A (2:47)

Oh, very good. You need scale for the majority of AI, But a few of us are also focusing on how to resolve what used to be called the curse of dimensionality, having too many variables and not enough data. And so there's two facets to it one facet that we hear for ChatGPT and these large language models throw more data into model compute longer and you'll see more emerging properties. And that's true, that's correct. But at the same time we're facing for these large language models issues such as counterfactual. There's just not enough examples of what's wrong. But to tell the large language model that's a wrong way to write a sentence, a wrong way to conceive an idea. So there are other methods that could at scale create these negative data sets in which they could learn better. But we can also think of enemy countries or organization that would generate erroneous data set that appear to be true. And then unless we're very smart right now at building our next generation of large language models in us, we've got to ensure if we take our data from different sources, whether they have been seeded in order to torp us and have lesser models than those that are developed in enemy countries. But coming back to an area where we're really good, at the University of Utah, it's on addressing what's called large P here. P is not a statistical P value, but it's a feature space. It's large P in computer science and small N which is the number of subjects and number of instances you have. And that used to be called the curse of dimensionality. And for those of you that are familiar with maths that are offered in inductive deductive science like engineering, physics, maths and first grade of university called linear model, it's pretty straightforward. If you have the same number of equation as you have variables, you have a well defined data set, you have more equation than variables, it's over defined and sometimes there's not even a solution to that. And if you have less equations than variables, you're under defined. So there are instances for example in personal genomics where you may have 25,000 gene expression measured in a patient only twice, maybe before and during treatment. Can you make some sense out of that? Now we're speaking of n equal to two instances in one patient and 25,000 variables. In the past this was, it was painful, it was called curse of dimensionality. It was a statistical nightmare. But our team and some of our colleagues have reduced new statistics and it's called the blessing of dimensionality.

C (5:50)

And we can solve the blessing of dimensionality.

A (5:54)

Yes. And you can identify altered pathways in patients due to pain, due to some drugs. And with these n of 1 trials, single subject studies, we can resolve why are they Resistant to therapy or why are they super responders with a simple blood sample, maybe with an provocation cellular assay, ex vivo in petri dishes, in a control essay. And that's all it takes. And overnight you get a solution to their edge case, which can be in a super responder or a resistance to a therapy or predicting a resistance or super response before they're going three months of treatment and averting unnecessary costs and especially if it's cancer, unnecessary treatment and addressing the real treatment. So it's very exciting times where you do have this question of large data sets and mathematics that could be used with AI against large data sets. But there's also the pearls of wisdom to be obtained in smaller data sets as well, which in the past were impervious to our unedited cultures.

C (7:02)

This is fantastic. I love this discussion. And what have you seen at the meeting that has surprised you?

A (7:07)

Well, we've got some keynote speakers that speak with substance and brevity, which in science we need more of those.

C (7:15)

Brilliant. We love substance and brevity.

A (7:18)

Exactly. So we've got some of these corporate speakers that have this skill set. This is the skill set. Come more for the. I'm a physician, engineer, informatician. I come more from the scientific side though I'm also an entrepreneur.

C (7:31)

Oh, you are?

A (7:32)

Yes.

C (7:32)

What are you doing out there?

A (7:34)

A long time ago, when I was a resident in medicine, I conceived the first pen computer for physicians with an AI on it in 1991. A what? An AI over pen computing with handwriting recognition. The system is called Purkinje. The company still exists. It was used in about 4,000 clinics and hospitals. And we had.

C (7:54)

So it's like digital writing.

A (7:56)

Digital writing, yes. And handwriting recognition.

C (7:59)

And handwriting recognition.

A (8:00)

So instead of an electronic record based on. On a system of record, it was an electronic record based on a system of intelligence. Conventional electronic record that were conceived earlier and that are still mainstream in US such as Cerner and hipc, use standard terminologies and flat data dictionaries as their building blocks. We at that time inspired by the systematized nomenclature of medicine conceived by Roger Coulte, an alumni from Harvard that was actually faculty at my alma mater. We use a graph which has a relationship between concepts of medicine for which there's many synonyms. And then you would have 30 years ago already, what would it look like if we printed a print and all to what does it look like on a display where the real state is very small, especially in those days and therefore you don't identify the term of the concept with every form in which you communicate it, whether it's paper based or visual on a display. And it's got a graph for which AI can operate on. Basically what it gave us is within the end of the note. The physician would know how much it could bill for it was they could attribute the reason for which it would bill. And you may not be familiar with the billing models, but 30 years ago they were on paper. So the account receivables, since we bill daily electronically, would be a month or two earlier than the conventional form. And it paid for the system. So that was the business model. That's awesome. And it's still around even nowadays when people bill, like in radiology, about 50% of the bills are denied. The physicians or hospitals have one month to respond or else legally the insurance companies need not pay. Often there's a bolus of rejection of bills in mid November to mid December. It just happens to coincide with holidays. And that means that administrative staff need to be staffed to such a height yearly to address these deliberate bonuses from health insurers that deny on specific times, because if you forget to bill within 30 days, they're legally not obliged to pay for it. So there's a lot of contortion into the billing system. So I'm an expert at billing as well as well as genomic medicine.

C (10:31)

Man, look at that. Wow, that's impressive. Very impressive. Eve, what's the riskiest thing you've ever done?

A (10:42)

That's a good one. So I'm considered a transformative leader. I've started the units of biomedical informatics in five institutions. And at Columbia University, my alma mater, I started the unit of the Northeast Biodefense center, which was across the northeast large institution, Princeton, Columbia University, NYU, you name it, Yale, etc. Etc. And then afterwards, at the University of Chicago, Illinois and Arizona, we built the first clinical data warehouses, as well as the service course for these institutions, as well as the collective intelligence that would be scalable so that I guess as I moved on with my career, there would be the next generation that would take over these systems. So we had sustainable informatics and AI in these institutions. So the riskiest part was always the next challenge. Because you promote within, you don't hire out when there's no problem. Right. So every promotion or every change of institution was actually a very high risk. And all these were starting new units. So I thought taking on the challenge of building up the oldest department of biomedical, Phrex, could be an issue because it's an old culture. And I had Been from my startup days when I was a resident in medicine, pretty much doing little startups in each institution and this was instead a big bureaucracy to move. So perhaps I would say it's always the latest challenge that's been the. The hardest.

C (12:15)

But you're like a bioinformatics remodeler. You take old departments and you shine them up and you make them modern.

A (12:21)

That's my last iteration. Right before that it was put in the sea. Putting the seed there was. It was scorched earth.

C (12:28)

Love that man. Super interesting. If you had the opportunity to go back to 20 years old, what advice would you give yourself?

A (12:37)

Well, first of all, I did my first company. I was 18. I was already president of the engineering student at that time. So the advice when I was 18, I don't know at that time I already had penned out that I'd make more money by dropping out of engineering or medicine. I was admitted in medicine as well then continuing on this academic trend.

C (12:56)

So you made that choice.

A (12:57)

You're like, I made that choice. You had to go down the route brought to multiple startups.

C (13:01)

So you were very bright from the beginning. So knowing everything you know now, what would you tell yourself at 20?

A (13:07)

I. I think it's the same thing I tell myself all the time. You have to brainwash me in order to. I knew that between engineering and medicine there'd be an opportunity. I thought at that time, my first company is about to bring the language of medicine to computers because it didn't work so well and we're doing statistics there and I didn't quite realize. Oh, but then there's the residency and then there's a postdoc and there's always another step of learning. So you'd have to brainwash me on the number of years it takes to train all these multiple disciplines to be effective and otherwise. It's just risk it, drop out and. And risk it all. Because I was already very skilled. I had started programming when I was 12 and I was very skilled. I could just drop out and done something. But then it's the advice I would not have known at that time would have been you've got to go to Silicon Valley or to. Or to Boston where there's a culture epicenter of creation, of creation, innovation and the funding for startups from one to the other y so that I didn't know that would be the best advice was drop, drop out and go to these sectors.

C (14:10)

That is fantastic.

A (14:12)

Holy smokes.

C (14:13)

That's great advice. That's great advice to your 20 year old self.

A (14:17)

So am I the most. The most provocative of your speakers?

C (14:20)

That's fantastic. Yeah, man, you're definitely at the top right now, folks. What do you think? I mean, we got Dr. Eve Lussier with us. Okay, so what's the place that people can reach out to you, learn more about your research, learn more about your companies and things like that.

A (14:36)

So I'm a public figure. You can find me on Wiki or LinkedIn.

C (14:40)

You've got a Wikipedia page, correct?

A (14:42)

Yes.

C (14:42)

You got to be a serious deal to have a Wikipedia page.

A (14:45)

Someone else felt like that. Right, because you can't invent.

C (14:48)

Exactly. So that's great.

A (14:50)

So I'm thankful to my alumni or other sponsors that have decided to put me there. It's a bit obsolete though. If you hear this, you could update the.

C (15:00)

They got to update it. Well, this will become a public record and we will have a transcript and so we could probably petition.

A (15:06)

The University of Utah has my. Has my email there as well. And we have a very successful program and the oldest program of training in biomedical informatics. So we have over 600 alumni. Many of our alumni have done, as I said, fantastic thing, pioneered informatics in Taiwan. I was there recently because the Czech Liquidity, which we will award an award next week at the American Medical Informatics association, also transformed the legal aspect of Taiwan. So believe it or not, you can get access to your electronic entire set of data across any provider in one place at any place anywhere in Taiwan, whether it's a pharmacist, a physician, you can decide which part of your record you have access to, another provider or the entirety, including everything, which is a hard thing to do. In the US Our interoperability is just on billing and surface level. Unless you're part of the network of IPIC or the network of cerner, where you do have surface interoperability between different implementation of EPIC or different implementation of cerners, you don't really have integrated interoperability at the concept level, but they have that in Taiwan. So Jack Lee is a phenomenal graduate of ours, but we had plenty of others. The founder, as I pointed out earlier, of the Department of Biomedical Informatics of Harvard, October Burnett, trained Zach Kohane, who's the founder of the New England journal AI of two years ago. Bring it back to AI. Right.

C (16:36)

Jeez. Well, there you have it, folks. We are with Dr. Yves Lussier. He's a department chair of Biomedical Informatics and professor of medicine at University of Utah. In the show notes, you'll find all the ways to get in touch with him, his Wikipedia page, everything that you want to do to find out what he's up to. And Eve, thanks so much for being with us.

A (16:58)

Thank you very much.

B (17:05)

Thanks for listening to Risk Never Sleeps. For the show, notes, resources and more information and how to transform the protection of patient safety. Visit us@SenseInet.com that's C E N S I N E T Com. I'm your host, Ed Gaudette. And until next time, stay vigilant because Risk Never Sleeps.