A multistrain probiotic for tau biomarkers and cognition in early cognitive impairment: a double-blind randomized controlled trial

Tap Document above to read the actual manuscript these findings cite.

12 findings7high5medium

What a reviewer will push on first

The most severe issues a reviewer will raise, named. Each jumps to the full critique below.

The catch a generic AI would miss

Internal consistencyTable 1 (GDS and NfL rows)

Table 1’s GDS total (8.73) falls below both group means — impossible for a weighted average — and the NfL total is inconsistent the same way

Table 1 reports GDS as 8.55 in placebo and 9.50 in probiotic, but a total of 8.73. A total mean is the sample-size-weighted average of the two group means, so it must lie between 8.55 and 9.50; a total of 8.73 sits below both arms, which no positive weighting can produce. This is not a rounding artifact — it points to a table-generation problem (a mis-merged column, or a row pasted from a different cut of the data). It is the kind of thing a statistical referee catches in the first five minutes.

Editorial verdict

Do not submit yet — rebuild the evidentiary spine first

Likely outcome if submitted today to a flagship AD journal: desk reject, or at best rapid rejection after statistical/CONSORT screening. There is an interesting signal — a plasma pTau181 decrease in the probiotic arm — but the trial is currently reported as a confirmatory RCT without the spine of one: no detected trial registration, no prespecified primary/secondary endpoint hierarchy, no sample-size justification, complete-case rather than intention-to-treat analysis, and no multiplicity control across roughly a dozen outcomes.

Read the full editorial recommendation

The single most important fix is not prose polishing — it is rebuilding the evidentiary spine. Regenerate every table from the locked dataset (Table 1’s GDS total of 8.73 falls below both group means, which no weighting can produce; Table 2’s TMT-B change scores do not match its baseline and follow-up means), add a CONSORT flow diagram and an ITT analysis of all 100 randomized participants, declare one prespecified primary endpoint with a multiplicity strategy, and report effect sizes with 95% confidence intervals instead of bare within-group p-values.

Then reframe the claims to match the data: limit "tau" language to "plasma pTau181" (pTau217 did not change and no PET was done), describe the cohort as "early cognitive impairment enriched for AD pathology" rather than "Alzheimer’s disease" (42.5% were amyloid-negative), and drop the "oral-immune–brain axis" mechanism unless microbiome or immune markers are measured. Recommended target after overhaul: Alzheimer’s & Dementia: Translational Research & Clinical Interventions is a more realistic first venue than the flagship journal.

Quality scorecard

Overall · 2.6/5

Originality

3/5

A probiotic RCT targeting AD-related plasma biomarkers and cognition in early cognitive impairment is a timely question — the gut/oral-brain axis is an active area — but the design is incremental rather than a novel mechanistic advance.

Importance of research question

4/5

Disease-modifying, non-pharmacological strategies for early cognitive impairment are a major unmet need, and a clean plasma-pTau181 signal from a low-risk supplement would be genuinely worth reporting.

Claims are well-supported

1/5

The central claims are not supported as stated: the title’s "ameliorate tau pathology" rests on a single plasma pTau181 change while the AD-specific pTau217 did not move, no tau/Aβ PET was performed, 42.5% of the analyzed cohort was amyloid-negative, and no microbiome or immune readout supports the "oral-immune–brain axis" mechanism.

Experimental soundness

2/5

The double-blind placebo-controlled design is appropriate, but there is no trial registration or prespecified primary endpoint, no multiplicity control across roughly a dozen outcomes, complete-case (not ITT) analysis with imbalanced 20% vs 6% attrition, and within-group p-values are presented as between-group efficacy.

Clarity of writing

2/5

Beyond prose, the data presentation is internally inconsistent: Table 1’s GDS total falls below both group means (impossible for a weighted average) and its NfL total is inconsistent the same way, Table 2 change scores do not match the reported baseline/follow-up means, and the ROCF-copy p-value differs across the abstract (0.027), the text (0.037), and Table 2 (0.369).

Value to community

3/5

A clean plasma-pTau181 result in a probiotic RCT would be useful to the field; the value is currently undercut by overclaiming and table errors rather than by the underlying data, which may well survive an honest reanalysis.

Prior work contextualization

3/5

The gut/oral-brain-axis and plasma-biomarker literature is engaged, but the Discussion does not situate the pTau181-without-pTau217 result or address non-AD explanations (inflammation, peripheral physiology, assay variability).

Showing the 12 highest-impact findings. The full Manusights review flagged 26 issues on this manuscript.

Detailed findings

High severity.
#1Internal consistencyTable 1 (GDS and NfL rows)
Table 1’s GDS total (8.73) falls below both group means — impossible for a weighted average — and the NfL total is inconsistent the same way
14.95 (4.74) 15.56 (5.83) 18.18 (29.1)
Table 1 reports GDS as 8.55 in placebo and 9.50 in probiotic, but a total of 8.73. A total mean is the sample-size-weighted average of the two group means, so it must lie between 8.55 and 9.50; a total of 8.73 sits below both arms, which no positive weighting can produce. This is not a rounding artifact — it points to a table-generation problem (a mis-merged column, or a row pasted from a different cut of the data). It is the kind of thing a statistical referee catches in the first five minutes.
The NfL row is inconsistent in the same direction: placebo 14.95 (n=47), probiotic 15.56 (n=40), but a total of 18.18 — well above the n-weighted average of (47 × 14.95 + 40 × 15.56) / 87 = 15.23. NfL is strongly right-skewed and the total SD of 29.1 dwarfs each group’s (4.74 and 5.83), so an untrimmed total dominated by a few extreme values could in principle sit above both group means — but then the table has to report the denominator and trimming that produce it. As written, the group rows and the total cannot all be correct. Two totals that cannot be reconciled point to the table, not to a typo.
Suggested fix
Regenerate every descriptive statistic in Table 1 directly from the locked analysis dataset, and add a numeric self-check (each total mean must equal the n-weighted average of the group means) before resubmission.
High severity.
#2Internal consistencyTable 2 (TMT-B rows)
Table 2 change scores do not equal follow-up minus baseline
469.31 (426.32) 419.30 (392.06) 8.85
For placebo TMT-B, Table 2 lists baseline 469.31, 24-week 419.30, and a change of 8.85 — but 419.30 − 469.31 = −50.01. For probiotic TMT-B it lists 455.62 → 362.18 with a change of −64.00, where the difference is −93.44. Neither gap is a rounding artifact. If different participants contributed to the baseline, follow-up, and change columns, the table must show outcome-specific n values — but the Methods state complete-case data with n=47 and n=40 throughout, which forecloses that explanation.
Suggested fix
Audit every change score in Table 2 against its own baseline and follow-up means; state explicitly whether change = follow-up − baseline (and for timed tests like TMT-B, that a negative change is an improvement); add outcome-specific n if missingness varies.
High severity.
#3Statistical auditAbstract; Results; Table 2
The same effect (ROCF copy) is significant or not depending on the abstract, the text, or the table — and a within-group p-value is used as efficacy
ROCF copy, p = 0.027) and maintained stable
ROCF copy is reported three different ways. The Results text says the between-group ANCOVA found "ROCF copy (z-score, p = 0.037)". Table 2 reports the between-group ROCF-copy p as 0.369. The Abstract headlines "ROCF copy, p = 0.027" — but that value is the within-probiotic paired test (change from baseline), not a between-group treatment effect. A reader cannot tell which inference is intended, and the one the Abstract leads with is the wrong kind: for an RCT, the question is not whether the probiotic arm changed from baseline, but whether it changed more than placebo under a prespecified model.
The same pattern recurs with the biomarkers: "GFAP and NfL remained stable in the probiotic group, whereas the placebo group showed significant increases (p = 0.014, p = 0.041)." Those are two separate within-group tests; a significant change in one arm and a non-significant change in the other is not evidence of a between-group difference. The between-group interaction must be tested directly.
Suggested fix
Choose one ROCF-copy estimand (raw / z-score / change), report a single between-group effect with its 95% CI, and reconcile the abstract, text, and table. Replace every "significant in one arm, not the other" claim with a direct between-group (group × time) test.
High severity.
#4Editorial framingTitle; Abstract; Discussion
Title and conclusion claim "tau pathology" and an "oral-immune–brain axis" the data do not show
modulating the oral-immune–brain axis with multistrain probiotics
The title says the probiotic "ameliorate[s] tau pathology" and the conclusion invokes "modulating the oral-immune–brain axis" to "slow AD-related pathological progression." The data underneath are narrower than that. Plasma pTau181 fell, but plasma pTau217 — the more AD-specific marker — did not change in either group, and the authors’ own limitations note that no tau or Aβ PET was performed, so the biomarker change cannot be tied to in-vivo brain pathology. The cohort is not uniformly AD: by the study’s own Table 1, only 57.5% were amyloid-positive, so "Alzheimer’s disease" claims do not hold across the 42.5% who were amyloid-negative. And nothing in the study measured oral or gut microbiota, microbial metabolites, or inflammatory markers — so "oral-immune–brain axis" is a mechanism asserted, not tested.
Engine-suggested honest retitle
Your manuscript
Multistrain probiotics ameliorate tau pathology and preserve visuospatial cognition in early cognitive impairment
Reviewer-ready
Multistrain probiotic supplementation reduces plasma pTau181 and is associated with preserved visuospatial performance in early cognitive impairment: a randomized placebo-controlled trial
Suggested fix
Scope the claims to the measurements. Limit "tau" language to "plasma pTau181," describe the population as "early cognitive impairment enriched for AD pathology," and drop the oral-immune–brain mechanism unless it is measured.
High severity.
#5Reporting guidelineMethods; whole manuscript
Reported as a confirmatory RCT, but the CONSORT spine is missing: no registration, no primary endpoint, no multiplicity plan
p-value < 0.05 was considered statistically
The engine detected a randomized controlled trial (high confidence) and applied CONSORT 2010; two critical items are missing. No trial-registration identifier (ClinicalTrials.gov NCT or an equivalent registry number) is detectable anywhere in the manuscript — for a prospective RCT, most flagship journals treat the absence of prospective registration as close to disqualifying. No single primary endpoint is prespecified: the trial reports four biomarkers (pTau181, pTau217, GFAP, NfL) and a battery of cognitive tests (CDR, MMSE, CERAD, Stroop, ROCF), then applies paired t-tests, independent t-tests, ANCOVA, and a mixed model with only "a p-value < 0.05 was considered statistically significant" — and no multiplicity adjustment across the dozen-plus comparisons.
Suggested fix
Supply the registration number (or state honestly that the trial was not prospectively registered and reposition the paper as exploratory), declare one prespecified primary endpoint with secondary and exploratory tiers, and add a multiplicity strategy (e.g. hierarchical testing or false-discovery-rate control).
Medium severity.
#6ReproducibilityMethods (analysis population)
Imbalanced 20%-vs-6% attrition handled by complete-case analysis, not intention-to-treat
complete-case data, without imputation
Of 100 randomized participants, 13 dropped out — but unevenly: 10 of 50 in the probiotic arm (20%) versus 3 of 50 in placebo (6%). The primary analysis then uses "complete-case data, without imputation" on the 87 completers, which is a per-protocol-style analysis, not intention-to-treat. Differential dropout of that magnitude can bias the efficacy estimate in either direction (for instance if less-tolerant or less-responsive participants left the probiotic arm), and the manuscript references ITT, PP, and safety sets without actually presenting the ITT analysis.
Suggested fix
Make an ITT-compatible analysis the primary one (mixed model for repeated measures, multiple imputation, or inverse-probability weighting), present complete-case as a sensitivity analysis, and report the CONSORT flow with randomized / treated / completed / analyzed denominators by arm.
High severity.
#7Statistical auditTables 1–2; Methods
Not a single effect size or 95% confidence interval is reported — only p-values
Values are presented as the mean (SD) or number (%)
Every result is reported as a mean (SD) plus a p-value; no between-group mean difference, standardized effect size (Cohen’s d, partial η²), or 95% confidence interval appears anywhere. For a clinical trial, the p-value answers "is there an effect?" but not "how big, and is it clinically meaningful?" — and with this many comparisons, a p just under 0.05 with no effect size is exactly what a trialist discounts. A 0.3-point change on a cognitive scale can be "significant" and clinically irrelevant at the same time; without effect sizes and CIs a reader cannot tell which of these results matters.
Suggested fix
Report, for every primary and secondary outcome, the adjusted between-group difference with a 95% CI and a standardized effect size, and interpret each against a stated minimal clinically important difference.
High severity.
#8Internal consistencyResults; Table 2 (BNT, CERAD CR)
Boston Naming Test is significant in Table 2 (p = 0.003) but the text says nothing changed
either group for the remaining cognitive assessments, including AF, BNT
The Results state that "no significant changes were observed in either group for the remaining cognitive assessments, including AF, BNT, TMT A & B, Stroop…" — but Table 2 reports the CERAD Boston Naming Test with a between-group ANCOVA p = 0.003, a clearly significant result. Either the text is wrong and a significant BNT effect is being under-reported, or the table is wrong; both cannot stand. The same section has a smaller version of the problem: the text gives CERAD CR p = 0.025 while Table 2 lists 0.022. Discrepancies like these read as manual p-value transcription rather than tables generated from the analysis, which undermines confidence in every number on the page.
Suggested fix
Reconcile BNT between text and Table 2 (report it consistently and, if retained, place it in the multiplicity framework), and harmonize every p-value across the abstract, text, tables, and figures from the locked output.
Medium severity.
#9Statistical auditResults (biomarkers)
The GFAP/NfL "treatment effect" is two separate within-group tests, not a between-group comparison
placebo group showed significant longitudinal increases
The neuroinflammation claim is built like this: GFAP and NfL "remained stable in the probiotic group," while the "placebo group showed significant longitudinal increases (p = 0.014 and p = 0.041)." Those are two independent within-group change tests. A non-significant change in one arm and a significant change in the other is not evidence that the arms differ — this is the classic "difference in nominal significance" error. The only valid test of a treatment effect is the direct between-group (group × time) comparison, which is not reported for GFAP or NfL.
Suggested fix
Report the between-group (group × time interaction) estimate with its CI for GFAP and NfL; drop the "stable in one arm, rose in the other" framing as evidence of efficacy.
Medium severity.
#12Editorial framingTitle; Discussion; Table 1
"Alzheimer’s disease" is claimed across a cohort that was 42.5% amyloid-negative
Amyloid burden positivity was determined using 18F-florbetaben
The framing throughout is "Alzheimer’s disease," but amyloid positivity was assessed by 18F-florbetaben PET and Table 1 reports only 57.5% amyloid-positive — so 42.5% of the analyzed cohort was amyloid-negative and, by current biomarker definitions, does not have AD. The honest population label is "early cognitive impairment enriched for AD pathology." If the authors want to make AD-specific claims, the effect should be shown within the amyloid-positive subgroup; otherwise the disease attribution overshoots the sample.
Suggested fix
Relabel the population as "early cognitive impairment enriched for AD pathology," and report the primary outcomes within the amyloid-positive subgroup before making any "Alzheimer’s disease" claim.
Medium severity.
#13ReproducibilityMethods; Results
No CONSORT flow diagram and no data- or code-availability statement
A reader cannot reconstruct the analyzed sample: there is no CONSORT flow diagram tracing assessed → randomized → treated → completed → analyzed by arm, and no data- or code-availability statement for the trial. Both are standard requirements for a clinical trial report and both are needed before the imbalanced attrition (finding 6) and the table discrepancies (findings 1, 2, 8) can be independently checked.
Suggested fix
Add a CONSORT 2010 flow diagram and a data/code-availability statement (de-identified analysis dataset and analysis scripts, or a clear access procedure), so the reported denominators and table values can be verified.
Medium severity.
#14MethodologyMethods; Results
Safety and adherence are asserted but never reported for an ingested-supplement trial
Compliance was verified at each visit
The Methods say safety assessments were performed and that "compliance was verified at each visit," but the Results report no adherence rates, no treatment-emergent adverse events, no lab abnormalities, and no serious adverse events — only the dropout counts. For a non-pharmacological intervention pitched as a low-risk strategy, safety and adherence are part of the value proposition and have to be shown, not assumed; and without adherence data, a null cognitive result cannot be separated from simple non-compliance.
Suggested fix
Add a safety table (TEAEs by preferred term, severity, relatedness, SAEs, discontinuations, lab abnormalities) and report adherence as mean ± SD and the proportion meeting a prespecified threshold, by arm.

Why this isn’t ChatGPT

Citations and DOIs verified against Crossref, with retraction-database checks
Statistics re-computed deterministically (statcheck / GRIM-style), not guessed
Every finding tied to a specific equation, table, or line in the manuscript
Trained on real peer reviews from 35+ current Nature, Cell, and Science reviewers

Full Review · $39

Find issues like these on your own paper.

A desk rejection can cost months. Same engine as the review above, trained on real peer reviews from 35+ current Nature, Cell, and Science reviewers.

Start with a free preview to see Manusights on your paper. If you want the full submission package (simulated peer review, target-journal risk, citation checks, reviewer strategy, and ready-to-use submission materials), upgrade to the Submission-Ready Dossier ($99).

Run a free preview