WEBVTT - AI for Healthcare Administration

0:00:02.520 --> 0:00:06.840
<v Speaker 1>Bloomberg Audio Studios, Podcasts, radio news.

0:00:07.160 --> 0:00:10.440
<v Speaker 2>You're listening to Bloomberg Business Week with Carol Messer and

0:00:10.520 --> 0:00:14.239
<v Speaker 2>Tim Stenebek on Bloomberg Radio. Okay, get this, Carol. In

0:00:14.320 --> 0:00:16.640
<v Speaker 2>the US, more than twenty two point six million people

0:00:16.640 --> 0:00:18.959
<v Speaker 2>work in healthcare a lot. It's about fourteen percent of

0:00:19.000 --> 0:00:21.360
<v Speaker 2>the US working population. This is according to the Bureau

0:00:21.400 --> 0:00:25.040
<v Speaker 2>of Labor Statistics. Yeah, healthcare pretty huge. There's a ton

0:00:25.040 --> 0:00:27.480
<v Speaker 2>of money there and there are a lot of inefficiencies.

0:00:27.920 --> 0:00:28.680
<v Speaker 1>We hear about that.

0:00:28.760 --> 0:00:31.200
<v Speaker 2>You ever tried to go get an MRI like sent

0:00:31.400 --> 0:00:33.040
<v Speaker 2>from one healthcare provider to another.

0:00:33.280 --> 0:00:34.080
<v Speaker 1>It's like impossible.

0:00:34.240 --> 0:00:38.680
<v Speaker 2>You got a cd ROM seriously, like way they still

0:00:38.680 --> 0:00:40.839
<v Speaker 2>make that. Don't even get me started on faxes because yeah,

0:00:40.880 --> 0:00:42.000
<v Speaker 2>doctors are still doing that too.

0:00:42.280 --> 0:00:42.479
<v Speaker 3>I know.

0:00:42.720 --> 0:00:45.320
<v Speaker 2>Well. Anterior is trying to make the system more efficient.

0:00:45.320 --> 0:00:48.080
<v Speaker 2>It's an AI company. It's focused on healthcare administration. It's

0:00:48.120 --> 0:00:51.720
<v Speaker 2>backed by Anya, Sequoia Capital and more. We've got with

0:00:51.800 --> 0:00:55.400
<v Speaker 2>us doctor abdel Mamod, CEO of Interior. He joins us

0:00:55.520 --> 0:00:58.920
<v Speaker 2>from London, Doctor Mahmood, good to have you with us

0:00:59.040 --> 0:01:02.640
<v Speaker 2>this after this evening. Thanks for staying up late. We're

0:01:02.840 --> 0:01:07.319
<v Speaker 2>in sort of the chain of the healthcare ecosystem, do

0:01:07.360 --> 0:01:07.759
<v Speaker 2>you fit.

0:01:09.400 --> 0:01:10.960
<v Speaker 3>Yeah, well, lovely to meet you and thank you for

0:01:11.480 --> 0:01:15.040
<v Speaker 3>having me on. So we tut of the main area

0:01:15.080 --> 0:01:18.399
<v Speaker 3>is administration. So we were born to tackle the trillion

0:01:18.440 --> 0:01:20.720
<v Speaker 3>dollars that I think the US spends on healthcare administration

0:01:20.800 --> 0:01:23.800
<v Speaker 3>every single year. And administration is made up of many,

0:01:23.800 --> 0:01:25.360
<v Speaker 3>many workflows, but at the core of it, and I

0:01:25.400 --> 0:01:27.640
<v Speaker 3>think you mentioned this statistic, there's a lot of manual labor.

0:01:27.640 --> 0:01:31.000
<v Speaker 3>There's a lot of very manual, tedious things involving factes,

0:01:31.000 --> 0:01:34.120
<v Speaker 3>and yes that's true. Doctors still use factors today. So

0:01:34.600 --> 0:01:36.640
<v Speaker 3>we've started mainly on the health insurance because we feel

0:01:36.680 --> 0:01:38.880
<v Speaker 3>like a lot of that administrative workflows and burdens starts

0:01:38.880 --> 0:01:41.840
<v Speaker 3>from that side, and we focus there, and maybe they're

0:01:41.880 --> 0:01:43.360
<v Speaker 3>the easy way to say it is we're focused on

0:01:43.360 --> 0:01:45.480
<v Speaker 3>wherever there's a fax machine that a kind of a

0:01:45.520 --> 0:01:48.000
<v Speaker 3>highly trained nurse or adoptor needs to spend hours reading,

0:01:48.360 --> 0:01:49.560
<v Speaker 3>that's the bit that we make faster.

0:01:51.160 --> 0:01:54.160
<v Speaker 1>Okay, So I feel like folks have been trying to

0:01:54.200 --> 0:01:59.880
<v Speaker 1>figure out how to modernize healthcare records and then also

0:02:00.120 --> 0:02:04.720
<v Speaker 1>protect patient privacy. So tell us about how you guys

0:02:04.840 --> 0:02:07.560
<v Speaker 1>are going to actually figure this one out. Why you

0:02:07.760 --> 0:02:10.640
<v Speaker 1>versus everybody else who's been talking about it and trying it.

0:02:10.960 --> 0:02:12.000
<v Speaker 1>Why you can do this.

0:02:13.400 --> 0:02:15.280
<v Speaker 3>Yes, that's a great question, and I think you touched

0:02:15.280 --> 0:02:19.079
<v Speaker 3>on something important there, which is patient privacy and interoperability, right.

0:02:19.280 --> 0:02:21.440
<v Speaker 3>And the reason we have facts today is actually because

0:02:21.440 --> 0:02:23.840
<v Speaker 3>it's one of the most secure forms of getting two

0:02:23.919 --> 0:02:26.320
<v Speaker 3>systems that don't know each other kind of aren't connected

0:02:26.360 --> 0:02:29.120
<v Speaker 3>electronically to be able to communicate in a very safe way.

0:02:29.440 --> 0:02:31.160
<v Speaker 3>The result in issue of that is that when you

0:02:31.200 --> 0:02:33.760
<v Speaker 3>send over a fax, the system on the other side

0:02:33.880 --> 0:02:36.640
<v Speaker 3>receives the facts, but it can't ingest the data. Right.

0:02:36.680 --> 0:02:39.200
<v Speaker 3>It's four hundred pages of medical information that only a

0:02:39.240 --> 0:02:41.239
<v Speaker 3>human today can read, so you hire a doctor or

0:02:41.360 --> 0:02:44.160
<v Speaker 3>nurse to just read it. And I think the bit

0:02:44.160 --> 0:02:45.600
<v Speaker 3>that's made us stand out in the market and we

0:02:46.080 --> 0:02:47.919
<v Speaker 3>have the leading product in the market that we be

0:02:48.000 --> 0:02:50.760
<v Speaker 3>able to understand unstructured medical data and we're almost like

0:02:50.840 --> 0:02:53.240
<v Speaker 3>recreate the structured version of that and do interesting things

0:02:53.280 --> 0:02:57.760
<v Speaker 3>with that downstream, such as bio authorizations, payment integrity, and

0:02:57.760 --> 0:03:01.160
<v Speaker 3>a few other certain workflows. And that's the the bit

0:03:01.240 --> 0:03:04.240
<v Speaker 3>that's the most valuable here is you can maintain if

0:03:04.280 --> 0:03:06.120
<v Speaker 3>you still use facts, and we should be getting rid

0:03:06.120 --> 0:03:07.960
<v Speaker 3>of facts, but it remains a kind of a core

0:03:08.000 --> 0:03:10.520
<v Speaker 3>part of of the US healthcare system. It allows you

0:03:10.520 --> 0:03:12.480
<v Speaker 3>to kind of still have that digitized and still have

0:03:12.480 --> 0:03:15.200
<v Speaker 3>that interoperable to that you know, of most valuable assets

0:03:15.240 --> 0:03:18.920
<v Speaker 3>and resources nurses, doctors, pharmacists on not spending hours reviewing facts.

0:03:18.840 --> 0:03:22.200
<v Speaker 2>Is this for the US healthcare system only?

0:03:23.560 --> 0:03:26.080
<v Speaker 3>Yes? I think the US healthcare system is quite unique

0:03:26.080 --> 0:03:28.040
<v Speaker 3>in the sense that it spends so much on these

0:03:28.080 --> 0:03:30.760
<v Speaker 3>administrative workflows that are quite outdated in a sense.

0:03:30.880 --> 0:03:31.120
<v Speaker 1>Yeah.

0:03:31.400 --> 0:03:34.680
<v Speaker 2>Yeah, So where were you in your medical school journey

0:03:35.520 --> 0:03:39.280
<v Speaker 2>or in your career where you discovered that the US

0:03:39.320 --> 0:03:41.400
<v Speaker 2>has these inefficiencies. Take us to that moment.

0:03:42.360 --> 0:03:45.680
<v Speaker 3>Yeah. So I trained in the UK, and somewhere along

0:03:45.880 --> 0:03:47.920
<v Speaker 3>the journey I realized that, you know, the impact you

0:03:47.960 --> 0:03:50.400
<v Speaker 3>can have in healthcare is probably less about how better

0:03:50.440 --> 0:03:52.760
<v Speaker 3>I got as a doctor clinically, but more about the

0:03:52.840 --> 0:03:55.320
<v Speaker 3>environment and ecosystem and the tools that we use as condisions.

0:03:55.400 --> 0:03:58.800
<v Speaker 3>So I embarked on a master's degree in computer science

0:03:58.800 --> 0:04:01.040
<v Speaker 3>and spend some time actually Facebook and then later at

0:04:01.040 --> 0:04:03.720
<v Speaker 3>Google working on some healthcare applications, and that's where I

0:04:03.760 --> 0:04:06.200
<v Speaker 3>got exposed as you know, those are American companies. I

0:04:06.280 --> 0:04:09.080
<v Speaker 3>got exposed to the US healthcare system and I saw that, yes,

0:04:09.120 --> 0:04:10.520
<v Speaker 3>there are issues of this in Europe, and there are

0:04:10.520 --> 0:04:13.080
<v Speaker 3>issues in Canada and other places, but in the US

0:04:13.120 --> 0:04:16.080
<v Speaker 3>it's really magnified and it's blown up in the last

0:04:16.120 --> 0:04:18.080
<v Speaker 3>ten years or so. And that's what got me exposed.

0:04:18.080 --> 0:04:20.119
<v Speaker 3>And that's when some of the large language model staff

0:04:20.120 --> 0:04:22.520
<v Speaker 3>and GENERALLYVII that was coming out, felt like that was

0:04:22.520 --> 0:04:24.920
<v Speaker 3>the most exciting thing you could apply large language models to,

0:04:25.200 --> 0:04:28.520
<v Speaker 3>which is this huge problem of unstructured data. That guess what,

0:04:29.120 --> 0:04:31.800
<v Speaker 3>large language models inherently have that capability that we never

0:04:31.800 --> 0:04:34.240
<v Speaker 3>had before of structuring unstructured data.

0:04:35.200 --> 0:04:37.800
<v Speaker 1>Go ahead, well, you know, it's interesting. I do think

0:04:37.839 --> 0:04:40.640
<v Speaker 1>about AI and large language models and how they might

0:04:40.680 --> 0:04:43.200
<v Speaker 1>be able to look at an X ray, you know,

0:04:43.320 --> 0:04:46.200
<v Speaker 1>and compare it to all these others, you know, in

0:04:46.240 --> 0:04:49.000
<v Speaker 1>a database and maybe figure out when there really is

0:04:49.040 --> 0:04:52.039
<v Speaker 1>something wrong and do it rather quickly versus waiting for

0:04:52.120 --> 0:04:54.240
<v Speaker 1>humans or maybe there's a doctor who's tired and it

0:04:54.279 --> 0:04:57.880
<v Speaker 1>doesn't come out so well. Having said that, I'm reading

0:04:57.880 --> 0:05:01.400
<v Speaker 1>from the Sequoia website says that what you guys are

0:05:01.440 --> 0:05:04.200
<v Speaker 1>doing handles all the work that comes before the clinical

0:05:04.240 --> 0:05:07.520
<v Speaker 1>decision making, where it might take bless you a nurse

0:05:07.560 --> 0:05:10.440
<v Speaker 1>several hours to track down, organize, and summarize hundreds of

0:05:10.480 --> 0:05:13.839
<v Speaker 1>pages of records they need, Interior can do it in seconds,

0:05:13.960 --> 0:05:16.600
<v Speaker 1>allowing clinicians to increase their output tenfold and focus on

0:05:16.640 --> 0:05:19.680
<v Speaker 1>what matters most, helping patients love that love that mission.

0:05:20.040 --> 0:05:24.000
<v Speaker 1>How do we know that LLMS that AI is going

0:05:24.040 --> 0:05:26.320
<v Speaker 1>to make the right clinical decisions?

0:05:27.960 --> 0:05:30.640
<v Speaker 3>That's a great question, and that's because you never let

0:05:30.680 --> 0:05:33.560
<v Speaker 3>the large language model make the clinical decision. That is

0:05:33.560 --> 0:05:35.960
<v Speaker 3>the key bit here, which is everything up until the

0:05:36.000 --> 0:05:38.920
<v Speaker 3>clinical decision is the bit that takes the most time

0:05:38.920 --> 0:05:40.360
<v Speaker 3>and the most effort. So if you look at the

0:05:40.440 --> 0:05:43.480
<v Speaker 3>nurses workflow right, they'll spend kind of of ninety five

0:05:43.480 --> 0:05:46.640
<v Speaker 3>percent of the time logging into the twenty different systems

0:05:46.720 --> 0:05:50.080
<v Speaker 3>right downloading the right PDF files, finding the clinical reference

0:05:50.080 --> 0:05:53.400
<v Speaker 3>and documentation and guidelines, mapping it all out, finding needles

0:05:53.400 --> 0:05:55.960
<v Speaker 3>in a haystack across hundreds of pages to then just

0:05:56.040 --> 0:05:58.200
<v Speaker 3>have all the information gathered so that he or she

0:05:58.240 --> 0:06:00.720
<v Speaker 3>can make that decision right. So what we tell our

0:06:00.800 --> 0:06:02.680
<v Speaker 3>nurses that use the product and the reason they love

0:06:02.720 --> 0:06:05.080
<v Speaker 3>it is like each one of you gets an intern

0:06:05.240 --> 0:06:08.359
<v Speaker 3>that overnight really smart intern that overnight has prepped all

0:06:08.400 --> 0:06:10.360
<v Speaker 3>the nurs right, has gone through everything, and then lays

0:06:10.400 --> 0:06:12.520
<v Speaker 3>it up for you. So you make the call, right,

0:06:12.720 --> 0:06:15.680
<v Speaker 3>and that what that is is you ensure that nurses

0:06:15.720 --> 0:06:18.600
<v Speaker 3>and doctors go through administration much much faster. But they

0:06:18.640 --> 0:06:20.360
<v Speaker 3>focus on the bit that they're great at the top

0:06:20.360 --> 0:06:22.400
<v Speaker 3>of their license, the thing they went to medical school

0:06:22.440 --> 0:06:25.200
<v Speaker 3>or nursing school for many years for.

0:06:26.760 --> 0:06:29.480
<v Speaker 2>So I want to know about training these models because

0:06:30.680 --> 0:06:32.479
<v Speaker 2>there's a big issue when it comes to what this

0:06:32.480 --> 0:06:34.360
<v Speaker 2>stuff is trained on. How do you do this when

0:06:34.360 --> 0:06:36.400
<v Speaker 2>patient data is confidential?

0:06:37.600 --> 0:06:40.120
<v Speaker 3>Yeah, so flatter out. We don't absolutely don't train on

0:06:40.200 --> 0:06:42.640
<v Speaker 3>any customer data, and I don't think companies would even

0:06:42.640 --> 0:06:44.440
<v Speaker 3>allow you to, right, So that's all in the contract

0:06:44.520 --> 0:06:46.279
<v Speaker 3>that we sign and so on. I think the world

0:06:46.279 --> 0:06:49.080
<v Speaker 3>has moved on in a sense of the previous issue

0:06:49.120 --> 0:06:51.240
<v Speaker 3>with AI was that you needed to train a lot

0:06:51.240 --> 0:06:53.680
<v Speaker 3>of medical data to even have something viable. But I'm

0:06:53.680 --> 0:06:56.840
<v Speaker 3>sure you've seen chat, gipt passes, the USMLI and all

0:06:56.880 --> 0:06:58.960
<v Speaker 3>these things. A lot of these general models that we

0:06:59.080 --> 0:07:02.520
<v Speaker 3>use off the shelf right, they're actually pretty good clinically.

0:07:03.000 --> 0:07:05.200
<v Speaker 3>The trick is that last ten percent though, And the

0:07:05.200 --> 0:07:07.479
<v Speaker 3>reason why chat gipt isn't doing any of this stuff

0:07:07.520 --> 0:07:09.640
<v Speaker 3>today is because it maybe gets sixty percent of the

0:07:09.640 --> 0:07:12.440
<v Speaker 3>way there. How you orchestrate them, How you get multiple

0:07:12.480 --> 0:07:15.080
<v Speaker 3>large language models, including also built on top of the

0:07:15.120 --> 0:07:17.400
<v Speaker 3>decades of machine learning progress that we've had, on top

0:07:17.440 --> 0:07:21.600
<v Speaker 3>of databases and and other kind of computational techniques. How

0:07:21.600 --> 0:07:23.120
<v Speaker 3>do you stitch that all together in end to end

0:07:23.160 --> 0:07:27.000
<v Speaker 3>experience that feels like almost almost like a kind of

0:07:27.000 --> 0:07:29.400
<v Speaker 3>completes the picture with a great user experience as well.

0:07:29.520 --> 0:07:31.000
<v Speaker 3>That's the bit we focus on, and that's the bit

0:07:31.040 --> 0:07:32.239
<v Speaker 3>that actually delivers them value.

0:07:32.480 --> 0:07:35.120
<v Speaker 1>Well, let us know how things are going. Doctor Abdel Mahmoud,

0:07:35.400 --> 0:07:37.600
<v Speaker 1>Chief Executive Officer of Interior, joining US