The harm-channel on the same exposure whose CV-event benefit is dose-dependent: marine omega-3 supplementation raises the risk of incident atrial fibrillation, and the risk rises with dose. This is the off-target counterweight to the ischemic benefit held on Vitamin and Mineral Supplements for Disease Prevention (VITAL null at 1 g/d; REDUCE-IT positive at 4 g/d) and Fish and Seafood Consumption (Lombardi 2020 CV-NMA). It does not touch the fish-as-food recommendation — this is an isolate-supplement finding. (inferred from Gencer et al., 2021)
The effect — dose-dependent AF risk
Gencer 2021, a gold SR+MA of 7 marine omega-3 CV-outcome RCTs (81,210 patients, weighted mean age 65, 39% female, weighted median follow-up 4.9 y), pooling incident/adverse-event AF:
«The weighted average follow-up was 4.9 years. In meta-analysis, the use of marine omega-3 fatty acid supplements was associated with an increased risk of AF (n=2905; HR, 1.25 [95% CI, 1.07-1.46]; P=0.013). In analyses stratified by dose, the HR was greater in the trials testing >1 g/d (HR, 1.49 [95% CI, 1.04-2.15]; P=0.042) compared with those testing <=1 g/d (HR, 1.12 [95% CI, 1.03-1.22]; P=0.024; P for interaction <0.001).» (Gencer et al., 2021)
«In meta-regression, the HR for AF increased per 1 g higher dosage of omega-3 fatty acids dosage (HR, > 1.11 [95% CI, 1.06-1.15]; P=0.001).» (Gencer et al., 2021)
- Overall: HR 1.25 (1.07-1.46) — a ~25% relative increase in AF (Gencer et al., 2021). Gencer computes RRs from cumulative incidence and states no annual event rate, so a pooled per-year base rate is NOT extractable from it; the absolute excess is nonetheless small because the base rate is low (inferred from Gencer et al., 2021)
- Dose split (prespecified 1 g/d cutoff): high-dose >1 g/d 1.49 (1.04-2.15) vs low-dose <=1 g/d 1.12 (1.03-1.22), P-interaction <0.001 — the interaction stratification collapsed the between-trial heterogeneity (overall I2=54.6% -> low-dose I2<0.01%, high-dose I2=9.9%), which is itself evidence the dose axis is real. (Gencer et al., 2021)
- Dose-response slope: per 1 g/d, HR 1.11 (1.06-1.15) over a studied range of 1-4 g/d (residual I2=0.00%). The slope is estimated only within 1-4 g/d — not below 1 g/d, and not above 4; read it as monotone-over-the-studied-range, not a law to extrapolate. (Gencer et al., 2021)
The trial-level forest — where the dose signal lives
| Trial | Dose / form | AF HR or RR (95% CI) | Dose stratum |
|---|---|---|---|
| VITAL Rhythm | 1 g/d EPA+DHA | 1.09 (0.96-1.24) | low |
| ASCEND | 1 g/d EPA+DHA | 1.23 (0.98-1.54) (rate ratio) | low |
| GISSI-HF | 1 g/d EPA+DHA | 1.10 (0.96-1.25) | low |
| RP (Risk & Prevention) | 1 g/d EPA+DHA | 1.22 (0.93-1.61) | low |
| OMEMI | 1.8 g/d EPA+DHA (elderly post-MI) | 1.84 (0.98-3.45) | high |
| STRENGTH | 4 g/d EPA+DHA (carboxylic acid) | 1.69 (1.29-2.21) | high |
| REDUCE-IT | 4 g/d icosapent ethyl (EPA-only) | 1.35 (1.11-1.65) (RR) | high |
(OMEMI is listed at 1.84 in the source’s forest plot but 1.94 in its discussion text — same CI 0.98-3.45, P=0.06, non-significant either way; the forest value is used here. (Gencer et al., 2021))
The two statistically-significant single-trial signals (STRENGTH, REDUCE-IT) are both high-dose; every low-dose trial is individually non-significant, positive-leaning, and only the pool clears significance at 1.12.
The dose-vs-formulation confound — and why it is NOT purely a REDUCE-IT/mineral-oil artifact
The high-dose arm is dominated by EPA-only REDUCE-IT (4 g icosapent ethyl vs a contested mineral-oil placebo) plus STRENGTH and OMEMI, so dose is entangled with formulation and with a sicker stratum. But the overall AF signal is not an artifact of REDUCE-IT alone — it survives dropping it:
«The effect of marine omega-3 fatty acid supplements on the risk of AF events after excluding REDUCE-IT (the only trial without DHA) remained significant (HR, 1.23 [95% CI, 1.02-1.49]; P=0.038).» (Gencer et al., 2021)
and dropping the AF-hospitalization-only RP trial (HR 1.26, 1.04-1.52). So the AF harm is a class effect, not one trial’s placebo controversy — but Gencer cannot separate whether the steeper high-dose slope is dose or EPA-only formulation (no trial randomized dose head-to-head; the >1 g/d arm and the EPA-only arm are the same trials). REDUCE-IT’s own prior report was the seed of the concern:
«REDUCE-IT also reported an increase in a prespecified tertiary outcome of AF hospitalization in those randomized to active treatment compared with placebo (3.1% versus 2.1%; P=0.004).» (Gencer et al., 2021)
Lombardi 2021 — a second, time-adjusted MA corroborates the AF direction (type-F, 2026-09-17)
Lombardi 2021 (an EHJ-CVP correspondence-format updated pairwise MA, systematic search to Nov 2020) pooled 5 marine-omega-3 CV-outcome RCTs (REDUCE-IT, Alpha-Omega/Kromhout, Risk & Prevention, STRENGTH, OMEMI) for incident AF, in patients at high CV risk or with established CVD and elevated plasma triglycerides:
«In the random effect model, O3FA supple- mentation was associated with an increased risk of incident AF as compared with placebo [IRR 1.37, 95% CI (1.22–1.54), P < 0.001] (Figure 1). There were no significant statistical heterogeneity between studies and no publica- tion bias, even if the funnel plot suggested some asymmetry.» (Lombardi et al., 2021)
Adding VITAL-Rhythm (from the AHA-2020 abstract) as a sensitivity analysis held the signal at IRR 1.29 (95% CI 1.13-1.48, P=0.0002) — same direction, same population. (Lombardi et al., 2021)
The metric choice is the type-F contribution. Lombardi deliberately reports an incidence-rate ratio (Poisson person-time), not RR/HR:
«Incidence rate ratio (IRR) and 95% confidence interval (CI) were chosen over relative risk as outcome measure because of the different follow-up of the selected studies.» (Lombardi et al., 2021)
This lands exactly on the axis the page already flags as a Gencer limit — Gencer computes RRs from cumulative incidence and reports no annual event rate, so its pooled figure is not follow-up-adjusted. Lombardi’s IRR accounts for person-time across trials of unequal duration, so the composite (a follow-up-adjusted IRR alongside Gencer’s cumulative-incidence HR, both landing ~+25-37%) is a better-specified harm metric on the time axis than either alone. (inferred from Lombardi et al., 2021)
Per-trial IRRs are figure-borne, not extractable. The forest plot (Figure 1) carries the per-trial estimates; the readable text gives only the two pooled IRRs above. Do not read per-trial numbers off this source’s text layer. (inferred from Lombardi et al., 2021)
Parameter commensurability — Gencer HR vs Lombardi IRR are NOT the same quantity
| Parameter | Gencer 2021 | Lombardi 2021 | Same quantity? |
|---|---|---|---|
| Effect metric | HR from cumulative incidence | IRR from Poisson person-time regression | NO — cumulative-risk ratio vs event-rate ratio; unadjusted vs follow-up-adjusted for exposure time |
| Pooled point estimate | 1.25 (1.07-1.46) | 1.37 (1.22-1.54); 1.29 (1.13-1.48) +VITAL | same DIRECTION, not poolable |
| Trial constituency | 7 RCTs (VITAL-Rhythm, ASCEND, GISSI-HF, RP, OMEMI, STRENGTH, REDUCE-IT) | 5 RCTs (REDUCE-IT, Alpha-Omega, RP, STRENGTH, OMEMI); +VITAL-Rhythm in sensitivity | overlapping (4/5 shared), not identical |
| Population | high CV risk / established CVD | high CV risk or established CVD + elevated TG | ~same |
| Outcome | incident/adverse-event AF | onset of AF | ~same |
The effect-size row is NO: Lombardi corroborates the direction (~+25-37% AF risk) with a time-adjusted metric; the magnitudes are not pooled or ranked as if identical. (Gencer et al., 2021; inferred from Lombardi et al., 2021)
Dose nuance — a refinement of Gencer’s formal dose-gradient, NOT a tension
Lombardi remarks qualitatively that the AF signal appears at lower doses too:
«Even lower doses of O3FA supplementation has shown a potential signal for increased risk for AF,2,4,5 proposing that regardless of the dose of O3FA supplementation implemented in the trials and potential-related benefits on CV outcomes, O3FA supplementation may in- crease the risk for AF.» (Lombardi et al., 2021)
Read against Gencer’s formal dose split (low-dose 1.12 vs high-dose 1.49, P-interaction <0.001) this
looks like a softening — the harm is not confined to high dose. But the two are not joined and this
is not filed as a [[tension]]: Lombardi ran no dose-stratification and no dose-interaction test in
this 5-trial pairwise correspondence, so the remark is a qualitative observation at a different
analytic grain (a narrative aside vs Gencer’s formal meta-regression / interaction estimate) — not the
same quantity, so it cannot contradict Gencer’s interaction. The two are in fact consistent: Gencer’s
low-dose arm is itself positive (1.12, CI excludes 1), so even low doses signal and the slope is
steeper at high dose are the same data read at two grains. Record it as an open refinement on the
dose question — the harm may be present across the studied dose range even as it rises with dose —
explicitly not a competing dose-interaction estimate.
(inferred from Lombardi et al., 2021)
The AF mechanism is not settled either way: Lombardi notes «the mechanisms through which O3FA supplementation may increase the risk for AF remain largely unknown» (Lombardi et al., 2021), observing that omega-3 has elsewhere been shown to protect against (ventricular) arrhythmias while some studies report higher post-operative AF with elevated omega-3 — so direction on AF specifically has no clean mechanistic warrant.
COI — clean, a relative strength
Only one author discloses (unrelated) industry ties, and no omega-3 manufacturer funded the work:
«Conflict of interest: G.B.-Z. has consulted for Cardionovum, Innovheart, Meditrial, and Replycare.» (Lombardi et al., 2021)
«The remaining authors have nothing to disclose.» (Lombardi et al., 2021)
Carbone’s funding is an AHA Career Development Award + NIH/VCU (non-commercial); Biondi-Zoccai’s consultancies are interventional-device / CRO firms unrelated to omega-3 products. Under symmetric standards this cuts both ways: a clean-COI MA on a harm finding removes one bias channel and is a relative strength versus any industry-funded omega-3 benefit MA. (inferred from Lombardi et al., 2021)
Synthesis — a same-exposure benefit/harm split by dose (type-A/G)
The decision-relevant move this opens (present in no single held page before it): omega-3 supplementation carries a dose-dependent AF harm that runs opposite its dose-dependent CV-event benefit. The held benefit evidence (Lombardi 2020 CV-NMA on Fish and Seafood Consumption; REDUCE-IT/VITAL on Vitamin and Mineral Supplements for Disease Prevention) puts the ischemic benefit above 1 g/d, concentrated in high-risk statin-treated strata; Gencer puts the AF harm on the same rising dose arm. So the two curves cross the same dose axis in opposite directions:
- General, replete adult on OTC ~1 g/d: the CV benefit is near-null (VITAL/low-dose stratum) and the AF harm is small (1.12, a small absolute excess given the low base rate) — a small-lever decision on both sides. The AF harm does not by itself contraindicate a low-dose capsule, but there is little benefit to weigh it against.
- High-dose prescription (4 g icosapent ethyl for high-TG established-CVD): the ischemic benefit is real (REDUCE-IT MACE ~25% down) but is bought with the 1.49 AF risk plus bleeding — a genuine net-effect trade, not a free lever. Gencer’s own framing:
«Because the benefit of omega-3 fatty acids also appears to be dose-dependent, the associated risk of AF should be balanced against the benefit on atherosclerotic cardiovascular outcomes.» (Gencer et al., 2021)
The clinical-implication the authors draw is a disclosure, dose-graded:
«The potential risk of developing AF should be discussed with the patients when prescribing marine omega-3 supplementation, especially when prescribing a higher dosage.» (Gencer et al., 2021)
(inferred from Gencer et al., 2021)
NOT independent of the held omega-3 evidence (shared trial base + authors)
This MA is not an independent type-E corroboration of the held omega-3 CV evidence — it re-pools the same trials and shares authors:
- Shared trial base: «Of those, 5 trials were already included in a previous meta-analysis of cardiovascular RCTs» (Gencer et al., 2021) (Hu 2019, the same pool the held CV-benefit evidence draws on), and its 7 trials include the held VITAL (Manson 2019), REDUCE-IT (Bhatt 2019), and the trials Lombardi 2020 pools.
- Shared authorship: Cook, Manson, and Albert are VITAL/VITAL-Rhythm investigators and co-authors here. The AUTHOR-LIST DIFF and MA-CONSTITUENCY tests both fail for independence.
So its relationship to the CV-benefit finding is a benefit/harm distinction on the same exposure,
not opposed claims (not a filed [[tension]]) and not independent backing — do not stamp
[E-independent]. It is the first source to weave the AF-harm cell, which was previously only a
one-line side-effect note (Lombardi’s AF 1.35 side-effect line on the Fish page; REDUCE-IT’s AF-hosp
3.1% vs 2.1%).
(inferred from Gencer et al., 2021)
Lombardi 2021 — also NOT clean type-E, on either axis
Lombardi 2021 does not add independent confirmation either, for two distinct reasons:
- vs Gencer: the author lists are disjoint (no shared author — Lombardi/Carbone/Del Buono/ Chiabrando/Vescovo/Camilli/Montone/Vergallo/Abbate/Biondi-Zoccai/Dixon/Crea vs Gencer/Djousse/ Al-Ramady/Cook/Manson/Albert), so Lombardi is author-independent of Gencer — but its trial base is heavily shared: 4 of its 5 primaries (REDUCE-IT, RP, STRENGTH, OMEMI) are among Gencer’s 7, and VITAL-Rhythm (also in Gencer) re-enters in Lombardi’s sensitivity arm; only Alpha-Omega/Kromhout is not in Gencer. Re-pooling largely the same primaries is not evidence-independent, so it does not raise confidence the way a genuinely separate dataset would.
- vs the held Lombardi 2020 CV-NMA: this 2021 paper is the same core group (Lombardi, Chiabrando, Vescovo, Del Buono, Carbone, Abbate, Biondi-Zoccai, Dixon are on both) and cites the 2020 paper as its ref 2. Across the wiki’s omega-3 holdings it is therefore not independent of Lombardi 2020 either.
So do NOT stamp [E-independent] and do NOT treat Lombardi 2021 as independent corroboration. It enters
as a type-F second MA: author-independent-of-Gencer but trial-base-shared, and same-group-as-held-
Lombardi-2020 — refining the harm metric on the time axis (the IRR above), not adding an independent
confirmation.
(inferred from Lombardi et al., 2021)
Limits
- AF ascertainment is heterogeneous and mostly not adjudicated. «differences exist in the AF outcome assessment between trials. In ASCEND (main article), REDUCE-IT, and RP trials, participants with preexisting AF were not systematically excluded from the analysis, whereas in other trials, the analysis reported new-onset AF events or excluded participants with preexisting AF.» (Gencer et al., 2021) AF outcomes were centrally adjudicated in only VITAL and OMEMI; prespecified in only VITAL, OMEMI, STRENGTH. The sensitivity analysis restricting to new-onset AF held (HR 1.26, 0.85-1.87; no interaction with the not-excluded set 1.28, 1.11-1.48).
- Subclinical-detection inflation in the elderly signal. The largest absolute AF risk difference (OMEMI, >3%, in those >=75 y) came from a trial that «also used monitoring to capture AF events; thus, it is likely that more asymptomatic, subclinical cases were detected» (Gencer et al., 2021) — more ascertainment, not necessarily more clinically-actionable AF.
- Aggregate, not individual-patient, data — «this report is unable to undertake subgroup analysis by age or other patient-level characteristics» (Gencer et al., 2021), so the stratum flip (elderly, cardiac-morbid) is a hypothesis, not an estimated effect-modification.
- Dose vs formulation not separable (above); the studied dose range is only 1-4 g/d, so the slope says nothing below 1 g/d or above 4.
- Lombardi 2021’s own stated limits (which apply to its IRR): no systematic haemorrhagic-risk assessment; «the lack of a systematic search for AF events in the individual studies»; some trials did not prespecify AF, «potentially resulting in under-reporting of AF-related events»; a study-level (not IPD) MA; placebo arms differed across trials (though no heterogeneity was found); and VITAL-Rhythm entered only via the AHA-2020 abstract, «as the fully data have not been disclosed yet». (Lombardi et al., 2021) It is a brief correspondence (tier high, not gold) — a 2-page updated MA, not a full PRISMA systematic review.
- Coherence, not validity (R1): the loop is open — no operation here grades the omega-3-dose -> AF association against a realized patient outcome, and AF is a mixed endpoint (hard arrhythmia burden vs subclinical/monitoring-detected).
Self-critique [run 2026-09-02, before commit — page minted]
- Not overclaimed. The AF risk is stated as an association with a dose-gradient, never a proven dose-causal law; the dose-vs-formulation confound and the studied-range limit on the meta-regression are stated explicitly. The general-adult decision-change is correctly the small one on both sides, and the high-dose prescription trade is framed as balance, not prohibition. No fish oil causes AF blanket reading survives; the food-vs-supplement firewall is stated.
- Not a fake tension. The relationship to the held CV-benefit evidence is filed as a benefit/harm
distinction on the same exposure (not-joined guard (ii): benefit and harm are different outcomes
on one exposure, not opposed answers to one question) — no
[[tension]]filed, nocontradicts:edge. - Not laundered-E. Independence is explicitly denied via the shared-trial-base (5/7 in Hu 2019;
includes held VITAL/REDUCE-IT) and shared-authorship (Cook/Manson/Albert) tests; no
[E-independent]token; confidence held atmedium(single gold MA, consistent dose-gradient and surviving-REDUCE-IT sensitivity, but heterogeneous unadjudicated AF ascertainment and the dose/form confound).
Self-critique [run 2026-09-17, before commit — Lombardi 2021 IRR MA folded (2nd source)]
- Not overclaimed on the corroboration. Lombardi’s IRR 1.37 / 1.29 is stated as direction
corroboration of Gencer, not as an independent confirmation and not pooled with Gencer’s HR. The
parameter table marks the effect-size row NO (cumulative-incidence HR vs person-time IRR), so the two
magnitudes are never summed or ranked as the same quantity. Confidence deliberately held at
medium, not bumped: the second MA re-pools a largely-shared trial base (4/5 primaries overlap Gencer) and is same-group as the held Lombardi 2020, so it corroborates without independently confirming — a bump would be exactly the laundered-independence error. - Independence denied on both axes, accurately. Author-independent-of-Gencer (disjoint lists) but
trial-base-shared; and same-core-group as the held Lombardi 2020 CV-NMA (cites it as ref 2). The
earlier draft claim that all 5 Lombardi trials are a subset of Gencer’s 7 was corrected against the
source: Alpha-Omega/Kromhout (ref 4) is in Lombardi but not Gencer’s forest, so it is 4/5 overlap, not
a strict subset — stated as 4/5 throughout. No
[E-independent]token;sources:gains Lombardi only because a distinct refining claim (the time-adjusted IRR + its limits) lives in the body (dual test). - Dose nuance is NOT a fake tension. Lombardi’s regardless of the dose is a qualitative aside with
no dose-stratification or interaction test, at a different analytic grain from Gencer’s formal
P-interaction<0.001 split — not joined (not the same quantity), and consistent with Gencer once read
at two grains (Gencer’s low-dose arm is itself positive). Filed as an open refinement, not a
[[tension]], nocontradicts:edge. The mechanism-unknown caveat is carried so the harm direction is not dressed with unwarranted mechanistic support. - COI note is symmetric, not a halo. The clean-COI point is stated as removing one bias channel on the harm direction, not as raising the finding’s grade; it is applied under symmetric standards (a strength here would be a weakness in an industry-funded benefit MA), not as special pleading for a result the page favours.