We Want to Retract Our Own Paper

My colleagues and I have done something unusual: we’ve asked to retract our own paper. What follows is a transparent explanation of why.

*Notice: What follows is an additional letter beyond our usual programing.

*Fair warning: It dives into an intense—and fairly niche—topic. If scientific drama isn’t of interest, or if you haven’t followed the controversial KETO-CTA study over the past year, this might not be the letter for you—and that’s okay.

*But for those who’ve been waiting nearly a year for updates, I can finally say more—and I’m excited to do so.

*This letter is free, but if you find it valuable, you can support the work by sharing it.

My colleagues and I have done something unusual: we’ve asked to retract our own paper. What follows is a transparent explanation of why.

First, make no mistake—this is, for me, a joyous moment! Not a moment of shame or regret. I am over-the-moon happy.

I’ve wanted this for months. If that seems unusual, it should. But it also shouldn’t, once you understand the details.

I also want to make clear that we’ve signaled about this in every capacity we could.

That includes an official statement from the Citizen Science Foundation—the public charity we used to fund this study—founded by my best friend and colleague on this paper, Dave Feldman.

The statement makes clear that the request came from us, after we raised concerns about externally generated analytical data, specifically from the company Cleerly^®, which performed an AI analysis of the coronary CT angiography (CTA) heart scans from our study.

After the paper was published, multiple issues came to light which I’ll discuss below.

And only after lengthy, good faith efforts to resolve this with Cleery^®, we took this to the journal and formally requested to retract the paper that contained their analysis.

*And, before we begin, I’ll highlight you can see Dave’s short (2 minutes) announcement regarding the retraction request, HERE.

The KETO-CTA Backstory: A Rapid Review

Now, this has been an ongoing drama for the last year.

Some of you may already be caught up.

But for those who are interested in this development and don’t yet have all the details, let me bring you up to speed as quickly as I can.

On April 7th, we published a paper in the Journal of the American College of Cardiology: Advances, titled “Longitudinal Data From the KETO-CTA Study: Plaque Predicts Plaque, ApoB Does Not”

It was a prospective study of 100 individuals with very high LDL cholesterol on a ketogenic diet, whose LDL increase appeared as part of a triad of markers: high HDL, low triglycerides, and generally otherwise excellent metabolic health.

This phenotype is known as the Lean Mass Hyper-Responder (LMHR).

LMHR are an important population to study because it is the first population ever observed in history with very high LDL levels without either:

• underlying metabolic dysfunction (such as metabolic syndrome), or

• a congenital disorder like familial hypercholesterolemia causing the high cholesterol.

In other words, this elevation appears to be a metabolic response.

To use myself as one extreme example, I can toggle my LDL levels between about 90 and nearly 600 through dietary changes. No drugs—just forcing my body to adapt to different dietary environments. The result is enormous swings in cholesterol.

*This is my personal latest unmediated lipid panel. To date, my arteries are perfectly clear.

In the cardiology and lipid world, this is entirely novel. And we learn a tremendous amount by studying outliers.

That is why this study is so important.

But even if this LMHR phenotype is an adaptive metabolic response, the question remains: what is the risk?

What About Cardiovascular Risk? Enter KETO-CTA

Prior to this study, we didn’t know whether this population was a high-risk population. That’s exactly what the study sought to uncover.

The major finding of the April 7th paper was that LDL cholesterol exposure and ApoB levels did not predict atherosclerotic plaque progression in this population. Instead, baseline plaque measurements predicted plaque progression.

Importantly—and I don’t want you to lose this thread—that finding remains incredibly robust.

In fact, since publication, multiple additional analyses—including applied analysis by Heartflow^® AI and by the pre-specified QAngio^® methodology—have found the same result.

That finding has not changed.

LDL cholesterol and ApoB do not predict plaque progression in this population, despite this cohort having the largest LDL cholesterol spread of any prospective imaging study ever conducted.

That is interesting.

However, there was a detail the initial paper got wrong—a detail we didn’t know about until after publication.

This is where our story begins.

What We Didn’t Know

But first, here’s something else weird that might not be intuitive.

Despite being co–first author on this paper I did not have access to the underlying data, nor did Dave Feldman, as he represents the Citizen Science Foundation funding body within the team.

This was purposeful.

Everyone who isn’t involved in the statistical analysis toward the original findings is appropriately “blinded” from the raw data in order to protect the integrity of the process.

Information we could only access after publication.

However, after publication, the raw, anonymized data was then provided to the Citizen Science Foundation, and thus Dave Feldman, who did a deeper analysis and found a number of anomalies.

He immediately alerted both the research Institute and Cleerly^®.

These concerns were further framed by the appalling revelation that the scans—the heart scans—had been run in an unblinded fashion.

This was a prospective trial looking at plaque progression in individuals with very high cholesterol. Properly done, Cleerly^® AI should have read the before-and-after scans in a blinded fashion. When this is not done, it can introduce a major source of bias.

This was concerning, so Cleerly^® was approached and asked to redo the scans—of course offering additional compensation for their labor, eventhough the scans should have been blinded in the first place. They refused. In fact, they continually refused.

*Author’s aside: And for those of you wondering why they refused, let me say only this: that’s an excellent question.

Failure by Cleerly^® to run the scans in a blinded fashion prompted the Citizen Science Foundation to expend additional resources and effort to obtain an additional blinded set of scans done by another company, Heartflow^®. (It’s worth emphasizing Heartflow is the most independently validated AI platform for CTA heart scans.)

The Heartflow^® results came back with a drastically different measure for overall plaque progression. Much lower. Much, much lower.

*Sample AI-guided Heartflow^® analysis of an LMHR patient, (total cholesterol >700 mg/dl for > 5 years). Images used with patient’s consent.

But that’s not all.

When the pre-specified QAngio^® methodology was also used—independent of the blinded Heartflow^® analysis—the QAngio^® and Heartflow^® data agreed with each other and disagreed with the unblinded Cleerly^® analysis.

So, to step back and review:

• The Cleerly^® analysis contains multiple anomalies.

• The Cleerly^® analysis was inappropriately unblinded, introducing a major source of potential bias.

• The company refused to perform a fully blinded analysis, upon multiple requests.

• Two independent analyses agreed with each other and disagreed with Cleerly^®. This included the pre-specified QAngio^® methodology and a blinded analysis by Heartflow^®.

*Author’s aside: Layered on top of all this were conflicts of interest with Cleerly^® that were not disclosed prior to publication—disclosures that absolutely should have been made, such as the Chief Medical Officer being a co-author.

*This is not new information. You can Google it. This occurred outside my scope of control given my position within the project, despite being co–first author on the paper. At worst, I suppose you could fairly accuse me of naïveté for assuming those conflicts would be disclosed and that the scans would be properly blinded.

The Next Big Plot Twist…

And then something brilliant happened.

Something, a maneuver, even I didn’t know about—being kept in the dark for my own legal protection until the information was ready to go public.

A subset of participants, through their cardiologists, independently resubmitted their scans to the same company, Cleerly^®.

This time, the scans were submitted in a properly blinded fashion, given Cleerly was likely unaware it was happening.

So, we effectively obtained a blinded reanalysis from Cleerly^® via a backdoor.

What did they find?

The difference was gargantuan!!! (Yes, three “!” are warranted. Actually, let me toss in two more for good measure: !!)

When I saw the numbers, I was stunned. I was doing backflips in my bedroom. Here.

*These are screenshots from the moment I saw the data for the first time earlier this week. I was genuinely stunned and overjoyed because it knew this was the nail in the coffin… the check mate moment.

Roughly half of the participants who resubmitted their scans not only saw their estimated plaque progression shrink—it inverted!

Of the eight who resubmitted (we will get to that n = 8 number momentarily), the mean change in plaque went from +20.9 mm³—substantial progression—to –5.1 mm³.

From clear progression to actual regression.

And below are shown the group median changes in non-calcified plaque volume as measured by coronary CT angiography (CCTA) run through Cleerly^® (unblinded), Heartflow^®, QAngio^® and Cleerly^® (blinded).

• TL;DR: one of these is not like the others.

So now step back and look at the full picture:

• The initial analysis—unblinded, with undisclosed conflicts—stands alone as an outlier.

• Three independent analyses largely agree with each other.

• This now included the blinded repeat analyses conducted by Cleerly^® itself.

Why We Asked to Retract Our Own Paper

By now, maybe you can understand why we want to retract our own paper.

The initial Cleerly^® analysis is unreliable to anyone with a modicum of reason. The analysis was conducted unblinded, inappropriately—and the company didn’t check its work. In fact, it refused to.

Multiple independent analyses—including applied Heartflow^® analysis, the pre-specified QAngio^® methodology, and even a randomly selected sample of participant scans rerun by Cleerly^® itself in a blinded fashion—all contradict the original Cleerly^® analysis as far as plaque progression is concerned.

As the evidence continued to pile up over months, Dave was hard at work trying to get Cleerly^® to finally agree to do a proper, fully blinded re-analysis – which is technically not a “re”-analysis since it could be correctly argued we have yet to get the original blinded analysis expected in every longitudinal study.

Regardless, the highlighted finding in the initial paper—that LDL cholesterol and ApoB levels don’t predict plaque progression in this population—remains robust.

Stated another way, every single analysis – even the Cleerly^® analysis – on these same 200 scans substantiates this major finding.

But the absolute level of plaque progression reported in the original April 7^th, 2025 paper based on the Cleerly^® analysis has been unequivocally and demonstrably refuted.

I also want to add a human element to this…

Imagine being that patient—the patient told you have extraordinarily high plaque progression.

The news can feel like hearing from your doctor that you have three years to live because of a terrible cancer. But then you come back after those three years and your doctor says, “Whoops—we misread the CT scan. You’re actually totally fine.”

Sure, you’d be excited to hear you’re going to live to a ripe old age. But you’d also be frustrated that the error wasn’t caught earlier and the information corrected.

But in this case, it’s worse—because the failure to redo the analysis in a blinded fashion was willful on the part of the company.

For all of these reasons, my colleagues and we have wanted to retract the paper for months.

Unfortunately—and now I’m being very diplomatic and polite—the wheels of scientific bureaucracy move slowly. So, after months, we decided to make an announcement that we would request the paper be retracted. We waited far longer than was reasonable for the journal to do so. (As Dave announced, this was approved on January 12).

Response to Critiques

Now, in the melee that has ensued on social media since we announced earlier this week that the blinded Cleerly^® analysis also contradicts their initial findings, the responses by critics have been disappointing—but expected.

I’m all for vigorous intellectual discourse and debate, but the seemingly willful ignorance of some of our medical and academic detractors is, frankly, embarrassing.

Yes, this was a highly controversial study, for reasons you can probably now understand. And yes, this centers on one of the deepest dogmas in medicine: the relationship between cholesterol and heart disease.

But that does not excuse what, from parts of the academic community, amounts to intellectual insincerity—or, if I’m giving people the benefit of the doubt, stupidity.

So let me run through some of the common counterarguments and explain why they do not hold water.

In an attempt to provide clarity amid the confusion, what follows are responses to common critiques about KETO-CTA.

1. The New Data “Are Just 8 People”

FALSE!

We are not hanging our hats on eight participants. We are hanging our hats on:

• Three coherent analyses that agree with each other, including:

  • The pre-specified QAngio^® methodology

  • The independent, fully blinded analysis conducted by Heartflow^®

  • The de facto blinded analysis conducted by Cleerly^®

*Author’s aside: Also consider the probability that the original unblinded analysis showed effectively zero regressors—and yet a random subset of those very same scans, including eight participants, suddenly shows 50% regressors.

*By way of an honest—if admittedly ridiculous—analogy: there is a higher probability that I could blindfold myself, stand in the middle of a hurricane, punt a ping-pong ball the length of a football field, and have it land directly into a red Solo cup. That’s how improbable it would be for the results observed in the blinded analysis of those eight participants to arise purely by chance if the original unblinded analysis were correct. In other words: this is not a chance event.

2. “They All Had High LDL”

People continue to invoke the argument of a limited dose-response effect.

The claim is that LDL levels were high across participants, and therefore we should not expect to see an association between LDL/ApoB exposure and plaque progression.

This is not a logical argument. It has simply become a meme that is propagating through popularity rather than reason.

Let me explain—using this critic’s analogy about NBA players.

In this analogy “high LDL” in the KETO-CTA study is like being tall (over 7 feet).

Quoting my response on X:

“Apologies, but your analogy is flawed. High LDL doesn’t equal high LDL, just like tall doesn’t equal tall. To make your analogy a fair one in this instance, we would have to imagine that the distribution of human heights above 7 feet ranges from 7 feet to ~20 feet.

The KETO-CTA study has the largest LDL spread of any prospective imaging study ever.

So, if you’re invoking range restriction or a saturation effect, then that is in direct contradiction to the current lipid–heart hypothesis.”

Another extension of this logic would be that lean mass hyper-responders at the upper range of LDL would derive essentially no benefit from treatment, since they would still remain above the saturation ceiling.

So:

• Does 7 ft tall ≈ 20 ft tall?

• Does LDL 600 ≈ LDL 200?

• Should an LMHR with LDL 600 not be treated because they’d still be above the “saturation ceiling”?

These aren’t my claims—they’re logical extensions of the dose–response counterargument that has itself become a sort of meme.

3. “The authors are untrustworthy. This study was botched. Ignore, ignore, ignore.”

This one is fairly easy to address because it is a transparent avoidance of the facts.

These claims rely on ad hominem attacks, driven either by ego defense or ignorance or both.

The reality is that Dave, Adrian, and I did not have all the information prior to April 7, largely to protect the integrity of the project.

Since April 7, we have been doing everything possible to uncover and present the scientific truth.

There have been obstacles—including Cleerly^®’s refusal to redo their analysis after we discovered it had been unblinded.

It is certainly anyone’s prerogative to withhold the benefit of the doubt. But that does not excuse ignoring the facts in favor of personal attacks. Frankly, to those who choose that path, you are only strengthening our hand.

4. The “Preponderance of Evidence Says…”

Another common counterpoint is some version of the statement that the “preponderance of evidence” shows a causal relationship between LDL/ApoB and cardiovascular disease, and that this cannot be overturned by one study—especially one that is not a randomized controlled trial.

Again, this is not actually a counter-argument. It is an intellectual dodge.

Science and medicine 101: you should not generalize findings from one population to another.

The entirety of the so-called preponderance of evidence does not study lean mass hyper-responders, who are unique in several important respects.

Unlike the general population, they tend to be metabolically healthy and insulin sensitive.

And unlike patients with familial hypercholesterolemia, their elevated LDL and ApoB are not due to a congenital genetic defect, but rather to a metabolic response.

In effect, lean mass hyper-responders represent the first population of people with isolated high LDL that is not caused by a broken lipid metabolism.

So why would we expect “preponderance of evidence,” drawn from populations on mixed diets, with genetic defects in lipid metabolism, and/or who are metabolically sick, to apply?

Another foundational rule in science and medicine is that we often learn the most from studying outliers. Ignoring this population is therefore not only intellectually lazy—it is scientifically irresponsible.

Furthermore, it is unequivocally true that there is inter-individual variability in the relationship between LDL/ApoB exposure and plaque accumulation. There are individuals with extensive LDL exposure who accumulate little to no plaque, even over many decades.

Understanding this variability is essential for advancing science, improving population health, and—right now—individualizing clinical care.

As for my position on ApoB, it is provocatively moderate:

And please feel free to quote me on this:

LDL and ApoB are necessary but not sufficient to cause cardiovascular disease. Treatments that lower LDL and ApoB have variable absolute benefits and variable costs, depending on the individual. Therefore, individualized risk–benefit analysis requires a careful calculus—one that takes into account far more than a person’s ApoB level.

At the end of the day, the goal here is to understand the truth.

The responsible path forward is simple: examine the evidence honestly, test competing hypotheses rigorously, and follow the data wherever they lead.

That is what we have been committed to from the beginning, and that is what we will continue to do.

These are the four main areas of pushback—and what I’ll just say are my definitive responses. I’ll add a fifth and sixth, more for my own amusement, but I’ll put them in a different category because they are next-level absurd.

5. It’s Unethical to Run a Blinded Analysis

Some individuals are trying to defend Cleerly^®, rationalizing that they shouldn’t run a blinded analysis because it would somehow be unethical to do so because baseline plaque levels were so low in this population that it was below the noise floor.

First of all, this is absurd at the outset and makes no sense why having low baseline plaque levels would make it unethical for Cleerly^® to do a properly blinded analysis.

Furthermore, this incredible set of mental gymnastics is implicitly arguing that the population in question—with extraordinarily high LDL cholesterol who had had that hypercholesterolemia for near half a decade prior to starting the study—was such a low-risk population that advanced AI that is supposedly accurate and precise couldn’t detect plaque in their vessels with any reasonable degree of reliable precision.

Is that the argument you really want to make?

6. “Too Hard to Follow” … But trust me, “Because I’m a Doctor”

And the final, perhaps most pathetic argument of all, is this: “this is too hard to follow.”

When all the attacks are expended and you push people to the end of the rope, they often resort to personal attacks, or simply saying this is too hard to follow and therefore not trustworthy anymore.

Even when you lay it out in four simple bullet points, some individuals who call themselves doctors suggest that they need the support of “5 LLMs” to follow the trail.

I’ll grant them this: it’s confusing. But so are medicine and science—and we’re dealing with nuances that cut to the heart of both.

If a physician or scientist can’t read and follow four bullet points, the differential diagnosis narrows quickly: willful ignorance or simple incompetence. That’s not a personal attack; it’s just the range of plausible explanations.

That’s why I’ve written such a verbose letter on the topic—to address it comprehensively and clearly for those willing to invest the effort to process the facts.

The hard truth is that not only can biology and science be messy, but when you layer onto that human imperfection, ulterior motives and interests, teams that don’t agree, and a slow-moving scientific bureaucracy, you can get a mess like this.

And I apologize for whatever role I contributed to what has been a confusing process. I’m not promising that my communication on the matter has been perfect. What I am promising is that every step of the way, I have represented myself honestly, shared what I was permitted to share, and always tried to bring you clarity and truth.

At this point, trying to provide that clarity does require a lot of words, and that clarity will only be accessible to those who actually do the work to try to understand what went on here.

Still, consider this an evolving story.

The science is not settled, although this position certainly feels like a checkmate on this battleground.

But sometimes, to get your opponent in checkmate, it pays to sacrifice a queen.

That is why I—and I think I can speak for my friends Dave and Adrian—are proud and relieved to be announcing that we have formally and repeatedly pressed to have the Cleerly^® data retracted.

Any questions?

Post Script: Can Clinical Experience Make You a Worse Doctor?

Since I’ve clearly dedicated this morning to releasing my unfiltered—and at times less diplomatic—thoughts, let me address one more very common rebuttal I hear in academic medicine: the concept of clinical experience.

Of all the arguments and counterarguments I encountered in academic medicine, none is more confused than the topic of “clinical experience.”

Let me be clear: clinical experience has value. The insights gained from treating hundreds or thousands of patients hone a physician’s instincts in ways nothing else can. I know many experienced clinicians—including dear mentors and family members—who are extraordinary doctors and nothing short of heroes.

But that does not excuse using “clinical experience” as a shield against data.

Too often, I see the darker side of that phrase.

It’s invoked as a virtue signal for what is often something else entirely: confirmation bias.

Decades of exposure to the same dogma—seeing the same problems through the same lens—can create a visceral resistance to new ideas that challenge old models, even when those models have clear limitations.

When “clinical experience” becomes a reason not to engage with evidence, it stops being wisdom and becomes dogma.

And that is dangerous.

It is dangerous for patients. I know that personally, because I have been that patient. I have sat across from the physician who cites “clinical experience” while refusing to engage with the data—or even the reasoning.

That is medical paternalism at its worst, cloaked in the respectable language of experience.

So, let’s make an important distinction.

There is the humble, open-minded clinician—someone whose experience has taught them the limits of their knowledge and the importance of continuing to learn.

And then there are those who cite “clinical experience” as a final refuge when they have no better argument.

Some might ask whether this critique is disrespectful to the profession of medicine and the institution of science

My answer is simple: no. In fact, the opposite is true because I respect the profession of medicine, I believe can be better. And science isn’t an institution at all—it’s a process.

• If you found this letter valuable, please share it.

• If you’d like to learn more about heart health, please check out the StayCurious Metabolism Heart Health Library, HERE.