Kyle Brill, CTO at Trio Health
AI is going to change how long and how well people live. I want to help prove it's safe and that it works.
If AI in healthcare goes right, kids grow up healthier. People live longer, and more of those years are good ones. Fewer people die from diseases we could have caught. When something new hits, the response is fast. And the care reaches people no matter where they live.
To get there, every new treatment and every AI result has to show it's safe and that it works. That takes data people can trust and check, and teams built to do the checking.
How I think about it
These hold for AI in healthcare and for the software behind it.
-
If you can't check it, you can't trust it.
An AI result pulled from a patient record should trace back to the note it came from. Software should meet the same standard. My teams run automated tests on every change and record a demo every week, so anyone can see whether a thing works. It's either in production or it's not.
-
At scale, the checking has to be measured.
Nobody can read six million data points by hand, and nobody can review every line of code an AI writes. So we measure. In healthcare that means confidence intervals and a bar for good enough that's set up front. In software it means tests and evals that run in the pipeline, and AI turns out to be really good at writing the tests. You're never going to get to 100%.
-
Go fast, and keep the quality.
We can't not go fast. If we cut quality to get there, we pay for it later. Tests and code review stay on even when a coding agent writes most of the code. A faster answer from patient data still has to hold up when someone makes a decision about care on it. Most of the job is finding the right amount of rigor for the problem in front of us.
-
When building gets cheap, choosing gets harder.
The cost of execution is collapsing, for code and for analysis. That doesn't mean we should build everything. My teams plan a month at a time and demo every week, because watching people use the thing tells us what to build next faster than a plan does. Data works the same way. Running another analysis costs almost nothing now, so the work is picking the questions whose answers would change care.
Where I'm coming from
I lead product, engineering, data and quality at Trio Health, where we turn de-identified EHR data into datasets and patient cohorts for life sciences teams. That puts me between the people building AI and the people who need evidence before they can use it. I write about both, and about how teams change when AI does more of the work. The posts are where I get into the details.
Recent writing
Posts start in November.
Most of the conversation happens on LinkedIn. If this is your world too, connect with me there.