Skip to content
EMMANUEL P. BAZILE, PhD
All posts

AI in Healthcare ·

Integration Has Always Been Translation

In 2005, I spent my days reading messages that almost no one else wanted to look at. They arrived as long strings of pipes and carets, segment after segment, each one describing a patient, an order, a result, a transfer between departments. To most people they were noise. To me they were sentences. My job was to make sure the sentence a laboratory system had written could be understood, exactly and completely, by another system that was never designed to listen to it.

That year I was analyzing HL7 messages for a national bio-surveillance program, watching for the earliest signals of an outbreak in data that was never meant to be read that way. It taught me something I have spent the rest of my career proving over and over: in healthcare, the systems rarely disagree about the facts. They disagree about how the facts are named, ordered, and trusted.

Two systems, one patient, two languages

An integration engine sits between two applications that each believe they already know how the world works. One calls it a "patient identifier." The other calls it an "MRN." One sends dates as a timestamp, the other as a string. One assumes every order has a result; the other assumes nothing until a human confirms it. Neither is wrong. They were simply built by different people, in different decades, to solve different problems.

My work has always been to stand in that gap and translate, faithfully, without losing meaning and without inventing it. Over more than two decades I built upward of ten thousand interfaces across the major electronic medical record and integration platforms. The code was rarely the hard part. The hard part was understanding what each side actually meant, and then earning enough trust from the people who depended on both systems that they would let a new connection carry real patient data.

That is the quiet truth of integration. A new interface is a promise. It says: whatever you put in on your side will arrive on the other side meaning the same thing. Break that promise once, in healthcare, and you do not get a second chance to earn it back.

Why people resist good technology

When I returned to study this question formally, my doctoral research asked why healthcare professionals embrace some technologies and resist others. I expected the answer to live in the software. It did not. People resist systems they do not understand, systems that change their work without explaining why, and systems that ask them to trust an output they cannot interrogate. They adopt the ones that respect what they already know.

I have watched brilliant implementations stall, not because the architecture was wrong, but because no one translated the change into the language of the people living through it. And I have watched modest systems succeed because someone took the time to make them legible.

Which brings me to AI

A large language model is, at its core, a translation machine. It takes meaning expressed in one form and renders it in another. That is the same problem I have worked on for twenty-seven years, now operating at a scale and a speed that would have been unthinkable when I was reading HL7 by hand.

The temptation, and I see it everywhere in healthcare right now, is to treat AI as a replacement for understanding. Point it at the mess, let it sort things out, trust the output. But an AI that translates confidently and incorrectly is far more dangerous than an interface that simply fails. A failed interface raises an alarm. A fluent, wrong answer does not.

So I approach these tools the way I approached every integration engine before them. I ask what it actually means when it tells me something. I check its work against reality. I never let it carry patient-critical meaning until it has earned the same trust I would demand of any other system in the pipeline. The technology is new. The discipline is not.

What I am teaching now

In the classroom, I bring the same systems I built for hospitals into my lectures, so students learn from real messages, real schemas, and real implementation decisions rather than abstractions. Increasingly, that includes teaching them to work alongside AI: to use it as a capable, tireless translator, and to never once stop reading the output as critically as I read those pipe-delimited sentences in 2005.

Integration has always been translation. AI has not changed that. It has only raised the stakes of doing it well. The people who will shape the next decade of healthcare technology are the ones who can hold both things at once: real fluency with these new tools, and the hard-won discipline to never confuse a fluent answer with a true one.

That is the work. It is the same work it has always been. And after twenty-seven years, I am more convinced than ever that it is worth doing carefully.

Enjoyed the read?

Explore the books behind the ideas, or get in touch.