Healthcare AI Evidence Library

Why a Citation Beats a Case Study

A vendor case study is written by the vendor. That does not make it false, and it does make it selective. Nobody publishes the site where it did not work.

Peer-reviewed work is not perfect either. Plenty of it is small, funded by the company, or run in a setting nothing like yours. But it has been through people whose job was to argue with it, and you can read the method instead of the highlight reel.

So the useful move in a procurement conversation is not to reject the case study. It is to put a real study next to it and ask why they differ.

Three Questions to Bring to Any Study Here

Reading one study properly beats skimming ten. These are the three things worth checking before a result changes your mind:

1. Who was in it? A model that performed beautifully in one country's screening program, on one manufacturer's scanners, is telling you about that program. Look for the population and the equipment before you look at the accuracy figure.
2. Compared with what? Better than nothing is easy. Better than your current process, staffed the way it is today, is the comparison that matters. A study that tested software against unassisted readers is answering a different question from one that tested it against double reading.
3. What was actually measured? Detection rate, reading time, and patient outcome are three different claims, and a paper that improves the first says nothing automatic about the third.

How These Lists Are Built

Each subject is a saved search against Europe PMC, which indexes PubMed, MEDLINE, PMC and the main preprint servers. The search is pinned to the title and abstract fields rather than run loose across the full text, and the results are then ordered by citation weight.

That pinning is worth mentioning because the first version of this library did not have it. A loose search sorted by citations returned the most famous AI papers that merely mentioned the words somewhere, so a general image-segmentation paper turned up at the top of the colonoscopy list. It was correctly matched and completely useless. Tightening the fields fixed it.

Nothing on a study row is written here. Title, authors, journal, year, identifiers, citation count and the abstract all come straight from the source record, which is the only way you can trust a citation you did not look up yourself.