What If We're Overcomplicating Healthcare IT?
I make the case that SQL could be the solution to simplifying the layers of complexity in health data architecture. This is a written version of my talk at EHRCON 25.
I’ve been working on a side project.
It’s called TrialMatched. The idea is simple, I want to help patients find clinical trials that actually fit them.
The reason I built it is personal.
My mother-in-law went through a lot of treatment for cancer. At some point, we started looking at clinical trials. My wife is a doctor. I’m a doctor. And we still couldn’t make sense of what we were reading. The inclusion criteria on clinicaltrials.gov reads like research jargon written for researchers, not patients or families trying to make a decision under pressure. You search for cancer trials in the US and you get thousands of results. Thousands. And most of them are incomprehensible.
So I just wanted to see: Can we make this easier?
The website is pretty bare. You leave your name and phone number. We call you, ask about your situation, figure out which trials you might qualify for, and help you get through the process. That’s it.
We launched Meta ads. The ad is genuinely bad, it’s just a picture and the words “Clinical trials near you.” About $5 per registration. And we’ve had close to 40 people sign up in the first week, which honestly surprised me.
Right now we’re tracking registrations in a Google Sheet. Someone from our Boston office is calling each person and filling in their details. It’s about as manual as it gets.
Here’s where it gets interesting (and hard)
Once someone registers, we call them and figure out where their data lives. Which hospitals, which EHRs. Then we send them a magic link, they connect their records, and we pull the data through FHIR.
If you’ve done our FHIR Fundamentals course, this part will be familiar. We’ve done this kind of EHR integration many times at Medblocks.
But there are always gaps. Some hospitals don’t support SMART on FHIR. Some want you to fill out a PDF form and fax it. Some require a HIPAA release. We’re looking at TEFCA’s Individual Access Services as a better path, you verify the patient’s identity through something like ID.me or Clear, and then you can pull records across the major QHINs like CommonWell and eHealth Exchange in one shot. We’re not there yet, but it’s getting closer.
And then there’s the data itself. Epic, for example, will send you the same patient record as a PDF, RTF, CCDA document and a FHIR resource, sometimes all at once. It is a lot to make sense of.
Once we have the data, the next problem is matching it against trials on clinicaltrials.gov. Look at the inclusion and exclusion criteria. See if the patient fits.
It is not as straightforward as it sounds.
The accuracy problem
I’ve been testing how well AI agents actually answer questions on real EHR data. Using Claude Opus 4.6 with multi-turn reasoning, tool calling and the works, on a dataset of 2,000+ patients, we got about 59% accuracy on clinical questions.
That’s barely better than a coin flip.
And the failures aren’t random. The model confidently sums the wrong things. It invents value sets on the fly and gets them wrong. It misses codes that don’t match the obvious text description. I wrote about this in detail here, the short version is that this is a much harder problem than it looks. Frontier models are still not reliable enough for this kind of matching at scale.
The data retrieval side is getting better fast. The reasoning side still has a long way to go.
Medblocks is doing well, this isn’t about money. I just wanted to see if we could do something useful here, and I was honestly surprised that 40 people showed up in week one asking for help.
And we want to help. We’ve been through this, and we know how hard it is.
If you’ve worked on clinical trial matching, patient data retrieval, or anything adjacent to this, I’d genuinely love to hear from you. What have you tried? What’s worked?
Reply to this email.
P.S. A lot of this is built with FHIR. If you want to go deeper into that, our FHIR Fundamentals course is completely free.
Related articles
View all
I Am Terrible at Directions
But that hasn't mattered in years.
CDS Hooks Matter More Than You Think
CDS Hooks are more than alerts: they’re real-time EHR webhooks that can surface a card and launch a SMART on FHIR app, letting you run workflows outside the EHR. The key is relevance + low latency, and rising adoption.
