Research FAQ

Are these randomised controlled trials?

No. Our published work to date consists of prospective, interventional and real-world cohort studies, reported as such in each paper. They demonstrate that AI-guided personalized microbiome modulation is associated with clinically meaningful improvement in the populations studied. They do not establish superiority over a comparator, because there was no comparator arm.

We state this plainly because it determines how the results should be used: as evidence of feasibility, effect size range and biological mechanism — not as a substitute for a randomised trial.

Why cohort studies rather than RCTs?

The intervention is personalized by construction: each participant receives a different combination of dietary change, synbiotics, antimicrobials and micronutrient correction, derived from their own multi-omic profile. A conventional fixed-protocol RCT tests the wrong object. The comparator that matters is the personalization process itself versus standard care, which requires a different design.

Our next-phase programmes are being designed with control arms and pre-registered protocols. We are open to co-designing them with pharma partners.

Was anyone blinded?

Dietary and supplement interventions of this kind cannot be blinded to the participant. Where outcome adjudication was performed independently of the treating clinician, this is stated in the individual study's methods block on the studies page.

Are the studies registered, and were they ethically approved?

Ethics approval and informed consent statements are reported in each publication. Registry identifiers, where a study was registered, are listed in the per-study methods block on the studies page. Where a study was not prospectively registered, we say so rather than omitting the line.

What does "much improved" actually mean?

It is a patient-reported global impression of change, collected at the end of the three-month programme. It is a subjective endpoint and we report it as one. The instrument used and the response scale are given in the per-study methods block.

It should be read alongside the objective biomarker endpoints, not instead of them.

What counts as biomarker normalisation?

Normalisation is defined against the reporting laboratory's reference range for each analyte — principally hs-CRP and faecal calprotectin — measured at baseline and at end of programme. The thresholds, assay methods and reference ranges used are given per study.

A change from a very high value to a lower but still abnormal value is reported as a reduction, not a normalisation.

Which endpoints were pre-specified and which were exploratory?

Each study's methods block separates the pre-specified primary endpoint, the pre-specified secondary endpoints, and analyses that were exploratory or post-hoc. Microbiome compositional findings — including taxon-level shifts such as Bifidobacterium longum — are exploratory unless explicitly stated otherwise.

Where are the confidence intervals and p-values?

Reported in the publications and reproduced in the per-study results table on the studies page, together with the statistical test used, how missing data were handled, and whether multiplicity correction was applied.

How much of the effect is placebo and regression to the mean?

In an open-label single-arm design, some of it will be. Participants enrolled at a symptomatic point, which favours regression to the mean, and expectation effects on self-reported outcomes are real.

This is precisely why the objective inflammatory markers matter: hs-CRP and faecal calprotectin are far less responsive to expectation than a symptom diary. The convergence of subjective and objective movement is the strongest argument available from this design — and it remains weaker than a randomised comparison.

Isn't there selection bias in who enrolls and who completes?

Yes, in both directions. People who seek out a personalized programme are more motivated than a general clinic population, and completers differ from non-completers. Enrolled, analysed and withdrawn counts are reported per study so the attrition is visible rather than hidden in a denominator.

Are results reported per-protocol or intention-to-treat?

Stated per study. Where an analysis is per-protocol, we say so, because it inflates apparent effect relative to intention-to-treat.

You funded and ran your own studies. Why should we trust them?

You should not trust them on our word. Every paper is peer-reviewed and openly accessible, funding and conflicts of interest are declared in each publication, and the underlying methods are summarised on this site so the design can be judged before the results are.

For partners under agreement we can provide de-identified analysis-level data and the analysis plan for independent re-analysis.

What does the model actually decide?

It integrates baseline stool metagenomics, blood biomarkers, micronutrient panels, genetics, medical history and longitudinal patient-reported signals into an estimate of biological state, and generates a candidate intervention plan from that estimate.

Candidate plans in clinical studies were reviewed by a clinician before being issued. The model proposes; a human accepts, edits or rejects.

How is predictive performance measured?

By comparing the model's predicted response trajectory against the measured trajectory at the next observation point, on data it did not see. The relevant metrics are calibration of the predicted response, error on the predicted biomarker change, and the rate at which clinicians override the proposed plan.

Aggregate performance figures are shared with partners under agreement and are being prepared for publication rather than being asserted here.

How do you avoid overfitting and leakage?

Evaluation is on held-out patients, not held-out timepoints from the same patient, so a patient never appears in both training and evaluation. Model updates are versioned, and a plan issued to a patient is tied to the model version that produced it, so any result can be traced to the exact system state that generated it.

What happens when the model is wrong?

Every issued plan generates a measured response, and a wrong prediction is a labelled training example — that is the point of a closed loop. Clinically, safety is held by clinician review and by hard constraints that no generated plan may cross, independent of what the model proposes.

Is the system regulated?

NostraBiome operates as an EU MDR Software Medical Device. Clinical evidence, risk management and post-market surveillance obligations follow from that classification.