AudioScholar·Find Trials ·Method note
A free tool for trainees: type a topic in any specialty and get the randomised trials, systematic reviews and guidelines it is built on — each with a direct link through your own university library, and one tap to save into Zotero, Mendeley or EndNote. Every number on every card comes from PubMed or the NIH. A model (Claude) chooses and orders the shortlist and writes the one-line reason on each card — it can drop or reorder the papers the search returned, and it cannot add one. The tool shows you the numbers it ranked on.
How a paper is surfaced
Each step either widens the net or narrows it. Nothing is hand-picked: PubMed and the NIH supply every paper and every number, and a model is allowed to do exactly two things at the end — choose and order a shortlist from what the search already returned, and say in one line why each paper is there.
One search per lane, taking up to three thousand papers rather than the first page or two. Depth is the whole point: the papers you are looking for are usually not the ones PubMed shows first.
Randomised and phase III trials in one lane; systematic reviews and meta-analyses in a second; guidelines in a third. These labels are the National Library of Medicine's own indexing, not our guess. A systematic review that earns a place in the shortlist is also listed under reviews, so a reader looking for the definitive meta-analysis finds it where they expect it.
Multicenter Study is left out of the trials lane on purpose — it lets large
observational registries into a list labelled "Trial". Removing it cut one topic's pool
from 2,626 papers to 1,041 without losing a single known landmark.
Ranking uses the Relative Citation Ratio from NIH iCite — a citation rate normalised for both research field and year of publication. That is what lets a 2001 nephrology trial and a 2019 oncology trial be compared without one burying the other, which raw citation counts cannot do.
Ranking on citations alone lets a merely famous paper drift to the top of a loosely worded search. So on a large pool a paper qualifies only if PubMed ranked it near the top for your topic, or its own title carries two of your topic's words — including the synonyms the search was widened with, so a trial titled "renal" is not penalised for a reader who typed "kidney" — or PubMed indexes your topic as one of its major subjects. Papers that fail are pushed below every paper that passes — and labelled — but they are never deleted. On a small pool, where there is nothing for fame to hijack, the test stands down.
An older guideline out-cites its replacement simply for having existed longer, so a plain citation sort would put superseded guidance above the document that replaced it. Guidelines from the last eight years therefore come first, ordered by how heavily they are cited, and everything older follows by date. The same guideline published in two journals is shown once, and where an older edition shares its replacement's title only the newer one is kept. That is what lifts a society's current guideline above the expert opinions and national digests that share its dates.
A paper from the last few weeks has no citations yet, so it cannot compete in a citation-ranked list — and the National Library of Medicine often has not yet tagged it as a trial at all. The newest papers from about seventy major journals, over the last four months, get their own section ordered by date, with no study-type gate.
Three narrow jobs go to a model (Claude). Before the search, it turns what you typed into the concept groups PubMed actually needs — "SGLT2 inhibitor" also as empagliflozin and dapagliflozin — and says so above the results, with one tap to search your exact words instead. After the search, it reads the forty most-cited candidates and picks up to twenty-five, ordering them by how much a trainee needs to have read each one and writing one line per paper on what it established or changed. It may drop a paper and it may reorder them; it can never add one. Every identifier it returns is checked against what the search found, and anything not in that set is discarded. The one-line reasons are the model's words and are marked as such; every number on the card is the NIH's.
How we check they are real
A trainee reading this list may not yet have the background to sense when something is wrong. So the checks have to hold on their own.
Excluded in the search itself. This matters more than it sounds: a citation-ranked list is enriched for retracted work, because a paper the field spent years rebutting accumulates citations while being rebutted.
These carry no publication-type marker at all — the only trace is inside the paper's own corrections record. Of sixty such papers we checked, sixty would have displayed with no warning. They now carry an amber badge, and a retracted paper that slips past the search filter gets a red one with its citation count removed.
Where a correction exists, the card says so and points at the record. RE-LY carries one that revised its event counts — worth knowing before you quote a number from the abstract.
A PubMed record also lists the IDs of every paper it cites. Reading those by accident would badge a card with a different paper's data. Every ID here is anchored to the article's own record, and every citation figure is matched by identifier, never by position in a list.
What it cannot do
This is the part worth reading twice, and it is printed on the tool itself, above the results — not buried in a footnote.
Search perioperative beta blockade and the top result is a 1999 trial from a group later found to have committed research misconduct — the finding that made perioperative beta blockade standard practice, and that was subsequently overturned.
Its PubMed record carries no retraction, no correction and no expression of concern. The trial that reversed the conclusion does not even appear in the results, because it uses a different word for the drug. Every number on that screen is true, and a reader without the background reads it exactly backwards. That is why the warning sits above the list rather than beneath it.
That is the honest test — search something you can already judge, and see whether the list matches what you would have taught. Then set your library once, and every result opens full text through your own institution.
Free · No account · No login · Nothing to install