A number can be wrong in two different ways. It can be inaccurate, which is a problem of measurement. Or it can be presented with more confidence than the underlying data can support, which is a problem of communication. April 2026’s literature is mostly free of the first failure and unusually full of the second.
Consider three numbers from this month’s selections. A Nature Medicine modeling study attributes 4.06 million ischemic heart disease deaths to suboptimal diet — with a 95% uncertainty interval running from 0.74 million to 6.22 million, an eight-fold span. A meta-analysis reports that adolescents eating more ultra-processed food have 63% higher odds of overweight or obesity, drawn from studies whose heterogeneity statistic reaches 97.58%, a figure so high that the pooled estimate is barely a coherent single quantity. A randomized trial of a herbal formula in Alzheimer’s reports its primary outcome as “P < 0.05” and never states the effect size at all, leaving a clinician unable to judge whether the improvement is meaningful or merely detectable.
None of those papers is fraudulent or even careless. In each case the authors reported their uncertainty honestly — the interval is right there in the abstract, the I² is stated plainly. What happens next is the problem: the point estimate travels and the interval does not. “Four million deaths” is a headline. “Somewhere between 0.74 and 6.22 million, depending on assumptions about what an optimal diet is” is not.
This month also offers the counter-example, and it is worth reading for contrast. A meta-analysis of oral nutritional therapy in cirrhosis pooled 3,128 patients across 18 trials and returned an I² of zero for three of four clinical outcomes. That is what a tight estimate looks like. The distance between that paper and the ones above is the subject of this issue.
Studies at a Glance
Reading the Interval, Not the Estimate
Three papers whose point estimates are quoted widely and whose uncertainty is quoted almost never. In each case the authors reported the imprecision plainly; it simply did not survive the trip from abstract to practice recommendation.
The Global Burden of Disease programme produces the numbers that anchor most population-level nutrition advocacy. This analysis narrows the lens to ischemic heart disease across 204 countries and 33 years. Understanding what kind of number it produces — and what kind it does not — matters more than the number itself.
Comparative risk assessment within the GBD 2023 framework, covering 204 countries from 1990 to 2023. Thirteen dietary factors were modeled, including low intake of fruits, vegetables, whole grains, nuts and seeds, and seafood omega-3s, and high intake of red meat, processed meat, sugar-sweetened beverages, trans fats, and sodium. Funded by the Gates Foundation, the Australian National Health and Medical Research Council, and Queensland Department of Health.
Show full methodology
This is not an observational study of individuals. No patients were followed and no diets were measured in the population described. The output combines three modeled inputs: estimated dietary exposure per country-year (often imputed where survey data are sparse), relative-risk estimates imported from the separate observational epidemiology literature, and a theoretical minimum risk exposure level — a counterfactual “optimal” diet against which every country’s intake is compared. Uncertainty intervals derive from probabilistic simulation propagating uncertainty across all three, which is conceptually different from a confidence interval around a measured statistic. The authors concede reliance on observational evidence, uneven data quality, and possible unmeasured factors.
Leading individual contributors, expressed as deaths per 100,000: low nuts and seeds 9.87 (95% UI 2.84–17.12), low whole grains 9.22 (4.73–13.67), low fruit 7.25 (1.54–13.34), and high sodium 7.15 (0.92–17.97). Burden concentrated in low- and middle-SDI countries. Every one of these intervals is extraordinarily wide relative to its point estimate — the sodium lower bound is roughly 8% of its central value. The age-standardized attributable death rate fell substantially over the study period, which is the genuinely encouraging finding and the one least often quoted.
Strengths
- 204 countries across 33 years
- Disaggregated by specific dietary factor rather than a composite
- Uncertainty intervals reported transparently rather than suppressed
- Consistent with the broader GBD dietary-risk literature
- Public and philanthropic funding
Limitations
- A modeled counterfactual, not measured mortality
- Exposure data sparse or imputed for many country-years
- Relative risks imported from observational epidemiology, inheriting its confounding
- The “optimal diet” benchmark is an assumption, not an empirical finding
- Uncertainty intervals span up to eight-fold
- Says nothing about any individual’s risk
Nothing here changes clinical counseling. The dietary patterns implicated — more nuts, whole grains, and fruit; less sodium — are already standard advice supported by better-suited evidence. The legitimate use of this paper is motivational and political rather than protocol-level: it is a population-health argument for food policy and affordability, which is how the accompanying commentary framed it. Using it to quantify an individual client’s risk imports a precision that does not exist in the source.
The single biggest misread is treating 4.06 million as a body count — as though researchers counted deaths and traced each to diet. They did not. This is an estimate of how many fewer deaths would be expected if every population’s average diet matched a theoretical benchmark. Quoting the point estimate without the interval overstates the paper’s own epistemic confidence by a wide margin.
Ultra-processed food and adiposity is among the most-cited relationships in contemporary nutrition, but most syntheses have addressed adults. This review claims to be the first focused specifically on adolescents, a population where dietary patterns are still forming and intervention could plausibly matter most.
Systematic review and meta-analysis of 23 observational studies comprising roughly 155,000 adolescents, published 2008–2025 across fourteen countries. Prospectively registered (PROSPERO CRD420251121568) and PRISMA-compliant. Random-effects pooling, heterogeneity by Cochran’s Q and I², publication bias by Egger’s regression with trim-and-fill. Quality appraised by Newcastle-Ottawa Scale, with all studies scoring 6–8.
Show full methodology
The majority of included studies were cross-sectional; two were case-control. Dietary exposure was assessed by food-frequency questionnaire in 19 of 23 studies. Outcomes used WHO BMI-for-age z-scores, CDC growth charts, or IOTF cutoffs, with some studies adding waist circumference. Individual study sizes ranged from 76 to 97,536 participants. The review received no specific funding and the authors declare no conflicts. GRADE certainty was not reported, and the trim-and-fill-adjusted pooled estimate was not stated.
Subgroup estimates diverged sharply: by region, Europe 1.31 (95% CI 1.00–1.70) against Africa 2.98 (1.88–4.73); by design, cross-sectional 1.55 (1.29–1.86) against case-control 4.71 (2.09–10.58). That three-fold spread by study design alone is a warning that the pooled figure is averaging genuinely different things. The most recent studies (2024–2025) produced the highest estimate at 2.09 (1.24–3.51). At I² near 98%, the summary odds ratio is less a single effect than a rough midpoint of a heterogeneous cloud.
Strengths
- Prospectively registered and PRISMA-compliant
- Roughly 155,000 adolescents pooled
- Formal quality appraisal of every included study
- Multi-continent representation
- Publication bias formally tested and reported
- No funding or conflicts declared
Limitations
- I² = 97.58% makes the pooled estimate statistically fragile
- Predominantly cross-sectional — no causal inference
- Egger’s test indicates likely publication bias
- Total energy intake not addressed as a confounder
- Socioeconomic status not addressed as a confounder
- GRADE certainty not reported
- Adjusted trim-and-fill estimate not stated
Reducing ultra-processed intake in adolescents remains reasonable advice on nutrient-density grounds independent of this paper. What this paper cannot support is a quantified risk claim. The unaddressed confounders are the crux: ultra-processed foods are calorically dense and hyperpalatable, so a substantial share of any association could run through simple excess energy intake rather than processing itself, and UPF consumption is strongly patterned by socioeconomic status in most countries represented here. Neither is discussed.
Quoting “63% greater odds” as a precise, causal, generalizable figure is the error. With heterogeneity near 98%, subgroup estimates ranging from 1.31 to 4.71, and two of the most obvious confounders unexamined, this is evidence that ultra-processed intake correlates with adolescent adiposity in a confounded literature — not an established dose-response relationship.
Menopausal clients ask about soy and cognition more than almost any other supplement question, and the existing evidence is genuinely conflicting. The landmark WISH trial found no global cognitive benefit from long-term isoflavone supplementation. This meta-analysis pools a larger and more recent set of trials and, importantly, separates cognitive domains rather than reporting a single composite.
Systematic review of 37 studies with 16 randomized controlled trials pooled in meta-analysis. PubMed, Scopus, and Web of Science searched. Studies rated by quality with 27 judged positive and 9 neutral. Outcomes were a composite cognitive score plus four separate domains: memory, executive function, psychomotor speed, and processing speed.
Show full methodology
PROSPERO registration was not reported and could not be located. Critically for interpretation, the review does not appear to stratify food-based soy intake from isolated isoflavone supplements, nor to harmonize dose ranges — meaning the pooled exposure blends tofu consumption with concentrated genistein-daidzein extracts at unspecified doses. Heterogeneity (I²), publication bias assessment, and GRADE certainty ratings were not reported in accessible text. The stored PubMed record captured no conflict-of-interest or funding statement, which should be read as a gap rather than as an absence of conflicts.
Both significant effects are small by conventional benchmarks, and both carry confidence intervals whose lower bounds sit very close to zero. The narrative synthesis across all 37 studies found 15 positive-significant, 13 non-significant, 7 bidirectional, and 2 negative-significant — a distribution that should temper confidence rather than reinforce it. An effect size of 0.25 describes a population-level signal most individual clients would not subjectively notice.
Strengths
- 37 studies with 16 pooled in meta-analysis
- Domain-specific analysis rather than composite only
- Formal quality appraisal conducted
- Reports the null domains explicitly rather than burying them
Limitations
- Food-based and supplemental isoflavones apparently pooled together
- Dose ranges not harmonized or reported
- I², publication bias, and GRADE all unreported
- Wide confidence intervals relative to point estimates
- Sits in tension with the null WISH randomized trial
- No funding or COI statement captured
For a menopausal client asking about soy for memory, the honest answer is that pooled evidence shows a small effect on memory specifically, no effect on executive function or processing speed, and comes from studies of inconsistent dose and formulation. That supports including soy foods as part of a reasonable diet. It does not support a therapeutic claim or a supplement recommendation. Whether food-based intake generalizes from supplement-heavy trials is not answerable from this review, because the two were not analyzed separately.
“Significant” here means distinguishable from zero, not clinically meaningful — and that distinction decides whether a client should start a supplement. Three of four cognitive domains showed nothing, and the composite effect runs counter to WISH, a well-powered randomized trial that found no global cognitive benefit.
The Value of a Null Result
Two well-designed studies whose most useful findings are the things that did not happen — and where the negative result is more practically important than the positive one buried alongside it.
The hypothesis that dietary emulsifiers drive gut inflammation has moved from mouse models into human trials over the past decade, and has become a fixture of clinical advice about packaged food. This trial does what the field needed: tests five chemically distinct emulsifiers head to head, in the same protocol, against a genuine placebo.
Double-blind, placebo-controlled randomized trial in 60 healthy participants at KU Leuven, Belgium. All participants followed an emulsifier-free diet for a two-week run-in, then continued it for four further weeks while consuming daily brownies containing one of carboxymethyl cellulose, polysorbate-80, carrageenan, soy lecithin, native rice starch, or no additive. Registered NCT06552156.
Show full methodology
Endpoints spanned cardiometabolic markers, gut microbiota composition and alpha diversity, short-chain fatty acids, fecal calprotectin, serum CRP, serum lipopolysaccharide-binding protein, an Olink inflammatory protein panel, and transcellular intestinal permeability. With 60 participants across six arms, roughly ten per arm, the trial is explicitly labeled exploratory by its authors and is underpowered for secondary comparisons. Investigators hold Research Foundation – Flanders grants; no food-industry funding was identified. The dose of each emulsifier relative to typical dietary exposure was not reported in accessible text, which is a material gap.
Alpha diversity remained stable throughout. All short-chain fatty acids were lower with carboxymethyl cellulose than placebo, a pattern mirrored by other emulsifiers without reaching significance. The permeability finding for carrageenan is reported against baseline, not against placebo — a distinction that matters considerably for how much weight it can carry. The largest and tightest effect in the entire trial belongs to the emulsifier-free run-in diet itself, which lowered cholesterol before any randomization occurred.
Strengths
- Genuine double-blind placebo control, rare in emulsifier research
- Five chemically distinct agents tested in one protocol
- Multi-domain endpoints spanning microbiome, permeability, and systemic inflammation
- Objective biomarkers rather than symptom reporting
- Publicly funded with no industry involvement
Limitations
- Roughly ten participants per arm
- Explicitly labeled exploratory by the authors
- Healthy adults only — not an IBD or IBS population
- Carrageenan permeability result is versus baseline, not placebo
- No multiple-comparison correction across five agents and many endpoints
- Dose relative to real-world dietary exposure not reported
- Four weeks cannot address long-term exposure
This does not support blanket emulsifier avoidance. It supports narrower, agent-specific caution: carrageenan for clients where intestinal permeability is a clinical concern, and carboxymethyl cellulose for clients whose protocol depends on short-chain fatty acid production. Because the population was healthy, the null inflammation result cannot be used to reassure clients with existing gut inflammation — that is precisely the group the hypothesis concerns and precisely the group not studied. The run-in cholesterol finding is interesting but reflects a whole-diet change, not any single emulsifier.
Reading this as confirmation that emulsifiers harm gut health inverts the result. The primary systemic findings were null for all five agents. The two positive signals are narrow, exploratory, uncorrected for multiple comparisons across five emulsifiers and many endpoints, and found in healthy people over four weeks.
As GLP-1 and dual GLP-1/GIP agonists move into routine use, the question has shifted from how much weight patients lose to what tissue they lose with it. This is the first systematic synthesis to pool body-composition data across incretin trials using multiple measurement methods rather than DXA alone.
Systematic review of 35 randomized trials of liraglutide, semaglutide, tirzepatide, or dulaglutide in adults with obesity, median duration 26 weeks and median 78 participants. Databases searched January 2003 to February 2026. Prospectively registered on the Open Science Framework. Body composition assessed by bioimpedance, DXA, CT, or MRI. Risk of bias by Cochrane RoB 2.0 with duplicate assessment.
Show full methodology
No formal meta-analysis was performed — heterogeneity precluded pooling, and the authors declined to force it. Findings were synthesized against prespecified benchmarks drawn from the existing literature: roughly 25% of weight loss as fat-free mass by bioimpedance or DXA, and roughly 15% as skeletal muscle by CT or MRI. Only 4 of 35 studies had a mean age of 60 or above, and none focused on adults over 65 — the group most vulnerable to sarcopenic consequences. Half the studies did not report race or ethnicity. Primary funding source: none. Note that the final typeset version reports slightly different figures (36 studies, median 34.9%) than the indexed abstract used here; the qualitative conclusion is identical under both.
By bioimpedance or DXA the median proportion was roughly 29%; by CT or MRI roughly 25.3%, with two-thirds exceeding the 15% benchmark for that method. Comparator arms lost a median of only 2.5% body weight, yet more than a third still exceeded their muscle-loss benchmark. Critically, no included study reported objective physical function — no strength, gait speed, or mobility outcomes anywhere in the evidence base, despite function being the entire reason muscle loss matters clinically.
Strengths
- First multi-method synthesis rather than DXA-only
- Prospectively registered with duplicate extraction
- Declined to force a meta-analysis onto heterogeneous data
- No industry funding despite examining flagship drug products
- Transparent benchmark methodology
Limitations
- No included trial measured strength or physical function
- Most trials 26 weeks or shorter, likely underestimating steady-state loss
- Almost no data in adults over 65
- Only about 43% of studies at low risk of bias
- Over half had no structured resistance training or protein co-intervention
- Indexed abstract and final version report slightly different figures
Counsel every client starting an incretin agent that a quarter to a third of the weight lost should be expected to be muscle-related tissue, and that this warrants active countermeasures — adequate protein and resistance training — from the start rather than after the fact. That protocol is extrapolated from bariatric surgery practice, not tested by this review. What the review cannot tell you is whether any specific protein target or training dose changes the proportion, because no trial has asked. The absence of any functional outcome data across 35 trials is the most actionable gap here.
The tempting misread is that incretin drugs are uniquely damaging to muscle. The paper’s own data undercut that — a substantial share of diet and lifestyle arms exceeded the same benchmark while losing far less weight. Disproportionate muscle loss appears to be a feature of weight loss without deliberate preservation, not of these drugs specifically. They simply make it larger in absolute terms.
Botanical Trials and the Disclosure Question
Two randomized herbal trials from the same journal in the same month, both methodologically respectable, both raising questions their abstracts do not answer — one about funding, one about whether a reported effect has any stated magnitude at all.
Diabetic peripheral neuropathy has few satisfying treatments, and the guideline-recommended options are symptomatic rather than disease-modifying. This is the first adequately powered multicentre placebo-controlled trial of Mudan granule, an NMPA-approved Chinese patent medicine, added to standard mecobalamin therapy.
Multicentre, double-blind, placebo-controlled randomized trial across 13 centres in China. 400 patients randomized 1:1 to Mudan granule 7 g three times daily plus mecobalamin, or matched placebo plus mecobalamin, for 24 weeks. Complete analysis set 357 (89.25%). Primary endpoint was change in Michigan Diabetic Neuropathy Score, analyzed by covariate-adjusted ANCOVA.
Show full methodology
Mudan granule contains nine herbs including Astragalus membranaceus, Corydalis yanhusuo, Panax notoginseng, Salvia miltiorrhiza, and Carthamus tinctorius. No marker-compound standardization was reported — dosing is by granule weight, not standardized active constituent content. Enrollment required not only diabetic neuropathy but a specific traditional Chinese medicine syndrome diagnosis (Qi deficiency and collateral obstruction), meaning the result does not generalize even to all neuropathy patients in China. Missing primary data were handled by last-observation-carried-forward, a dated approach by current standards. Secondary outcomes and adverse event data were not available in accessible text.
Both arms improved; the Mudan arm improved roughly twice as much on the primary score, and the confidence intervals do not overlap. That is a real signal from a properly designed trial. The critical context is funding. The trial’s own published protocol paper states it was funded by Liaoning Aoda Pharmaceutical, the manufacturer of Mudan granule, while the published trial’s conflict-of-interest statement declares no relevant relationships. Industry-sponsored trials of herbal products have consistently reported larger effects than independently funded replications, and that base rate applies here.
Strengths
- Multicentre design across 13 sites
- Adequately powered at 400 randomized
- Double-blind with matched placebo
- Prespecified primary endpoint
- Add-on design against active background therapy
Limitations
- Manufacturer-funded per protocol, while the paper declares no relevant relationships
- No standardization to marker compounds
- Enrollment restricted to a TCM syndrome subtype
- Last-observation-carried-forward for missing data
- Attrition of 10.75% with reasons unreported
- Secondary and safety outcomes not accessible
- No glycemic control data to confirm arm balance
For a US-based practitioner this is not an actionable protocol. Mudan granule is a Chinese domestic patent medicine with no legitimate standardized supply chain here. More importantly, the individual constituent herbs — astragalus, salvia, notoginseng — are widely available, and this trial provides no evidence whatsoever about any of them alone or in a different combination. The result belongs to one manufacturer’s specific nine-herb formulation at its specific preparation, in patients selected by a diagnostic framework most Western practitioners cannot apply.
The headline invites a practitioner to extrapolate to astragalus or salvia products already on the shelf. That fails on two counts: the trial tested one patented multi-herb granule, not its constituents; and it was funded by that granule’s manufacturer while disclosing no relevant relationships — a gap that should raise skepticism about the effect magnitude, not lower it.
Behavioral and psychological symptoms of dementia are among the hardest problems in geriatric care, and the pharmacological options carry real risks. Yokukansan has a substantial Japanese trial literature for this indication, almost all of it using one manufacturer’s standardized preparation. This is an independent replication in a different manufacturing and regulatory context.
Single-centre, randomized, double-blind, placebo-controlled trial at Shanghai University of Traditional Chinese Medicine. 100 patients with Alzheimer’s disease and behavioral symptoms enrolled, 89 completed. Yokukansan 2.5 g three times daily or matched placebo granules for 4 weeks, with assessment at baseline, weeks 2, 4, and 8. Registered ChiCTR2300069404. Primary outcome was change in Neuropsychiatric Inventory score.
Show full methodology
Secondary outcomes included MoCA, MMSE, ADCS-ADL, HAMD-17, and a TCM pattern-differentiation score. Funded by the National Natural Science Foundation of China; no industry funding was identified, and the authors declare no competing interests. The specific product, manufacturer, and standardization method were not reported — and this matters, because essentially all prior Yokukansan evidence uses the Japanese Tsumura TJ-54 preparation. Whether this Chinese-manufactured product matches its composition or marker compounds could not be verified. Adverse event rates were not reported, notable given the formula’s typical licorice content and the associated pseudoaldosteronism risk with prolonged use.
The primary outcome reached significance and the authors report trends toward improvement in daily living and MMSE. No mean scores, between-group difference, effect size, or confidence interval appear anywhere in the accessible report — only a threshold p-value. For a clinically trained reader that is close to uninformative: it establishes that a difference was detectable, and says nothing about whether it was large enough to matter to a patient or a caregiver. Cognition and mood showed no significant change.
Strengths
- Randomized and double-blind, uncommon rigor in this literature
- Prospectively registered
- Validated primary instrument (NPI) is the field standard
- Follow-up extended four weeks beyond treatment
- Government rather than industry funding
Limitations
- No effect size or confidence interval reported for the primary outcome
- Product, manufacturer, and standardization unverified
- Single centre with 89 completers
- Authors themselves call for larger confirmatory trials
- No adverse event data despite licorice content
- Concomitant dementia medication use not reported
This adds a data point supporting Yokukansan as an adjunct for agitation and anxiety in Alzheimer’s, consistent in direction with the Japanese literature. It does not support positioning it as cognitive or antidepressant therapy, since neither improved. If citing a dose, 2.5 g three times daily is what this unverified preparation used, and it should not be presented as interchangeable with a different manufacturer’s product or a compounded decoction. Standard monitoring for hypokalemia and blood pressure applies with prolonged use of licorice-containing formulas — a practice standard, not a finding of this trial.
The headline pairs “Yokukansan” with “Alzheimer’s” and invites the conclusion that the formula treats the disease. It does not. Cognition showed no significant effect; only behavioral symptoms improved, and even that result is reported as a bare p-value with no magnitude attached — leaving no way to judge clinical meaningfulness.
Where the Evidence Actually Holds
Three papers that earn their conclusions, with the caveats that still apply. The first is the tightest estimate in the issue; the other two are rigorous work whose headline claims outrun their own statistics.
Malnutrition in cirrhosis is common, consequential, and under-treated. Individual trials have generally been too small to detect effects on hard clinical outcomes, which is precisely the situation meta-analysis exists to address. This review pools frailty and function outcomes alongside the clinical endpoints.
Systematic review and meta-analysis of 18 randomized trials comprising 3,128 adults with cirrhosis, drawn from 21 identified trials published 2000 to June 2025, each with at least four weeks of oral nutritional therapy. Random-effects modeling with sensitivity, subgroup, and meta-regression analyses. Prospectively registered (PROSPERO CRD420251073182). No funding reported; authors declare no conflicts.
Show full methodology
Interventions pooled branched-chain amino acids, high-protein high-calorie diets, β-hydroxy β-methylbutyrate, and general high-protein diets under one heading — a genuine heterogeneity in intervention type even where statistical heterogeneity was low. Outcomes spanned overt hepatic encephalopathy, decompensation, hospitalization, mortality, handgrip strength, and Liver Frailty Index. GRADE certainty ratings, publication bias assessment, and risk-of-bias tool results were not reported in accessible text.
Decompensation fell similarly (RR 0.69, 95% CI 0.53–0.90, NNT 16, I² = 0%). Functional outcomes were far noisier: handgrip strength improved 2.41 kg (95% CI 0.58–4.24) but at I² = 87%, and Liver Frailty Index improved −0.35 (95% CI −0.70 to −0.0005) at I² = 84.4% — an interval whose upper bound is essentially zero. Higher MELD scores predicted greater mortality benefit. Trials of six months or less showed significant mortality reduction while longer ones did not, but duration was not a significant moderator in formal testing — a distinction that matters.
Strengths
- 3,128 patients pooled across 18 trials
- I² of zero for three of four clinical outcomes
- Prospectively registered
- Numbers needed to treat reported for every clinical outcome
- Meta-regression for effect modifiers
- No funding or conflicts
Limitations
- Pools genuinely distinct interventions under one heading
- Functional outcomes carry I² of 84–87%
- Liver Frailty Index interval effectively touches zero
- GRADE, publication bias, and risk-of-bias results not accessible
- Cannot identify which specific intervention drives benefit
This is the most directly actionable paper in the issue. Oral nutritional support in cirrhosis reduces encephalopathy, decompensation, hospitalization, and mortality on tight pooled estimates across more than three thousand patients, with numbers needed to treat between 8 and 23. Higher-MELD patients derive more mortality benefit. What the paper cannot tell you is which modality to choose — branched-chain amino acids, high-protein high-calorie diets, and HMB are pooled, and the subgroup comparison suggesting high-protein high-calorie superiority is not a randomized head-to-head.
The one trap is the apparent duration effect. Short trials showed significant mortality reduction and longer ones did not, which invites the conclusion that nutritional support is a short-term intervention. Formal meta-regression found duration was not a significant moderator — the pattern is more likely an artifact of fewer, noisier long-duration trials.
Most practitioners carry a working heuristic that omega-6 fatty acids are uniformly pro-inflammatory and should be minimized. This analysis tests that heuristic against the largest available cohort by separating linoleic acid from the remaining n-6 species rather than treating them as one exposure.
Prospective observational cohort within UK Biobank. 273,795 participants free of dementia at baseline, with 5,799 incident cases over a median 15.06 years of follow-up. Plasma linoleic acid and non-linoleic n-6 polyunsaturated fatty acids analyzed separately by quintile and continuously, using Cox proportional hazards models. Conducted under UK Biobank application 85092.
Show full methodology
Arachidonic acid comprises roughly 77% of the non-linoleic n-6 fraction, with five other fatty acids contributing about 5.8% — but the paper analyzes the fraction as a composite rather than reporting per-fatty-acid hazard ratios. Fatty acids were quantified by high-throughput NMR metabolomics. The specific covariate adjustment list was not reported in accessible text. The authors concede the platform is not precise enough to track arachidonic acid accurately, meaning they cannot attribute the higher-risk signal to it specifically, and they state plainly that the study cannot establish causation.
The two associations run in opposite directions with non-overlapping intervals, which is the paper’s substantive contribution: treating omega-6 as a single category obscures a real divergence. In age-stratified analysis the linoleic acid association held among participants aged 50–59 at baseline but not those 60 and over. The funding warrants explicit attention. Support included the Soy Nutrition Institute and the United Soybean Board — soybean oil being the dominant US dietary source of linoleic acid — and the senior author holds stock in a company selling omega-3 and omega-6 blood testing. Neither appears in the conflict statement itself.
Strengths
- 273,795 participants with 15 years of follow-up
- Objective plasma biomarkers rather than self-reported diet
- Separates linoleic acid from other n-6 species
- Both quintile and continuous modeling
- Authors concede the assay limitation candidly
Limitations
- Observational — no causal inference available
- Funded in part by two soy-industry bodies
- Senior author holds stock in a fatty-acid testing company
- Assay cannot track arachidonic acid accurately, per the authors
- Covariate adjustment list not reported
- Reverse causation plausible given dementia’s long prodrome
- UK Biobank carries known healthy-volunteer selection bias
The defensible takeaway is to stop treating omega-6 as a monolithic category in client education. Plasma linoleic acid, which is essentially diet-derived since humans cannot synthesize it, tracked lower dementia risk; the remaining n-6 fraction, which reflects both diet and genetically variable endogenous synthesis, tracked higher. That is a meaningful refinement. What it does not license is a dietary recommendation to increase linoleic acid intake — this is biomarker association data, not a trial, and the authors themselves frame intake recommendations as future work.
Two misreads to avoid. The first is treating this as vindication that seed oils prevent dementia; it is observational, reverse causation is plausible over a 15-year window, and the authors cannot confirm which fatty acid drives the harmful signal. The second is reading the funding as disqualifying rather than as context — the finding may well be right, but a soy-industry-supported study producing a linoleic-acid-favorable result deserves independent replication before it reshapes practice.
Choline intake falls below the adequate intake level for a large share of the population, and there is no validated way to identify who is deficient. A blood test that reliably reflected habitual intake would be genuinely useful for pregnancy, plant-exclusive diets, and genetic variant carriers. This trial tests whether plasma markers can do it under ideal conditions.
Double-blind, randomized, crossover controlled feeding trial at UNC Chapel Hill. Participants completed three 15-day arms providing choline at 550, 275, and 137.5 mg/day — 100%, 50%, and 25% of adequate intake — with at least two weeks washout between arms and betaine matched across all arms. Registered NCT03726671. Final analytic sample 73–75 participants.
Show full methodology
Plasma metabolites were quantified by stable-isotope-dilution LC-MS/MS with defined limits of detection and interassay CV under 10%. Compliance was verified objectively by urinary riboflavin rather than self-report. Liver fat was measured by FibroScan controlled attenuation parameter. Enrollment figures differ between sources — the trial registry reports 101 enrolled while the paper describes 96 randomized — and the discrepancy is unreconciled. Multiple-comparison correction was applied inconsistently: false-discovery-rate correction for the isotope analyses, Tukey elsewhere, and raw uncorrected p-values for some secondary findings. Funded by NIH/NIDDK; two authors including the first author report support from Balchem Corporation, a commercial choline manufacturer.
Plasma choline and betaine both tracked intake level robustly in mixed-effects models — betaine especially, at −24.05 (95% CI −27.72 to −20.39) for the lowest intake arm against the highest. Phosphatidylcholine showed no significant difference at any intake level. The headline claim that combining markers improves discrimination is not statistically supported by the paper’s own bootstrap test. In the liver-fat analysis, 41.6% of participants showed at least a 10% increase in controlled attenuation parameter on the lowest choline arm, though with no consistent group-level effect.
Strengths
- Double-blind randomized crossover with objective compliance verification
- Validated LC-MS/MS assay with defined detection limits and QC
- Betaine matched across arms to isolate the choline signal
- Adequately powered against the a priori calculation
- Carryover and sequence effects formally tested and absent
- Primary analyses prespecified and conducted blinded
Limitations
- Combined-panel improvement not statistically significant (P = 0.055–0.080)
- Validated only under controlled feeding, not free-living conditions
- Healthy volunteers with liver and renal disease excluded
- Multiple-comparison correction applied inconsistently
- Enrollment figures differ between registry and paper
- First author reports support from a choline manufacturer
- Sample skews predominantly White from a single site
This is analytic validation, not clinical validation, and the distinction decides how it should be used. The markers track a controlled dose under compliance-verified conditions; whether a single blood draw in a free-living patient reflects habitual intake is explicitly the authors’ stated next step, not something demonstrated here. Do not adopt plasma choline testing into practice on this basis. The incidental liver-fat observation is worth noting — over 40% of participants showed early hepatic fat signals within 15 days at the lowest intake — but with high individual variability and no group-level effect.
Treating an AUC of 0.85 as proof that a blood draw can now assess a patient’s choline status is the error. This was validated under an artificial, isotope-supported, compliance-monitored 15-day protocol in healthy volunteers, says nothing about free-living accuracy, and nothing at all about clinical outcomes. The combined-panel advantage the abstract highlights did not actually reach significance.
Synthesis & Emerging Themes
Heterogeneity Is the Number That Should Travel
Set this month’s two largest syntheses side by side. The cirrhosis nutrition meta-analysis pooled 3,128 patients and returned an I² of zero for three of four clinical outcomes — meaning the included trials were estimating the same underlying effect, and the pooled figure means something. The adolescent ultra-processed food meta-analysis pooled roughly 155,000 participants and returned an I² of 97.58%, with subgroup estimates ranging from 1.31 to 4.71. The second study is far larger and far less informative. Sample size is the number that gets reported; heterogeneity is the number that determines whether the estimate is real.
Null Results Are Doing the Most Work This Month
The emulsifier trial found nothing on every systemic inflammation marker for all five agents tested — a result that meaningfully constrains a hypothesis that has driven a decade of clinical advice about packaged food. The incretin review found that lifestyle comparator arms lost muscle at the same disproportionate rate as drug arms, which reframes the entire concern from a drug-safety question to a weight-loss-management question. Neither of those is the sentence that will be quoted from either paper, and both are more useful than the findings that will be.
Disclosure Gaps Are Not Always Where You Expect
April’s clearest funding problem is not in a supplement trial. It is a 400-patient multicentre trial whose own protocol paper names the product’s manufacturer as funder while the published paper declares no relevant relationships. Meanwhile the UK Biobank dementia analysis — a large, competent, academically framed cohort study — was supported in part by two soy-industry bodies and led by an author with equity in a fatty-acid testing company, none of it in the conflict statement. The lesson is not that either finding is wrong. It is that the conflict-of-interest statement is not a reliable index of who paid, and checking the funding line separately is now part of reading a paper.
What Remains Unresolved
Whether the disproportionate muscle loss on incretin therapy translates into functional decline is unanswerable from the current evidence base, because not one of 35 trials measured strength, gait speed, or mobility. Whether emulsifier effects that are absent in healthy adults over four weeks would appear in inflamed guts over years is exactly the question the field needs and exactly the one this trial could not ask. And whether plasma choline and betaine can identify deficiency in a free-living patient — as opposed to under a compliance-verified feeding protocol — remains the stated next step rather than an achieved result.
The connective tissue across all ten papers is the distance between what a study established and what its abstract asserts. That distance was largest in the papers with the most impressive headline numbers, and smallest in the least glamorous paper in the issue: a meta-analysis of protein and calorie support in liver disease, with numbers needed to treat between 8 and 23 and no heterogeneity to speak of.
“A wide interval is not a flaw in the study. It is the study telling you how much to trust it.”
Next month, a reasonable discipline: before quoting an effect size, find its interval and its heterogeneity, and decide whether you would still quote it.