Why We Wanted to Retract Our Own Paper
In a Nutshell
Researchers behind a controversial JACC Advances paper on low-carb hyper-responders (lean individuals with sky-high LDL) are retracting it after discovering anomalies in the unblinded Clearly AI analysis, including zero regressors (everyone progressed), reversal of the "power of zero" (zero CAC predicted more progression), and disagreement with literature. Independent blinded reanalyses (Qangio, HeartFlow, participant-led Clearly) showed near-flat or regressing plaque, contradicting Clearly's results despite repeated offers for Clearly to reanalyze blindly (which they refused). They're proud of prioritizing integrity, with new studies (including controls) underway to clarify plaque progression in high-LDL low-carbers.
These notes were generated by AI and may contain inaccuracies.
Data was coming back not matching the provided dataset. Initial Clearly dataset overturned. If confident in analysis quality, why not do a check? Analyses outside theirs disagree. No good deed goes unpunished. Gambling heart, not social media points.
Special conversation, cathartic, putting everything on table. 11 months since publishing controversial study in JACC Advances, now pushing to withdraw. Proud and enthusiastic to pull from literature. Co-authors: host and Dave.
Dave Elman, engineer obsessed after high cholesterol on ketogenic diet. Did N=1 experiments, focused on reason and risk for cardiovascular disease in low-carb high-LDL people. That's why did Keto study.
Dave is best friend, like brother, inspiration. Grew up in academia, often disappointing lack of intellectual integrity. Dave from outside academic medicine, no MD/PhD, yet published more in lipidology than prominent critics. Attacks with ferocity of curiosity, always hard questions. Area of research his brainchild: low-carb hyper-responders with sky-high cholesterol.
Studied 100 lean mass hyper-responders on low carb, LDL through roof. Mean baseline LDL 272, top 0.1%. Last paper: average LDL 254, top 10% of top 1% in NHANES (1 in 1000). Wide spread.
Followed over year with high-resolution coronary CT angiography, assessing calcified and soft plaque progression and predictors.
April 7th JACC Advances paper: ApoB and LDL don't predict plaque progression; baseline plaque predicts future plaque (plaque snowball). Novel finding: LDL/ApoB no association. Measured progression over year larger than expected based on initial report.
Semi-quantitative (horseshoes and hand grenades scoring, quick). Quantitative Clearly. Robust findings: plaque predicts plaque; LDL/ApoB no association in both.
Concern: difference between semi-quantitative and Clearly. Usually correlate closely in other studies. In Clearly, majority (55/100) under PAV 1.0, low progressors.
Complicated: PAV percent atheroma volume (established cut points, most not rapid). Non-calcified plaque volume (absolute/percent change). Different metrics differ.
Semi-quantitative total plaque score: 6-7 rapid progressors (score change 3+ out of 100). Total possible score 45 (0-3 per 15 points). No double-digit advancement, no +10. Expected in middle-aged population.
Clearly AI read not focus of initial manuscript. Initial: plaque predicts plaque, ApoB/LDL don't. Prespecified Qangio for precise quantification (labor-intensive, in methods paper). All post-semi-quantitative are quantitative; Qangio longest.
Publishing as data comes in. Agreed novel findings: plaque predicts plaque; LDL/ApoB no association (consistent across analyses, largest LDL spread in prospective imaging study ever).
Discrepancies noted pre-publication but funders blocked from raw data for integrity.
Affiliated via Citizen Science Foundation funding. Blocks protected integrity, no thumb on scale. Funder gets raw anonymized data only post-publication (contractual with Lundquist Institute). Appropriate: funders shouldn't access raw data pre-pub.
Recent paper on GLP/GIP/glucagon agonist (trillion-dollar drug, e.g., retatrutide). Misinfo on lean mass preservation despite weight loss. Drug company (Eli Lilly) funded, analyzed data, interpreted, wrote manuscript.
Dave created integrity barriers, but entrusts data to others, validation post-facto.
Proud of embracing critics: percent change non-calcified plaque volume missing numeric value (not excited about metric; low baseline denominator skews, no good thresholds). Field now focuses on PAV.
Correction issued promptly, interviewed facing tough questions. Good scientist owns mistakes.
Lundquist delivered anonymized dataset to Citizen Science Foundation. Compiled 24-issue report in 24 hours. Brought immediately to Lundquist and Clearly.
Secured meeting with Clearly CMO (author on paper, undisclosed). Thought beneficial to resolve anomalies.
Clearly dataset anomalies vs. other longitudinal imaging:
- No regressors: everyone had more plaque at follow-up, no less. Expected bidirectional scatter (noise, device error, especially low baseline plaque). Even high-risk group should have noise below floor (bathroom scale example: can't distinguish 1000 vs 1001 marbles reliably).
If you weigh a bag of 1000 marbles and a bag of 1001 marbles a hundred times, you expect times where the 1001 marbles show less weight due to measurement error below the noise floor.
Power of zero: Zero calcification in coronary arteries (CAC scan or CT angiogram) indicates substantially lower likelihood of future events or plaque progression compared to those with calcification, per established literature.
In the Clearly dataset, this is reversed: participants with no calcification had 3.5 times more plaque progression than those with calcification—the opposite of prevailing literature, with huge magnitude.
These patterns contradict existing literature. Two possibilities: massive bias in dataset or all high-risk participants (not prespecified methodology).
Citizen Science Foundation funded additional independent fully blinded analyses:
- Qangio (prespecified methodology)
- HeartFlow (completely blinded)
Both agreed with each other and disagreed with initial Clearly dataset progression. Now one outlier dataset (Clearly) vs. entire literature and two independent datasets.
Initial Clearly analysis was not properly blinded. Blinding means analysts know scan pairings belong to a subject but not the order (baseline vs. follow-up).
Unblinding risks bias: If analyst believes lipid hypothesis, sees high LDL, baseline left/follow-up right, and measures lower progression on follow-up, they might unconsciously reanalyze the second scan more closely or adjust tools.
This applies to humans or AI (e.g., neural networks trained on chronology). Example: AI trained on skin lesions got 100% accuracy by detecting rulers left on melanoma images (known cancerous cases).
Unblinded read introduces enormous bias potential, human or AI.
From first anomalous report, Citizen Science Foundation repeatedly emailed Clearly requesting fully operationally blinded analysis as quality control, offering to pay all costs. Clearly refused.
Timeline: Anomalies identified before unblinding revealed (June); HeartFlow commissioned as refusal became obvious; Qangio completed August.
Behind scenes: Gave Clearly every opportunity over months, accumulated evidence (initial report + independents), involved lawyers, before deciding to withdraw.
Why wouldn't they check their work? Unblinded, anomalies, disagrees with independents—no good reason.
Clearly's response: Stand by independent analyses outside their study as validation (but those disagree with theirs).
HeartFlow engaged after Clearly refusal (not originally planned). Confirmed upfront they'd do quality control repeats if anomalous, fully blinded—no hesitation, emphasize accuracy as core product.
Even AI (e.g., Grok) deems blinding non-negotiable for prospective coronary scan studies.
Clearly has record of blinded analyses (2023 paper in JACC journal, James Earls senior author).
No reason why Clearly didn't blind initially or refuse compensated reanalysis when questioned.
Lost confidence in Clearly analysis post-publication (April 7). Requested withdrawal (synonymous with retraction).
Approved January 12; as of March 10 recording, not yet live (behind-scenes issues).
Gave accumulated evidence to Clearly; no quibbling—Clearly analysis specifically unreliable.
No one defends Clearly dataset anymore (post-Qangio).
Participants requested their scans, submitted via personal cardiologists with valid Clearly accounts (outside study, above board).
19 submissions total; 8 had both baseline/follow-up.
Study Clearly data: ~20 mm³ non-calcified plaque progression (median).
New Clearly reanalysis (blinded, clinician version): median change 0.7 mm³ (near flat), mean is regression (4/8 regressors, 50%).
Initial study: 0/100 true regressors.
Probability of 50% regressors from initial dataset in random 8: effectively zero (analogy: blindfolded punt of ping-pong ball in hurricane into red solo cup on first try).
All independent analyses agree against unblinded Clearly:
- HeartFlow (blinded)
- Qangio (prespecified)
- Participant-led Clearly reanalysis
Pulled every stop for reanalysis; participants deserve it (put hearts on line).
Proud of integrity focus; lessons learned for future (e.g., partnering with HeartFlow).
Speakers' reputations and health on line (e.g., TC ~700, choices against conventional cardiology despite Harvard MD exposure).
Goal: Justice for patients like themselves.
- 5-year (or 6-year) scans for original Keto CTA study (funding locked; some participants ready post-IRB).
- Larger companion study: More relaxed eligibility, control group, fundraising via COI conference (expected closed by video release).
- 3 new papers in production/under review.
- Movie: "The science isn't settled."
Growing interest: Influential international doctor wants independent study in their country.
If I'm the match, you are the powder keg.
Can't rely solely on journals/cardiologists network.
Some researchers are very interested. But that will not carry the day. That's why we need things like this documentary. We need both in play. We need the research, but we also need the social media campaign to carry the research and the good data forward.
To that extent, I'm just so thankful that you and all of your brilliance and all of your hard work have taken you to where you are now. And I can't wait to see where you go next as well.
Keep Nick Norwitz in your library
Save the videos and channels worth coming back to, and find them again in one place.
