WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=UyyOoJmuATU

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/UyyOoJmuATU.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.640 --> 00:00:14.280
Hey everybody.

s3
00:00:14.280 --> 00:00:16.760
How's everyone doing today?

s4
00:00:16.760 --> 00:00:18.760
Good.

s5
00:00:19.760 --> 00:00:30.200
So, this is a bit about my background, but I think it's always really good to learn what the audience background is if it's more on the technical side, which I know this

s6
00:00:30.200 --> 00:00:32.720
conference is, healthcare side.

s7
00:00:32.720 --> 00:00:38.360
Let's get a quick show hands to see who Who here is on the healthcare side?

s8
00:00:38.440 --> 00:00:38.920
Ooh, wow.

s9
00:00:38.920 --> 00:00:39.600
That's a lot of you.

s10
00:00:39.600 --> 00:00:40.400
There's more than I expected.

s11
00:00:40.400 --> 00:00:41.040
Wow.

s12
00:00:41.040 --> 00:00:41.400
Okay.

s13
00:00:41.400 --> 00:00:45.840
Who here is on the technical does the genetic workflows?

s14
00:00:46.600 --> 00:00:47.160
Wow, okay.

s15
00:00:47.160 --> 00:00:48.080
Overlap.

s16
00:00:48.080 --> 00:00:48.360
All right.

s17
00:00:48.360 --> 00:00:50.440
As should be expected at this conference.

s18
00:00:50.440 --> 00:00:55.240
Who here has models running right now somewhere doing some work?

s19
00:00:55.240 --> 00:00:57.040
Wow.

s20
00:00:57.040 --> 00:00:58.480
It's like the whole audience.

s21
00:00:58.480 --> 00:00:59.080
Okay.

s22
00:00:59.080 --> 00:00:59.440
All right.

s23
00:00:59.440 --> 00:01:00.880
So, I know who I'm talking to.

s24
00:01:00.880 --> 00:01:01.880
Wonderful.

s25
00:01:01.880 --> 00:01:03.840
This is the right crowd.

s26
00:01:03.840 --> 00:01:08.800
So, what is the goal of this?

s27
00:01:08.800 --> 00:01:15.960
Like if we're working in healthcare and we're doing some genetic workflows, we have to keep in mind the goal.

s28
00:01:15.960 --> 00:01:22.960
And that goal, at least from my perspective, is to drive the overall cost down.

s29
00:01:22.960 --> 00:01:31.800
So, in this talk will be about insurance cost specifically, the cost of of interacting with insurance.

s30
00:01:31.800 --> 00:01:41.400
Um, but also to improve the patient experience because insurance relates back to the patient and how they experience the whole process.

s31
00:01:41.400 --> 00:01:43.480
So, let's keep that in mind when we're solving problems.

s32
00:01:43.480 --> 00:01:56.400
I know that if we're on the technical side, we like to get really experimental with things and get excited about just this the technology itself, but has to be grounded in one of these two concepts.

s33
00:01:58.720 --> 00:02:06.960
So, a little trip down memory lane just starting from this long journey of AI machine learning and where we are today.

s34
00:02:06.960 --> 00:02:12.320
And we're obviously going to talk about the last point, this agentic execution layer.

s35
00:02:12.320 --> 00:02:23.280
But, we can see this evolution from the neuron convolutional neural networks large-scale unsupervised learning back in 2011 2012.

s36
00:02:23.280 --> 00:02:32.480
Then, the introduction of attention is all you need, one of my favorite titles for a paper, the introduction of the the transformer.

s37
00:02:32.480 --> 00:02:38.160
Then, we go into this modern chat interface with these large language models.

s38
00:02:38.160 --> 00:02:50.840
And then, finally with the Claude codes, the Codexes, and our system internally, and a lot of systems you have open claw, Hermes agent, all all that kind of stuff.

s39
00:02:50.840 --> 00:02:55.320
Which really brings into the into the picture this execution layer.

s40
00:02:55.320 --> 00:03:04.519
So, we're going to be talking about how to do this execution layer safely and reliably in health care.

s41
00:03:06.760 --> 00:03:21.320
So, let's think back to some of the earlier examples of like getting really excited about some AI technology and then realizing it has all these little bits and pieces which make it a lot more trickier than maybe it is is obvious at first.

s42
00:03:21.320 --> 00:03:25.400
So, like you have a check and you want to cash it.

s43
00:03:25.400 --> 00:03:27.640
You want to deposit it into your bank account.

s44
00:03:27.640 --> 00:03:32.000
Um you might say, "Oh, we have solved the handwritten digit problem.

s45
00:03:32.000 --> 00:03:34.640
We can recognize digits from zero to nine."

s46
00:03:34.640 --> 00:03:35.040
Right?

s47
00:03:35.040 --> 00:03:41.560
Wow, oh oh, so now we're ready to um deposit this check into this person's account and transfer money.

s48
00:03:41.560 --> 00:03:50.760
Well, not quite because as you dig in a little bit deeper you see that you have to identify all sorts of characters in the check.

s49
00:03:50.760 --> 00:03:56.040
You have to make sure it matches up with all these other pieces of the infrastructure.

s50
00:03:56.040 --> 00:04:04.360
You have to make sure that it is um that it is um going to the target account that you're interested in.

s51
00:04:04.360 --> 00:04:08.680
So, parts of this can be thought of as as the harness.

s52
00:04:08.680 --> 00:04:25.840
Um in claims, we have a similar challenge that there's a lot of these little AI steps involved in fulfilling that whole patient journey of eligibility to getting the insurance company to deposit

s53
00:04:25.840 --> 00:04:28.280
money in the provider's bank account.

s54
00:04:28.280 --> 00:04:30.240
A whole bunch of little steps.

s55
00:04:30.240 --> 00:04:39.560
And we have to make sure that we're safely doing this, that we're operating like in these tight con these strict confinements.

s56
00:04:42.160 --> 00:04:54.120
So, another thing that comes up, just sort of setting the stage, is that um multimodal context in comes up very frequently with claims.

s57
00:04:54.120 --> 00:05:04.919
So, you might have an image and it might seem like at at first for cost reasons or something else that you'd want to take that image and reduce it down to the findings, like

s58
00:05:04.919 --> 00:05:10.080
here's the anatomy in the image, and maybe even extract some geometries from that anatomy.

s59
00:05:10.080 --> 00:05:11.600
Here's pathologies.

s60
00:05:11.600 --> 00:05:20.840
And then you would take that and then combine it with some other machine learning with some other data in a different downstream machine learning model.

s61
00:05:20.840 --> 00:05:23.919
Like um EHR.

s62
00:05:23.919 --> 00:05:30.400
And that might make sense from a cost perspective and also just like model capabilities.

s63
00:05:30.400 --> 00:05:36.000
Um but in a lot of situations, it you lose context.

s64
00:05:36.000 --> 00:05:48.280
So, it might be that you're extracting all this information and missing something that relates to some downstream procedure that you didn't that wasn't the upstream model wasn't aware of it.

s65
00:05:48.280 --> 00:05:53.240
So, that introduces this concept of just multimodal processing.

s66
00:05:53.240 --> 00:06:01.600
So, uh uh another place this comes up in healthcare, but not related to anatomy or anything like that, is desktop use.

s67
00:06:01.600 --> 00:06:12.040
You can see that sometimes, you know, things are buttons or or or, you know, it might have some shapes that are only obvious when you do this multimodal.

s68
00:06:16.880 --> 00:06:17.360
All right.

s69
00:06:17.360 --> 00:06:22.000
What is the agentic execution layer?

s70
00:06:22.000 --> 00:06:25.720
So, this can this can take on a lot of different forms.

s71
00:06:25.720 --> 00:06:31.080
It is the ability of this model to take actions.

s72
00:06:31.080 --> 00:06:36.720
So, it might be you're starting out with um database queries.

s73
00:06:36.720 --> 00:06:38.480
And let's say it's just completely open.

s74
00:06:38.480 --> 00:06:48.919
You're querying the database, you're finding your schema, you're figuring out what this what the data looks like, and then you might even have access to your code.

s75
00:06:48.919 --> 00:06:53.200
So, then you're querying your code with the respect to your data.

s76
00:06:53.200 --> 00:06:58.680
Uh and you might actually, in our system or other systems, you might be making insurance transactions.

s77
00:06:58.680 --> 00:07:00.600
You might be making a phone call.

s78
00:07:00.600 --> 00:07:02.919
You might be looking at a web portal.

s79
00:07:02.919 --> 00:07:05.800
You might be interfacing with an EHR.

s80
00:07:05.800 --> 00:07:08.120
These are all actions you can take.

s81
00:07:08.120 --> 00:07:11.480
And some of these actions have right implications.

s82
00:07:11.480 --> 00:07:20.160
If you're interfacing with different PMSs, you're going to the desktop, you can have at least user logs at the minimum.

s83
00:07:20.640 --> 00:07:23.520
Uh and then the next concept is memory.

s84
00:07:23.520 --> 00:07:28.919
So, cloud code or codex, they use local memory, they write to your desktop.

s85
00:07:28.919 --> 00:07:38.240
In enterprise healthcare, we can't really do this, so we do memory in a database, just so we have that logical separation.

s86
00:07:40.320 --> 00:07:56.240
Uh important concept here is that when you're introducing new and improved better models more sophisticated more parameters you can't you can't just replace the model and assume it's going to be better

s87
00:07:56.240 --> 00:08:10.320
it's different right it's a it's a on certain evals it's a better model as measured by these different metrics but it doesn't necessarily mean it's better right for all the situations that you want it to be better at because of the way you've designed your system

s88
00:08:10.320 --> 00:08:22.120
so you really have to redo everything from scratch just make sure your evals your testing your validation is all set up so that you can introduce these new models

s89
00:08:22.120 --> 00:08:24.640
and not break your system.

s90
00:08:31.120 --> 00:08:47.040
So this concept of harness different groups have different definitions of this so I'm going to use a super broad definition here which is like all the different nuts and bolts that that surround this agentic

s91
00:08:47.040 --> 00:09:01.600
reasoning and that is the concept of memory that we discussed the different tools the checks the permissions the handoffs the evals but also in the context of health care and claims

s92
00:09:01.600 --> 00:09:04.160
it's x12.

s93
00:09:04.160 --> 00:09:24.400
So just like we have these old school languages or formats like COBOL or other stricter maybe strict languages typescript llms really thrive they work well and when they're confined they have clear

s94
00:09:24.400 --> 00:09:39.800
limited values that they can predict and x12 is exactly this so it provides this underlying structure this contract between what you're trying to communicate and the insurance company.

s95
00:09:44.200 --> 00:09:56.200
So, when you're reasoning in this in this healthcare, your your objective is to do something with handle a claim or research your EHR with respect to claims.

s96
00:09:56.200 --> 00:09:59.760
It might be that you have like 50 steps or something like that.

s97
00:09:59.760 --> 00:10:01.720
There's a lot of different steps.

s98
00:10:01.720 --> 00:10:10.840
And so, you can um you at at each of those steps as you make mistakes, those mistakes can propagate down your system.

s99
00:10:10.840 --> 00:10:15.920
Um and so, it's very good to have something grounded that can be rejected to.

s100
00:10:15.920 --> 00:10:24.480
So, if you have a really strict got guardrails, you can reject something that happens that's incorrect.

s101
00:10:27.600 --> 00:10:35.680
So, if you're reasoning over, let's say, the previous example, 50 steps, that and they're multimodal.

s102
00:10:35.680 --> 00:10:37.560
You're considering and everything like that.

s103
00:10:37.560 --> 00:10:39.320
That can get really expensive.

s104
00:10:39.320 --> 00:10:41.320
It can also take a really long time.

s105
00:10:41.320 --> 00:10:43.560
And folks might not want to wait.

s106
00:10:43.560 --> 00:10:47.080
You know, it could be too expensive and people don't want to wait that long.

s107
00:10:47.080 --> 00:10:53.640
And each time it each step is an opportunity to introduce an error and you can have problems.

s108
00:10:53.640 --> 00:11:04.600
Um but, if you hardcode your whole system, you say we're you're going to throw out this whole agentic process, you limit yourself or your code can explode to be just unmanageable.

s109
00:11:04.600 --> 00:11:12.600
So, now all of a sudden, you just have this crazy bloat and you have to have this giant engineering team, which poses its own problems.

s110
00:11:12.600 --> 00:11:28.200
Um so, what we want to do is strike this balance between what we should be completely free, like um with just pure agentic reasoning and execution, and what is hardcoded.

s111
00:11:28.200 --> 00:11:41.120
So, we do that internally with introducing memory, just this uh partner level memory, organizational organized memory and user memory.

s112
00:11:41.120 --> 00:11:50.080
So, we say if a user, we find people in in multi-site health organizations, they tend to do the same thing day after day.

s113
00:11:50.080 --> 00:12:01.000
And it might be if they mention a few words, "Oh, they usually do eligibility and they usually do it within this context, they probably mean this."

s114
00:12:01.000 --> 00:12:01.280
Right?

s115
00:12:01.280 --> 00:12:05.200
Where another user, they probably mean that.

s116
00:12:05.200 --> 00:12:16.440
So, we want to be really careful here because as you introduce memory, you also persistent memory across chats, across days, you also introduce bias.

s117
00:12:16.440 --> 00:12:23.080
So, maybe that person doesn't want to do the exact same thing that they did yesterday and now you steer them to do the exact same thing they did yesterday.

s118
00:12:23.080 --> 00:12:24.160
That's a problem.

s119
00:12:24.160 --> 00:12:31.600
So, you want to strike a balance somewhere in there and you want to make sure that the use any user can break out of this.

s120
00:12:35.880 --> 00:12:42.080
So, for folks that are unfamiliar with the whole claim life cycle, it's many steps.

s121
00:12:42.080 --> 00:12:46.800
So, each step does have an X12 correspondence with it.

s122
00:12:46.800 --> 00:12:52.440
Starting with the schedule, when let's say you're showing up to the doctor's office before you even show up.

s123
00:12:52.440 --> 00:12:54.560
That's insurance starts then.

s124
00:12:54.560 --> 00:13:04.000
Um when you're getting treated, that also relates to insurance, what you you know, the different procedures that you are candidate for depending on your insurance.

s125
00:13:04.000 --> 00:13:10.880
Um your documents, sometimes the x-ray itself is the document and you would send proof of that in.

s126
00:13:10.880 --> 00:13:12.400
Submitting the claim.

s127
00:13:12.400 --> 00:13:18.280
And then finally getting the payment in the provider's bank.

s128
00:13:22.320 --> 00:13:32.240
So, this I think this concept is a little bit I I found it to be obvious in retrospect but let me talk you through it.

s129
00:13:32.240 --> 00:13:45.840
Maybe you find it's it's interesting or not but if you're calling an insurance company that it that boils down to a transaction an X12 transaction.

s130
00:13:45.840 --> 00:13:48.480
You'd say hey this is the patient I'm talking about.

s131
00:13:48.480 --> 00:13:48.840
Great.

s132
00:13:48.840 --> 00:13:52.800
That's a like an eligibility request a 270.

s133
00:13:52.800 --> 00:14:01.520
Oh I need to do you're requesting a claim status or whatever it is you're doing that has an X12 grounding.

s134
00:14:01.520 --> 00:14:05.720
And that is the whole concept here this this X12 harness.

s135
00:14:05.720 --> 00:14:22.640
So you call the insurance company you have an agent interact with a desktop you have an agent interact with the browser your imaging system that's a 275 and and and your insurance your bank your ACH so

s136
00:14:22.640 --> 00:14:27.520
that's not X12 but it's still that structured transaction.

s137
00:14:32.470 --> 00:14:32.920
[snorts]

s138
00:14:32.920 --> 00:14:37.360
So this is just a reiteration of these different transactions.

s139
00:14:39.240 --> 00:14:49.240
And the other beautiful thing about it it's not you know it can you ask an agent to do something let's say you're genetically programming or let's say you're just you're you're programming

s140
00:14:49.240 --> 00:15:01.280
how you know maybe half the companies I spoke with here still program today just everything by hand and they use these clock code or code X4 research.

s141
00:15:01.400 --> 00:15:04.440
If you look up any of these transactions they're all public.

s142
00:15:04.440 --> 00:15:08.240
Like this is not the beautiful thing about this is like this is not my schema.

s143
00:15:08.240 --> 00:15:13.080
If you ask agents to make a schema for you you're going to get like all sorts of stuff.

s144
00:15:13.080 --> 00:15:20.160
But now if we ground it in something standard you can look up all of these and you would know just right off the bat my schema.

s145
00:15:20.160 --> 00:15:24.240
Let's say you're a new engineer coming in, like, you know.

s146
00:15:30.640 --> 00:15:43.320
So, X12 is a is a system of rules and it doesn't mean that when an insurance company gives you an X12, it's true.

s147
00:15:43.320 --> 00:15:52.120
So, that concept is when insurance company tells you something, it's coming from different teams potentially.

s148
00:15:52.120 --> 00:15:56.200
They can have an engineering team that's It could be even a different company.

s149
00:15:56.200 --> 00:16:07.440
A different company that the insurance company contracts out designed their web browser, their phone system, or their X12 layer, or their fire.

s150
00:16:07.440 --> 00:16:12.080
And we have to understand that there's no ground truth.

s151
00:16:12.080 --> 00:16:19.280
They also within all of these systems, they can they can all actually agree on the wrong information as well.

s152
00:16:19.280 --> 00:16:21.720
Like, let's say they all say this patient is covered.

s153
00:16:21.720 --> 00:16:23.560
All three You call them.

s154
00:16:23.560 --> 00:16:26.960
You look in the browser and the X12 and they all say, "Yes, this patient is covered."

s155
00:16:26.960 --> 00:16:32.680
And then you treat the patient, they say claim is denied due to the patient wasn't covered during that time.

s156
00:16:32.680 --> 00:16:42.080
Um so, they can all disagree, but sometimes you'll learn some idiosyncrasies of these different payers that some of these systems are more reliable than others.

s157
00:16:42.080 --> 00:16:51.960
But, regardless of if it originates as an X12 or not, you can boil all those transactions down to your own internal semi-correct X12.

s158
00:16:51.960 --> 00:16:59.480
Correct until downstream evidence proves it otherwise uh to be incorrect.

s159
00:17:01.080 --> 00:17:02.760
So, just a little bit more on that.

s160
00:17:02.760 --> 00:17:07.720
So, and any of the X12, any of the information coming from the insurance company, any time can be wrong.

s161
00:17:07.720 --> 00:17:09.640
It can be updated later.

s162
00:17:09.640 --> 00:17:11.880
So, have fun.

s163
00:17:13.520 --> 00:17:19.400
This is just an example of what it would look like if you're um in that patient journey.

s164
00:17:19.400 --> 00:17:29.760
You're you're trying to figure out how much you would pay as a patient up front, and it's very important for your experience going to the doctor.

s165
00:17:32.520 --> 00:17:40.320
And then the different treatments that you have in that clinic can oop.

s166
00:17:40.480 --> 00:17:41.160
Yeah.

s167
00:17:41.160 --> 00:17:45.760
The different treatments that you have in that clinic can be the evidence that you need.

s168
00:17:45.760 --> 00:17:47.600
Like you might get a CBCT.

s169
00:17:47.600 --> 00:17:54.840
Well, that those images and slices of those images might be the evidence that they're asking for.

s170
00:18:00.440 --> 00:18:08.200
So, ultimately, if you're delivering that treatment, you're sending that claim, that claim is like a receipt of what you did.

s171
00:18:08.200 --> 00:18:10.920
I did this, like here's the invoice.

s172
00:18:10.920 --> 00:18:11.320
Right?

s173
00:18:11.320 --> 00:18:15.000
Like you send it to the insurance company this invoice, and they would pay you back.

s174
00:18:15.000 --> 00:18:26.320
So, that is your ultimate like contract of you're saying you did this work, it's sealed um and now the ball is in insurance company's court.

s175
00:18:28.960 --> 00:18:40.560
And just a little bit about this progression of the claim from you're sending it, you're getting some acknowledgement, has like the syntax is right with that 999.

s176
00:18:40.560 --> 00:18:42.200
The status has been updated.

s177
00:18:42.200 --> 00:18:44.240
Hey, cool, we received it.

s178
00:18:44.240 --> 00:18:48.440
Um maybe you call them and you verify the status didn't come in.

s179
00:18:48.440 --> 00:18:54.240
Then you have this EOBs 835 receipt of payment.

s180
00:19:00.160 --> 00:19:09.800
And then we're getting to the end of this uh this talk here, but I think that, you know, LLMs, I'm fully AI pilled, right?

s181
00:19:09.800 --> 00:19:20.720
But we want to make sure that we introduce la- language models, small tiny models in a very skeptical conservative way.

s182
00:19:20.720 --> 00:19:26.000
So, being AI filled is great, but you should also be very AI skeptical.

s183
00:19:26.000 --> 00:19:31.880
Like these things, they make mistakes and it's not even you can't even say they make mistakes.

s184
00:19:31.880 --> 00:19:33.680
Like we make mistakes designing them.

s185
00:19:33.680 --> 00:19:35.280
We might set them up to fail.

s186
00:19:35.280 --> 00:19:41.720
So, we have to be very skeptical of them and we have to use them in a way that's also cost-effective.

s187
00:19:41.720 --> 00:19:52.360
You can't throw I mean you don't you don't want to use an overpowered over expensive model cuz then if you're going back to if you're reducing costs or not.

s188
00:19:52.360 --> 00:20:00.440
Let's say it's ends up being super super expensive to deliver one of these routine things that need to be done a thousand times a day.

s189
00:20:00.440 --> 00:20:03.160
You definitely don't want that.

s190
00:20:04.920 --> 00:20:05.440
All right.

s191
00:20:05.440 --> 00:20:07.800
Thank you very much.

s192
00:20:22.899 --> 00:20:24.899
[music]