Dr. Michael Gao

And it was sort of actually, funnily, the doctors were arguing that the care that they were providing wasn't sufficiently captured by the coders.

And the coders and clinical documentation improvement specialist says, well, yeah, but that's because you're not documenting it, not because we're not capturing it after your documentation.

Right.

So to try to resolve this, I started trying to understand kind of how to build algorithms that could take in the raw clinical data and use that rather than just pure documentation as a source of truth.

So if your blood pressure is 180 over 100 and you get prescribed a high blood pressure medication, whether or not somebody wrote the words high blood pressure,

you can infer that it was likely something that was managed during that visit.

You still have to confirm it, get it documented, you know, double-check it, but you can sort of have an inkling that that must be, you know, true.

And at the same time, what was really fascinating was that...

So this is like 2017, 2018.

One of the sort of data science approaches, the Transformers approach, which is kind of the underpinning of LLMs, started making the waves.

So it was still a very early version called the BERT models.

Okay.

But...

It was actually really interesting to see how even very primitive implementations of this approach allowed models to understand clinical data and patterns in a way that, you know, kind of very sophisticated implementations of traditional statistics and machine learning didn't allow for.

Yeah, so I think one example of this is, let's say a traditional natural language processing approach is the, I'm going to ask you to identify all times the word bank appears in a book.

And so what you might do is you might just search for the letters B, A, and K and then, you know, kind of see how many times it pops up.

And that's kind of in the clinical world that might be akin to if the doctor writes the letters, you know, kidney injury, then you can have some sort of dictionary that it corresponds to and it sort of looks that up.

But if I say, hey, Jonathan, can you search for all the times banks appear?

But what I mean is specifically not just a word, but the bank that you go to deposit money as opposed to a river bank or a bank shot or a promise you can bank on.

The context is actually very hard to understand.