Science · Method
Everything published here can be checked: by a reader with a school diploma on the plain-language part, by a doctor on the technical part, and by anyone on the sources.
We start from a written research protocol that names the question, the areas, the databases and the journals, and asks for negative and contradictory results as much as for confirmations. A study enters only if it exists in PubMed or an equivalent bibliographic database with a DOI, if its samples or measurements were taken before the clinical event when the claim is about prediction, and if its outcome is defined. Theoretical frameworks without data are cited as frameworks, not as evidence.
Title, authors, journal, year, volume, pages, DOI, PubMed identifier, checked against PubMed and Crossref, not against memory.
The PubMed page, the DOI and the open-access copy, following redirects. We record the address requested, the final address, the answer and the time.
A page that answers is not enough: the page must be that study, and the numbers we quote must be there. A working link to the wrong paper is a broken link.
Each figure on a study page is checked on the document we actually read, and we write down where: abstract, section, table, figure or supplement. If we only read the abstract, we do not report details that are not in it.
We look for errata, retractions, expressions of concern and published critical comments, in PubMed and Crossref, and we say what we found, including when the search could not be done.
Every check carries its date and time, and says what could be consulted. A source we could not open is declared as such, never presented as verified.
An article behind a paywall can have a verified title and abstract without an accessible full text. We never collapse these into one tick.
The reference exists and is what we say it is.
Each address answers, with its final destination recorded.
What the page says matches what we claim, number by number.
We read the whole article, or we say which version (published or author manuscript) and that supplements were not read.
Four things are judged separately and never merged into one score: how solid the evidence is, what kind of result it is (association, prognosis, early diagnosis, intervention), how mature the application is (from A, already usable in the right setting, to D, not demonstrated), and how relevant it is to KSM. A study can be excellent and irrelevant to the Twin, or relevant and weak.
The tools, named in full: PROBAST (Prediction model Risk Of Bias ASsessment Tool) for prediction models, with its 2024 extension for models built with machine learning; TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis), with its 2024 extension, for how completely a model is reported; GRADE (Grading of Recommendations Assessment, Development and Evaluation) for the certainty of a body of evidence; ROB 2 (Risk Of Bias 2) for randomised trials; ROBINS-I (Risk Of Bias In Non-randomized Studies of Interventions) for non-randomised studies of interventions; QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2) for diagnostic accuracy studies; AMSTAR 2 (A MeaSurement Tool to Assess systematic Reviews 2) for systematic reviews. Reporting quality and risk of bias are different things and are kept apart. When the available text does not allow a complete judgement, the judgement is marked preliminary.
An abstract is the authors' text. We reproduce it in full only when the licence allows it and we can document that licence (for example Creative Commons Attribution), always with attribution. Otherwise the page carries a "Summary of the abstract by KSM", labelled as ours, and the direct link to the original. A translation is labelled as a translation. We never call a rewrite "original abstract", and we never reconstruct a missing abstract.
Each study page says who funded the work and whether authors own the test or work for the company that sells it, when the paper states it. And ours: Klinik Sankt Moritz develops the Digital Twin and the services that go with it. These pages are written by KSM. That is a reason to check our claims against the sources, which is why every claim links to one.
What we changed, when, and why. A correction to a study record propagates to every page and language that uses it.
The first version of the record said the replication cohort was not named in the abstract. On reading the full text, it is the Malmö Diet and Cancer study (163 cases, 163 controls); the record now states it with the position in the paper.
The author manuscript in PubMed Central reports a median follow-up of 8.2 years; the published abstract reports 9.7. The page uses the published figure and declares the discrepancy; it will be settled by reading the typeset article.
Each page shows the date of its last scientific verification, separate from the date of its last edit. The research protocol proceeds in blocks: biomarker signatures (done, 20 studies identified, 3 fully verified so far), biological clocks and ageing, oncology, behaviours and wearables, data over time, Digital Twin trials, from prediction to prevention, validation and consolidation. Pages for the blocks not yet done do not exist on the site.
artificial intelligence: programs that read, organise and generate text and data.
Area Under the Curve and concordance index: how well a model ranks people by risk; 0.5 is chance, 1 is perfect. Not the share of correct diagnoses.
Body Mass Index: weight in kilograms divided by height in metres squared.
Creative Commons Attribution licence: the text may be reproduced with attribution.
Digital Object Identifier: the permanent identifier of a publication.
Homeostasis Model Assessment: estimates of insulin sensitivity (HOMA-%S) and beta-cell function (HOMA-%B) from fasting glucose and insulin.
95% confidence interval: the range within which the true value is estimated to lie.
Liquid Chromatography–tandem Mass Spectrometry: a laboratory method that separates and identifies molecules.
Oral Glucose Tolerance Test: blood sugar measured two hours after drinking 75 grams of glucose.
Odds ratio, hazard ratio, risk ratio: relative measures of risk; 1 means no difference. They do not say how many people in a hundred are affected: that is absolute risk.
PubMed Identifier and PubMed Central: the article's number in the bibliographic database of the US National Library of Medicine, and the free full-text archive.
The quality tools, spelled out in the section above.
UK Biobank Pharma Proteomics Project: the proteomic measurement of a subset of the UK Biobank, funded by a consortium of pharmaceutical companies.
