Evidence
Tested on real reviews
An AI screening tool is only useful if you can trust it on your review, not just on average. So we tested Cidexar against the decisions of the experts behind published systematic reviews, we looked hardest at where it did worst, and we built in a way to check it on your own review.
Three things that matter
Does it keep the right papers?
Of the papers that ended up in the final published reviews, how many did Cidexar keep?
Does it clear the noise?
Of the papers that were clearly irrelevant, how many did it set aside?
How much does it vary?
An average can hide a bad result. How did it do on the reviews it found hardest?
What we found
About98%
of the papers that made it into the final reviews were kept, at the standard setting.
About7 in 10
clearly irrelevant papers were set aside at the same setting.
On the cautious setting, it kept essentially all of the final included papers, while still clearing about half the noise. That's the setting to choose when a topic is hard or unusual.
These are results across many reviews. Some topics are harder than others, which is why you choose the setting, why nothing is ever deleted and why you can check its work on yours.
How we kept it honest
Real expert decisions.
We compared Cidexar with the actual choices made by the teams behind published systematic reviews, across a range of health topics.
Reviews it had never seen.
We tested on hundreds of papers from dozens of reviews that weren't used to build or tune it.
Weak spots, not just averages.
We checked results at every setting and looked closely at the reviews where it did worst. We'd rather tell you about them than hide them in an average.
Consistent results.
The same paper and criteria get a consistent score. Small differences between scores don't mean much, so we group papers into clear tiers.
Check it on your review
Our testing tells you how Cidexar did on other people's reviews. Yours may be easier or harder. So you don't have to take our numbers on trust.
Sample what was set aside.
Pull a random sample of the "likely exclude" pile and screen it yourself.
Adjust if you need to.
If your sample turns up papers you'd keep, choose a more cautious setting or read more of the pile.
Report what you did.
Your setting, your check and the numbers for your flow diagram all go in your methods. What to report
What Cidexar doesn't do
It reads titles and abstracts, not full text. Some papers that look eligible from the abstract will still fail at full text. That's normal at this stage.
It doesn't make final decisions. Your team does.
It isn't perfect. That's why we tell you the numbers, let you choose the setting and keep every paper.
Be first to screen with Cidexar
Cidexar is coming soon. Leave your email and we'll let you know the moment it's ready.