I've been testing Iris.ai for the past three months, specifically for matching patients to clinical trials. My team handles a lot of observational study data, and manually sifting through ClinicalTrials.gov for potential matches is a massive time sink.
My initial impression? The tool is powerful, but the learning curve is steeper than I expected. It's not just a keyword search engine; you really have to learn how to "train" it with your own documents and define your research questions properly. I used it to take a cohort of 50 patient profiles and find relevant interventional trials. The process wasn't instant, but after some tuning, it surfaced a few highly relevant, recent trials we had completely missed in our manual process.
A few key observations from my workflow:
* **The "Workspace" setup is crucial.** You feed it your patient cohort criteria (anonymized, of course) as your "context" documents, and then you point it at the trial registry data. The AI then maps the relationships.
* **Precision vs. recall is a real trade-off.** Out of the box, it gave me a huge list of possible matches. I had to use the filtering tools extensively to narrow down to the truly relevant ones, focusing on specific eligibility criteria like biomarker status.
* **It's fantastic for discovery,** but **not** for final verification. You absolutely must have a human (preferably a clinician) review the logic behind each match. The tool shows you *why* it made a connection, which is great for transparency.
Has anyone else used it for a similar use case? I'm particularly curious about:
- How you structured your "research question" for best results.
- Whether you integrated it with any internal patient databases (we're considering this).
- If the time saved in the long run justified the setup time.
The potential here feels huge, especially as trial criteria get more complex. It feels less like a search tool and more like a research assistant that can highlight connections you might not see.
"Training" the tool sounds like the real bottleneck. So you spent three months tuning it to find a few relevant trials you missed. What's the ongoing maintenance cost for that model? Does it drift when the trial registry updates?
I've seen these bespoke AI matching systems become a full-time job to maintain. A clean SQL view on a refreshed ClinicalTrials.gov feed and some well-defined exclusion criteria often gets you 80% of the way there for 5% of the ongoing headache.
But I'm curious, what was the actual time spent? Was the "massive time sink" of manual searching truly greater than the setup and tuning time?
SQL is enough
That initial tuning period is real. We had a similar experience when setting up our own process. The time investment upfront is substantial, but for us, it became a force multiplier.
It's less about maintaining a single model and more about refining a repeatable process you can run for each new cohort. Once you have your workspace template and filtering rules dialed in for a specific therapeutic area, you can apply it again much faster. The real question is how often your core data model changes.
Did you document your setup steps? I ended up creating a quick checklist for my team - "define research question, upload context docs, set inclusion/exclusion as filters" - which cut the setup time for subsequent projects by about half.
> "surfaced a few highly relevant, recent trials we had completely missed in our manual process."
That's the claim that always gets me. How do you know you "missed" them? Manual review has its own recall floor, but it's a known floor. Iris.ai gives you a black-box suggestion set that you then have to manually verify anyway. So you're swapping one manual step for another manual step plus a training phase.
> "Precision vs. recall is a real trade-off."
You had to filter extensively to get to the "truly relevant" ones. So the tool's main value was generating a candidate list that you then manually pruned. That's not matching - that's a pre-filter. A good SQL query with some trial ID exclusions and a few boolean conditions on eligibility criteria can do the same without the workspace templating overhead.
What was the actual false positive rate? If you started with 50 profiles and got a "huge list" before filtering, how many of those were garbage? And how many hours did you spend on the filter tuning for those 50 patients?
I'm not saying the tool is useless. But I'm tired of hearing "we found stuff we missed" without a baseline comparison. Did you run the same 50 profiles through a manual process with a stopwatch to see if the total time was actually lower?
Question everything