Stem Cell Studies
Log inSign up

How we sort studies

This site is an index. It finds published studies, sorts each one by what kind of study it is, and links you to the source. It does not tell you whether a treatment will work for you.

Where the studies come from

Published studies come from PubMed, the U.S. National Library of Medicine's index of medical journals. Registered trials come from ClinicalTrials.gov. We search both each week. We store the identifier, title, journal, year, and our labels, and we link to the original record. We do not republish abstracts.

The evidence ladder

Every study gets exactly one of five labels. The order reflects how much a study of that type can tell you about whether a treatment works in people.

  1. Randomized controlled trial

    Patients were assigned at random to the treatment or to a comparison, such as a placebo or another treatment. Random assignment is the best protection against a result that only looks good because of who chose the treatment.

  2. Controlled study

    Treated patients were compared with a separate group, but not by random assignment. Useful, though the groups may differ in ways that affect the result.

  3. Case series

    A report on one or more treated patients with no comparison group. It can show that something is possible and how often problems occurred. It cannot show that the treatment caused an improvement.

  4. Animal or lab study

    Work in animals, cells, or tissue samples. No patients were treated. Results here often do not carry over to people.

  5. Review

    A summary of other studies. Some are systematic and pool results carefully; others are one author's overview. A review is only as strong as the studies it covers.

The label describes the type of study, not its quality. A small, poorly run randomized trial can be less informative than a large, careful controlled study. Trial protocols, editorials, and commentary are kept off the ladder.

Who does the sorting

An AI model does the first pass. For each study it reads the title and summary and answers a fixed set of questions: what type of study this is, what condition it covers, where the cells or blood product came from, whether patients were involved, whether there was a comparison group, and whether side effects were reported.

The model reports how sure it is of each answer. When it is not sure, or when its answers contradict each other (for example, a randomized trial with no comparison group), a person checks the study before it is counted. Until then the study is left out of the totals, and each condition page says how many are waiting.

Before using the model we tested it on 200 studies against a full second set of labels. It matched on study type 96.5% of the time. It was less reliable on some other questions, which is why those go to a person more often. The model can still be wrong when it is confident. If you see a study in the wrong place, tell us and we will check it.

Published studies in search

When you search a condition, the count of published studies and the evidence ladder come straight from PubMed, using the National Library of Medicine's own indexing of each paper: its publication type, and whether it involved people or animals. That covers every matching study, not a sample. Very recent papers are often not indexed yet, so they are shown as not sorted. For the newest randomized and controlled studies listed, an AI model reads each paper's summary and sorts what it reported, the same way it does for trials below.

How search works

When you ask a question, an AI model reads it and picks out the condition, any place, and filters such as “recruiting” or “randomized”. It does not write an answer. We then search ClinicalTrials.gov for trials of cell-based and blood-product treatments, and the model checks each trial the registry returns so that off-topic ones are left out. The page shows how your question was read, how many trials the registry returned, and how many we kept. Everything you see about a trial comes from its registry record.

The check looks at up to 60 trials per search. For broad questions, add a place, a status, or a specific treatment to see the rest. The model can misread a question or wrongly drop a trial, so treat the counts as a guide and follow the links.

Trial summaries and reported results

The short description under each trial is assembled from its registry record: the design, the number of participants, what was tested, what was measured, and how the trial ended. No AI model writes it.

For trials that have results, an AI model reads either the summary of the published paper linked to the trial, or the results posted on the registry, and sorts the trial into one of five groups: reported a benefit over the comparison group, reported no benefit over the comparison group, reported mixed results, participants improved but there was nothing to compare against, or reported on safety only. The model goes by the comparison and the statistics in the text, not by how hopeful the authors sound. When it is not confident, we say so and link to the source instead of showing a label.

Keep three things in mind. Most registered trials never post or publish results, so most trials show none. Trials with favorable results are more likely to be published, so any tally leans favorable. And a label describes what one trial reported; it does not tell you whether a treatment works or would work for you.

The written summary

Above the numbers, most searches show a short written summary. A general-purpose AI language model writes it, and it is given only the records on the page: the study counts, the listed papers and trials, and the label each one received. It is told to report what studies found, to use numbers, to cite the studies it names, and never to give advice or a verdict on whether a treatment works.

Before you see it, a second model reads the draft against the same records and blocks it if it gives advice, predicts what would happen to a person, states a general verdict, or includes a number or claim that is not in the records. When a draft is blocked, the writer gets one more try. If that fails too, the page shows the numbers with no summary. A summary can still be wrong, which is why every study it names links to its source.

What we do not do