue occupying the space between them — the plaque. It then attempts to classify that plaque into subtypes: calcified, non-calcified fibrous, and low-attenuation plaque, which is considered the highest-risk category due to its association with lipid-rich necrotic cores.

The technical achievement is real. Tracing coronary vessel walls across hundreds of cross-sectional images, accounting for cardiac motion and breath-hold variation, is genuinely difficult. HeartFlow has trained its algorithm on an enormous dataset and employs certified CT analysts to review and modify segmentations — it’s a semi-automated system with human oversight, not a fully autonomous AI diagnostic. The output is a quantified plaque burden report with volume measurements for each subtype, benchmarked against an age- and sex-matched reference population derived from 273,000 CCTAs.

The company’s primary validation study, REVEALPLAQUE, compared AI-derived plaque volumes against intravascular ultrasound (IVUS) across 432 lesions in 237 patients and reported a correlation coefficient of 0.91 for total plaque volume, which HeartFlow markets as “95% agreement.” This is the number you’ll see in every piece of HeartFlow’s marketing material. The actual Pearson correlation coefficients published in the paper are:

Total Plaque Volume (TPV) r = 0.91, calcified plaque r = 0.91, Non Calcified Plaque (NCP) r = 0.87, Low Attenuation Plaque (LAP) r = 0.28, lumen volume r = 0.93, vessel volume r = 0.94. PubMed

That LAP number — r = 0.28 — should stop you cold. This is the clinically most important plaque subtype, the one HeartFlow and the broader CCTA field are pitching as the key risk-stratifying feature (the “vulnerable plaque” marker), and the correlation with IVUS is essentially noise. To understand why, we need to understand this is a physics problem, not a software one.

CCTA-based plaque “characterization” is a misleading term, because the laws of physics makes current generation CT based plaque characterization impossible. Let’s review the current hierarchy of plaque imaging:

The test with the highest spatial resolution is Optical coherence tomography (OCT), which requires an invasive intracoronary catheter that achieves spatial resolution of 10-20 micrometers. At that resolution, you can see individual plaque components with genuine fidelity: thin fibrous caps (defined as less than 65 micrometers thick in pathological studies), macrophage infiltration, neovascularization, and the microstructural features that could potentially predict rupture risk one day. It is the closest thing to in vivo histology available.

Next up is IntraVascular Ultrasound (IVUS), also invasive, achieves resolution of approximately 100-200 micrometers. At this resolution, IVUS can robustly quantify plaque burden and identify positive remodeling — features with established prognostic significance — but it cannot resolve thin fibrous caps and is significantly limited by calcium shadowing, which blocks the ultrasound beam and obscures the plaque underneath.

CCTA achieves spatial resolution of approximately 1 millimeter.

That is a 50 to 100-fold resolution gap between CCTA and the gold standard for tissue characterization. At 1mm resolution, you are not seeing individual plaque components — you are seeing Hounsfield unit averages across voxels that each contain a mixture of tissue types. The algorithm’s ability to subtype plaque into “low-attenuation” versus “fibrous” categories rests entirely on HU thresholds applied to these averaged voxels, with the implicit assumption that HU values reliably reflect underlying tissue composition.

They don’t. Not reliably. Not across scanners. Not across institutions.

The literature is unambiguous on this point. Validation studies comparing CCTA plaque characterization to IVUS have found significant differences in mean HU values between plaque subtypes, but with considerable overlap between categories. The same tissue, with the same underlying composition, will register different HU values depending on the scanner manufacturer, reconstruction kernel, contrast timing and bolus volume, heart rate at acquisition, and whether iterative or filtered back-projection reconstruction was used. As one SCCT expert consensus document explicitly states, plaque HU overlap across subtypes “can be accentuated by variations in image quality, lumen enhancement, and tube potential“ — and the same document advises against using plaque classifications that imply mechanical or histological characteristics not reliably assessed by CCTA, such as “soft,” “mixed,” “vulnerable,” “lipid-rich,” and “fibrous” plaque. Furthermore, the HU threshold for low-attenuation plaque has been inconsistent across published studies, ranging from 30 to 90 HU — a range so wide that what one scanner classifies as high-risk LAP, another may classify as benign fibrous tissue.

The problem becomes acute at the plaque-calcium interface. When calcification is adjacent to non-calcified tissue, beam-hardening and partial volume effects cause systematic errors in HU measurement of the surrounding tissue. The algorithm sees what looks like low-attenuation plaque that is actually a measurement artifact. Given that the low-attenuation plaque designation is the highest-risk classification in the HeartFlow report — and the category most likely to trigger intensified treatment — systematic false positives in calcified plaques represent a clinically significant problem that the available validation literature cannot rule out.

There is also a size threshold below which CCTA simply cannot detect plaque at all. Small non-calcified plaques with a wall thickness below approximately 500 micrometers are invisible to CCTA regardless of how sophisticated the segmentation algorithm is. This isn’t a software problem, it’s a fundamental constraint imposed by the physics of CT imaging.

And then there is the microcalcification problem. Pathological studies of sudden cardiac death victims have found that approximately two-thirds of ruptured plaques contain microcalcifications — tiny calcium deposits that destabilize the fibrous cap from within. Current CT scanners cannot resolve microcalcifications below about 0.5mm in diameter, and only calcifications greater than 0.5mm are visible on CCTA. Most of them are below that threshold. The plaques most likely to cause the next myocardial infarction contain a feature that is invisible to the very technology being marketed as the key to identifying them.

What HeartFlow’s AI is actually measuring is not tissue composition. It is volume — the quantity of material within vessel wall boundaries, binned into categories based on HU averages that have overlapping ranges across tissue types and inconsistent values across acquisition parameters.

That is a meaningful measurement. It is reproducible, it correlates with disease burden, and it may add something over a visual gestalt assessment. But it is categorically different from what the marketing language implies when it talks about plaque “characterization” and identifying “vulnerable” lesions.

Return now to that REVEALPLAQUE correlation. The study validated AI-derived CCTA plaque volumes against IVUS. IVUS is described as the “reference standard for in vivo intracoronary plaque characteristics.” That language is carefully chosen — it’s the best available invasive tool. But IVUS itself cannot see thin fibrous caps, is blinded by calcium, and has its own inter-vendor and inter-observer variability. When HeartFlow reports 95% agreement with IVUS, it is reporting agreement with a measurement standard that shares many of CCTA’s own limitations. Neither modality can see what matters most at the level of the vulnerable plaque: the fibrous cap thickness, macrophage infiltration, and microcalcification pattern that distinguish the lesion about to rupture from the one that will remain stable for another decade.

True ground truth for plaque characterization is histology — tissue cut out of an artery and examined under a microscope. No large-scale validation of AI-CCTA plaque characterization against histology exists. The technology is validated against a surrogate that is itself limited by the same physics constraints the technology is trying to transcend.

Furthermore, the “agreement” metric in REVEALPLAQUE — a correlation coefficient — measures linear association, not agreement. High correlation is compatible with systematic bias. Bland-Altman analysis in a related validation study demonstrated that AI-QCPA underestimated total plaque volume by -9.4 mm³ and calcified plaque by -11.4 mm³ compared with IVUS — a systematic bias that the headline correlation number conceals. If AI-CCTA consistently underestimates plaque volume in calcified lesions, the most common lesion type in the symptomatic patients HeartFlow is targeting, the clinical significance is not trivial.

To evaluate any diagnostic technology fairly, you need to ask three distinct questions. Getting the first right doesn’t mean you’ve answered the second or third.

Gap 1: Does the measurement associate with adverse events? Yes. Total plaque burden on CCTA correlates with cardiovascular outcomes. This is well-established across multiple cohorts. Low-attenuation plaque features predict events in the SCOT-HEART substudy and others. I have raised a number of technical questions about the process of quantification of plaque volume, but for the purposes of this argument, will accede that Gap 1 is closed.

Gap 2: Does the measurement discriminate individual risk beyond existing tools? This is where the evidence gets thin. When AI-derived quantitative plaque analysis is compared to clinical risk factors plus coronary calcium score plus standard CCTA interpretation, the incremental discrimination is modest. The published NRI (net reclassification index) for adding AI-QCT plaque quantification over a competent standard CCTA read with plaque notation is in a range where confidence intervals cross zero. The technology adds automation and reproducibility over manual quantification. It does not appear to add new clinical information that a trained reader wasn’t already capturing, at least not at a magnitude that would be expected to change outcomes. Gap 2 is not reliably closed.

Gap 3: Does acting on the measurement improve outcomes? There is not a single randomized trial in existence that answers this question. Not one that tests plaque-volume-guided therapy versus LDL-guided therapy with hard cardiovascular endpoints. Not one that demonstrates that intensifying statin therapy in response to a high plaque burden report reduces myocardial infarction rates beyond what LDL-targeted therapy would have achieved anyway. Gap 3 is wide open.

HeartFlow’s commercial model requires all three gaps to be closed to justify its reimbursement position. The evidence closes only Gap 1.

The DECIDE registry — HeartFlow’s primary outcomes dataset, claiming ~22,000 patients enrolled — has published its primary results as a conference abstract in the Journal of Cardiovascular Computed Tomography, appearing in the supplemental proceedings of the SCCT 2025 meeting. It is not a full peer-reviewed publication. The abstract reports data from 972 patients, representing approximately 5% of the enrolled registry. The primary endpoint was management change rate, reported at 51.3%. The most common management change was statin intensification. The lead investigator himself described the results as providing “supportive analysis for future trials.” (It is worth pointing out that the study was entirely funded by HeartFlow and the principal investigators are paid HeartFlow consultants.)

A 51% management change rate in a population of patients with non-obstructive CAD on CCTA sounds impressive until you examine the baseline event rates. In the SCOT-HEART trial’s low-plaque subgroup — a close analog to this population — the five-year event rate was 1.4%, or approximately 0.28% annually. At that baseline risk, the number needed to treat with statin therapy to prevent one cardiovascular event over five years is in the range of 250-300. The DECIDE registry found that half of patients had their management changed in response to AI plaque analysis. But as we are soon to see that management change is largely driven by increasing the prescription of lipid lowering therapy that has outcome benefits that are oversold.

Proponents of AI-powered plaque quantification often argue that the technology’s true value lies in its ability to identify “high-risk” plaque features that prompt more aggressive LDL lowering, even in patients whose baseline levels are already at or below guideline targets. The logic is straightforward: seeing a quantified plaque burden report — especially one highlighting low-attenuation plaque — motivates physicians to escalate therapy, driving LDL down to levels that ostensibly prevent events that standard risk calculators might miss. In practice, this frequently means adding PCSK9 inhibitors like evolocumab (Repatha) or alirocumab (Praluent) on top of high-intensity statins, aiming for LDL values below 40 or even 30 mg/dL. This argument deserves scrutiny, particularly in the context of the patients most likely to receive plaque analysis: those with non-obstructive coronary disease on CCTA, a group whose baseline event rates are low (often <1% annually).

The landmark trials of PCSK9 inhibitors — FOURIER and ODYSSEY OUTCOMES — demonstrate that further lowering already low LDL does confer benefit, but the absolute risk reductions are vanishingly small, consistent with the law of diminishing returns that governs biological systems as surely as it does economic ones. Moreover, substantial residual risk for bad cardiovascular events persists even at profoundly low LDL levels, underscoring that plaque quantification may be prompting treatment without proportionally large outcome gains. FOURIER randomized 27,564 patients with established atherosclerotic cardiovascular disease and LDL ≥70 mg/dL despite optimized statin therapy to evolocumab or placebo. Evolocumab reduced LDL from a median baseline of 92 mg/dL to 30 mg/dL — a 59% relative reduction. The primary composite endpoint (cardiovascular death, myocardial infarction, stroke, hospitalization for unstable angina, or coronary revascularization) occurred in 9.8% of the evolocumab arm versus 11.3% in placebo over a median 2.2 years (hazard ratio 0.85, 95% CI 0.79-0.92), for an absolute risk reduction of 1.5%. The key secondary endpoint (cardiovascular death, MI, or stroke) showed a similar pattern: 5.9% versus 7.4% (HR 0.80, 95% CI 0.73-0.88), absolute reduction 1.5%. There was no mortality benefit — neither all-cause nor cardiovascular. The number needed to treat to prevent one primary event over two years was approximately 67.

ODYSSEY OUTCOMES, enrolling 18,924 post-acute coronary syndrome patients with LDL ≥70 mg/dL on high-intensity statins, found comparable results with alirocumab. LDL fell to approximately 40 mg/dL at 4 months and 66 mg/dL at 48 months. The primary endpoint (death from coronary heart disease, nonfatal MI, fatal or nonfatal ischemic stroke, or unstable angina requiring hospitalization) occurred in 9.5% of the alirocumab group versus 11.1% in placebo over a median 2.8 years (HR 0.85, 95% CI 0.78-0.93), absolute reduction 1.6%. All-cause mortality was nominally reduced (3.5% vs 4.1%, HR 0.85, 95% CI 0.73-0.98), absolute reduction 0.6%. The number needed to treat for the primary endpoint was 63.Both trials illustrate diminishing returns: the relative risk reductions (15-20%) are consistent with the Cholesterol Treatment Trialists’ meta-analysis, where each 39 mg/dL LDL reduction yields approximately 22% relative event reduction. But as baseline LDL falls, the absolute benefit shrinks. In FOURIER, patients in the lowest baseline LDL quartile (median 74 mg/dL) saw benefits comparable to the overall cohort, but the absolute event differential was smaller. In ODYSSEY, patients with baseline LDL ≥100 mg/dL derived a larger absolute reduction (3.4%) than those with lower starting levels. Extrapolating to plaque analysis patients — often with non-obstructive disease and baseline LDL already near 70 mg/dL — the incremental absolute benefit from adding a PCSK9 inhibitor could be as low as 0.5-1% over 2-3 years, with numbers needed to treat exceeding 100.

The benefit in these two large trials is entirely driven by non-fatal ischemic events : non-fatal heart attacks, strokes, and coronary revascularization. So this is an extremely small absolute gain that must be weighed against persistent residual risk. Even at LDL <30 mg/dL in FOURIER, events occurred — the trial’s Kaplan-Meier curves show no plateau, but the residual event rate remained substantial. Post-hoc analyses from both trials confirm a monotonic relationship: lower LDL correlates with lower risk down to <40 mg/dL or even <25 mg/dL, without safety signals like increased hemorrhagic stroke or neurocognitive decline. Yet pathology tells us why: ruptured plaques involve factors beyond LDL, including inflammation (hsCRP often unchanged by PCSK9 inhibition), thin fibrous caps, microcalcifications, and thrombotic propensity. In ODYSSEY, patients achieving LDL <25 mg/dL had event rates similar to those at 25-50 mg/dL after propensity matching, but events weren’t eliminated. For plaque proponents, this means the “actionable” report may drive PCSK9 use in lower-risk cohorts where the arithmetic doesn’t favor large gains.

The DECIDE registry’s 51% management change rate — often statin intensification or PCSK9 addition — occurs in a population with ~0.3% annual event risk. Applying FOURIER/ODYSSEY’s 15% relative reduction yields an absolute risk reduction of ~0.05% annually, or one non-fatal cardiovascular event prevented per 2,000 treated patients per year. At $5,000-$6,000 annual cost for PCSK9 inhibitors (post-rebates), the price per event averted approaches millions — a far cry from cost-effective thresholds.

This isn’t to dismiss PCSK9 inhibitors; in truly high-risk patients (e.g., recent ACS with LDL >100 mg/dL), the trials show meaningful benefit. But extending them reflexively based on a plaque report in stable, non-obstructive disease risks overmedicalization. The law of diminishing returns reminds us: biology isn’t linear. Pushing LDL from 70 to 30 mg/dL isn’t equivalent to dropping from 130 to 70 — the absolute payoff shrinks as you approach the floor, and the floor doesn’t even get you close to zero risk. Plaque quantification may create the illusion of precision, but the current state of cardiovascular therapy means that most of the benefit accrues to those selling CAT scans and plaque quantification software, not patients.

Let’s start with what famed economist Joseph Schumpeter would argue HeartFlow got right. In The Theory of Economic Development, published in 1911, Joseph Schumpeter argued that the entrepreneur’s defining role is not to respond to existing demand but to create it.

“It is the producer who as a rule initiates economic change, and consumers are educated by him if necessary; they are, as it were, taught to want new things, or things which differ in some respect or other from those which they have been in the habit of using.” - J Schumpeter

On this account, HeartFlow is doing exactly what a successful innovator is supposed to do. They identified a genuine limitation of existing tools — calcium scoring cannot see non-calcified plaque — built a technically sophisticated product to address it, trained physicians to recognize a data point they had never previously possessed, and created demand for something that did not exist before. That is Schumpeterian entrepreneurship operating as designed.

The problem is not the innovation. The problem is where it lands.

In a normal market, Schumpeter’s entrepreneur faces a discipline that HeartFlow does not. Consumers can defer purchase until value is demonstrated. Competitors can offer alternatives. Buyers can simply decline. The entrepreneur who teaches consumers to want something must eventually prove that what they are buying is worth having. The market, however imperfectly, applies that test over time.

Medicare reimbursement does not work this way. Once a Category I CPT code exists and a national coverage determination is in place, the patient has no meaningful role in the purchase decision. The physician’s financial incentive runs toward using the tool, not interrogating it. The entrepreneur has successfully induced demand. No physician was missing a quantified plaque volume report before Heartflow made one available. The felt clinical need was created by the product — not by a gap in patient outcomes. Schumpeter charitably would term this education. In medicine we call it a CPT code.

The institutional infrastructure — the CPT panel, the coverage process, the professional society guideline — has already rendered its verdict. The window for market discipline closes the moment the billing code is established, which is precisely why the sequence of events documented in this article matters so much. The entrepreneur who successfully navigates that sequence extracts payment not because the product proved its value to consumers, but because the advocacy was efficient.

This is where Economist Galbraith’s darker analysis takes over. In The Affluent Society, published in 1958, Galbraith introduced the concept of the dependence effect: the observation that in a wealthy society, the wants of consumers are increasingly created by the very process of satisfying them. His illustration was sharp. “The case cannot stand if it is the process of satisfying wants that creates the wants. For then the individual who urges the importance of production to satisfy these wants is precisely in the position of the onlooker who applauds the efforts of the squirrel to keep abreast of the wheel that is propelled by his own efforts.” The trap he’s identifying is that the want and the satisfaction are created by the same process. The squirrel doesn’t have an independent desire to reach a destination. The running itself generates the feeling that running is necessary.

Applied to consumer society: companies create products → marketing makes you feel you need them → you buy them → the act of buying reinforces that the need was real → companies make more products. The “need” was never independent of the supply. It was manufactured alongside it.

HeartFlow has operationalized the dependence effect at the level of federal healthcare reimbursement. The DECIDE registry measures management change as its primary endpoint. Physicians see quantitative plaque numbers they didn’t have before. The numbers look precise and actionable — total plaque volume in cubic millimeters, low-attenuation plaque as a percentage, a percentile ranking against age- and sex-matched peers. Confronted with data they did not previously possess, physicians change management. HeartFlow then cites the management change rate as proof that the tool fills an unmet clinical need.

But here is the circularity: before the report exists, the physician is not missing it. The clinical need for quantitative plaque characterization is created by the act of providing quantitative plaque characterization. The physician who never saw a HeartFlow report was not managing their non-obstructive CAD patient incorrectly by some prior standard. They were using LDL, clinical risk scores, and the qualitative CCTA interpretation. The new report creates a new data point, the new data point creates cognitive pressure to act, the action is recorded as a management change, and the management change is presented as evidence of clinical value.

Schumpeter’s entrepreneur teaches consumers to want new things. Galbraith’s squirrel runs harder to keep up with wants the wheel itself is generating. In a functioning market those two forces eventually reach equilibrium — products that don’t deliver lose customers. In Medicare reimbursement, once the wheel is spinning, no one turns it off. The house has no walls and no foundation. We have built a billing infrastructure on the belief that because the tool changes behavior, it must be improving outcomes. Those are not the same thing.

Here is how the reimbursement architecture was constructed for HeartFlow’s plaque analysis.

In 2021, a Category III provisional CPT code was established — a temporary code for emerging technologies pending evidence accumulation. Through local coverage determinations by Medicare Administrative Contractors, the technology began generating reimbursement at various rates. In the 2024-2025 reimbursement cycle, CMS established national reimbursement of approximately $950 per study. In January 2026, HeartFlow’s plaque analysis received a Category I permanent CPT code — the same designation given to established procedures with mature evidence bases.

Each step in this progression required orchestrated advocacy. The AMA’s CPT Editorial Panel, which assigns billing codes, relies heavily on professional society testimony about clinical utility. CMS’s coverage determination process relies heavily on professional society guidelines and scientific statements to establish whether a technology represents “reasonable and necessary” care. The Society of Cardiovascular Computed Tomography (SCCT) and the American College of Cardiology were the primary societies advancing this agenda — culminating in a December 2025 ACC Scientific Statement explicitly supporting HeartFlow Plaque Analysis, published just as the Category I code was being finalized.

What requires scrutiny is the relationship between those societies, their leadership, and HeartFlow.

Jonathon Leipsic, former president of SCCT, discloses HeartFlow consulting fees and stock options across multiple peer-reviewed publications and in roundtable expert opinion documents. He appeared as a featured speaker at a HeartFlow-sponsored symposium at SCCT 2023, held the morning before the REVEALPLAQUE and DECODE study presentations. He is a listed co-author on the 2025 ACC Scientific Statement on Quantitative Coronary Plaque Analysis that HeartFlow immediately cited in its press release announcing the Category I code approval. And his recent Medscape commentary “CT-First for Chest Pain: The Anatomical Truth Machine” — which a Medscape counterpoint piece by John Mandrola explicitly noted fails to cite supporting randomized trial evidence — advocates for CCTA as the frontline test for chest pain evaluation. That is precisely the patient population whose scans generate the add-on plaque analysis revenue.

The 2025 SCCT Clinical Trials Award went to a DECIDE registry investigator — the HeartFlow-funded study whose data was simultaneously being used to justify the Category I code application.

This is not, importantly, a story about any individual acting improperly. The individuals involved are accomplished scientists with legitimate expertise. The story is structural. CMS’s reimbursement process is designed to rely on professional society guidance as a proxy for clinical consensus. When professional society leadership has systematic financial entanglement with the companies seeking coverage, the guidance the societies produce is not independent of commercial interest — even when the individuals involved believe sincerely in the technology. The conflict is baked into the architecture, not the intentions.

The result is a system in which the same network of experts writes guidelines, testifies to CPT panels, lobbies CMS, runs sponsored symposia, receives society awards for company-funded research, and publishes in society journals — all simultaneously representing both professional clinical judgment and the commercial interests of the companies whose products they’re evaluating. At scale, if even 20% of the approximately two million coronary CTAs performed annually in the United States add an AI plaque analysis, the resulting Medicare spend approaches $400 million per year. That is $400 million contingent on the claim that the three evidence gaps described above have been closed — when only the first has been.

This critique is not an argument that CCTA lacks any diagnostic value. In the right clinical setting — evaluating a patient with new stable chest pain whose pretest probability warrants anatomical clarification — CCTA provides genuine information. It identifies non-obstructive disease that functional testing misses, it rules out obstructive CAD with high negative predictive value, and it changes clinical management in ways that are often appropriate. The SCOT-HEART investigators documented real diagnostic reclassification in nearly a quarter of patients who received CCTA. That is a legitimate contribution.

It is, however, an argument that the trial evidence for outcome benefit from CCTA is considerably weaker than the guidelines imply. PROMISE — the larger, NIH-funded, methodologically cleaner trial — found no outcome difference between CCTA and functional testing in 10,000 patients. SCOT-HEART’s findings are inconsistent with the PROMISE result, vulnerable to ascertainment bias in an open-label design, dependent entirely on 29 nonfatal MIs with no mortality signal, statistically fragile when appropriate corrections for multiple comparisons are applied, and mechanistically implausible given the modest medication differential that supposedly explained a 41% relative risk reduction. The broader meta-analytic evidence, absent SCOT-HEART, shows no mortality or hospitalization benefit. Reasonable cardiologists can and do disagree about the net value of CCTA as a chest pain testing strategy. The guidelines’ confident Class I recommendation is poorly supported.

This is not an argument that plaque biology is clinically irrelevant. Total atherosclerotic burden is a meaningful risk marker. The association between plaque volume and adverse events is real and robustly demonstrated across multiple cohorts.

This is not an argument that HeartFlow’s algorithm doesn’t work as described. The technology does what it claims to do — it quantifies plaque volume from CCTA with reasonable reproducibility and benchmarks it against a large reference population.

If patients and their doctors want to avail themselves of this new technology, I think they should go ahead and add this particular datapoint to their health file. The problem is when you ask me to pay for your test, which is what happens in the current American medical marketplace when a test is deemed to have clinical value and is “covered” by insurance/government programs. This puts a very heavy burden on making the sure the process to determine clinical value is working correctly.

Sadly it isn’t. The story about the $950 additional cost per coronary CTA on “AI analysis” is the price of admission to a movie that reveals systemic problems: a reimbursement system that treats professional society endorsement as a proxy for clinical evidence, that creates permanent billing codes before outcomes data exists, and that allows the financial entanglement between industry and academic leadership to flow unimpeded through the guideline-writing and advocacy process. HeartFlow's plaque analysis is one example. CT-FFR before it was another.

The pattern repeats because the system rewards it. Technologies are adopted not because good trials demonstrate that patients live longer or better, but because a network of credentialed experts — many with undisclosed or disclosed-but-ignored financial relationships with the relevant companies — successfully navigates the CPT panel, the CMS coverage process, and the guideline-writing committees. Once reimbursement is established, the technology becomes embedded in practice, the billing infrastructure creates its own momentum, and the window for rigorous evaluation closes. The $950 question is really a $400 million question, and behind it is a question about whether American medicine has the institutional will to insist that the things it pays for actually work. Galbraith’s squirrel is running very fast. It is not getting anywhere.

Anish Koka is a Cardiologist in Philadelphia. He cohosts a weekly medical-health policy podcast called The Doctor’s Lounge. He is on @X : @anish_koka