Latest / Pediatrics Now: Cases Updates and Discussions for the Busy Pediatric Practitioner / Revolutionizing Pediatric Care with AI: Grand Rounds Talk!

Transcript

Revolutionizing Pediatric Care with AI: Grand Rounds Talk! · Pediatrics Now: Cases Updates and Discussions for the Busy Pediatric Practitioner

0:00
Music.
0:14
So the mission of matrix ai consortium is
0:17
really to think about inventing deploying
0:21
and advancing ai at scale with the primary goal of building technologies that
0:28
advance or promote human well-being and quite often when we say human well-being
0:34
health is of prime importance but But in general,
0:39
it could be any aspect of human well-being that could be boosted or accelerated
0:45
with these technologies.
0:47
And in order to accomplish this, we really take a transdisciplinary approach.
0:53
So in the consortium, we have AI researchers, ML engineers.
1:00
Music.
2:45
How humans process information, how we make predictions, how we learn in different
2:51
environments, and how we do that in the most efficient way.
2:55
But most importantly, when we are deploying any of these technologies,
2:59
we have to think about whether the solutions we are building are trustworthy.
2:59
Music.
3:04
Whether the solutions are robust, fair, unbiased,
3:10
whether the models are interpretable.
3:12
So we have a trust on trustworthy AI.
3:18
And ethics is a cross-cutting theme across all of these.
3:22
And today, the consortium has centers of excellence in each of these trusts,
3:27
supported by different agencies.
3:31
So we have about $45 million in funding portfolio with 30-plus research labs
3:37
and, as I said, five centers and students engaged from multiple grad programs.
3:43
And about 500 students from different research labs were part of the consortium.
3:51
And we have AI compute infrastructure.
3:55
I think this is a modest infrastructure if you think of the resources that some
3:59
of the computational researchers would have.
4:02
We have DGX servers, Lambda servers, and we have a very niche infrastructure
4:07
that we recently got funding from National Science Foundation.
4:11
We are the national hub for.
4:14
Music.
4:36
Okay, so that's kind of a brief overview of Matrix, and I'm happy to have conversations
4:41
offline about any interests about what Matrix is doing and how to engage and so forth.
4:47
So for the rest of the talk today, I'm going to focus on kind of where AI is
4:53
today in a very specific aspect of how algorithms are doing today and what are
4:59
some of the more recent advancements and some opportunities specifically in pediatric care.
5:06
So if you look at the trends that are out there, I think this is publication
5:12
or data points from Harvard School for Public Health,
5:16
it shows that there is a projected reduction in treatment costs when using AI for diagnosis.
5:24
You could get like a 50% reduction in treatment costs.
5:28
That is a significant and reduction, high cost efficiency.
5:33
So why then not consider this in the loop for diagnosis?
5:40
And similarly, there is a projected improvement in health outcomes up to 40%.
5:47
I'm sure there are a lot of caveats to these projections, but in general,
5:53
there seems to be a huge opportunity in achieving some savings by using these tools.
6:00
And there is an estimated market price that you would see for healthcare and
6:07
that is a significant investment that is coming in.
6:11
So it's really, I would say these are not ordinary trends and I think this is
6:18
only going to increase in the coming years.
6:21
So it's very important to think about AI not as a passive tool,
6:27
but how do we actually actively engage in the diagnosis or in the prediction
6:34
or in your workflow, how to actively use it.
6:41
So when I say passively using it as a tool,
6:45
I am sure some of you are using ChatGPT or some form of large language model
6:52
today and maybe querying it in some forms,
6:56
but is it part of your workflow where it can help you in more active ways?
7:05
I think most clinicians are not doing that at this stage.
7:09
And I'm sure for several reasons, beyond the technical flows that are not designed
7:17
to support it, it's also infrastructure support and so forth.
7:22
But it's very important to think in this day and age to.
7:28
Actively use this in the design flow. So at Matrix, we kind of look at different
7:34
projects that are focused on AI centers of excellence that are related to health.
7:39
I'll give an example of one of it in the later slides, which is the chatbot
7:45
we are building for the trauma research supported by the state where we are
7:50
collecting data, trauma data for the entire state of Texas.
7:53
And there are a few other projects where we are trying to build tools for physicians
8:00
and clinicians in adopting AI in their workflows through AimAhead and other projects.
8:08
But I just want to give you a context in terms of the hype that is there about AI.
8:15
So several of you might be familiar with Mark Roberth's videos.
8:19
If you look at this video in here,
8:22
you see that this animal is able to really adapt to the changes in its environment,
8:31
come up with creative solutions to get to the destination,
8:35
and it is continuously improving on how it is doing this.
8:41
This is a complex environment. Whereas if you think of modern day AI,
8:47
it is not able to adapt in this form.
8:51
This kind of flexibility and real-time adaptation for different tasks is not there.
8:59
So if you think of current AI models, whether they're AI agents or not,
9:07
they're really incapable of such continual learning or adapting to unforeseen
9:12
challenges in real time.
9:14
And a lot of times there are huge data sets that are needed to solve these problems.
9:20
So we did this project with DARPA. This is almost like five years or four years old now,
9:27
where we were trying to solve this problem of how do you build AI models that
9:33
learn continually in different environments.
9:36
So I hope you're able to see this video.
9:39
So we built these new types of algorithms that are able to adapt,
9:45
and this is deployed on this drone in a photorealistic simulation where it is
9:50
able to forage through this dynamic environment and try to reach to the people
9:58
who need rescue operation.
10:00
So in search and rescue operations, you can perhaps consider these kinds of
10:05
algorithms. If you look at the current ones, this kind of adaptability is not there.
10:11
And of course, this has to be done in...
10:16
In a matter of milliseconds. So which means latency is key aspect as well,
10:23
not just adaptability, but how quickly it is able to do it.
10:28
So these were some early attempts in addressing this problem.
10:33
And since then, we have advanced the algorithms quite a bit in our work.
10:37
And there is a lot going on in the community in order to think about how do
10:41
we design AI that learns continually similar to humans, a lifelong learner.
10:47
So for the next part of this talk, I'm going to show some examples in settings
10:53
that might be familiar to you in how AI is being used or can be used and potentially used.
11:05
So very early on, I think one project that we worked on was with the city of
11:12
San Antonio, where during the peak of COVID,
11:17
they wanted support on trying to see if AI could help them in understanding
11:23
the patterns of, you know,
11:28
the mobility patterns, sorry, in the city.
11:32
And based on that, if we can predict how the COVID spread is going to be impacted
11:39
within the And these were some things we have done on almost a weekly basis early on,
11:46
and then it almost became a daily report that we gave to the city of San Antonio.
11:51
These models were built by graduate students in Matrix to make those predictions.
11:56
And like, you know, if you increase the mobility by this factor,
12:00
what would be the effect if you increase by 50 percent, 75 percent and so forth?
12:07
So this was, I think, helpful for the city at that point of time to have a real-time
12:14
dashboard and have a model that could make these predictions.
12:19
So if there are situations where real-time decision is needed with different
12:26
types of data in your workflows, I think this is something Matrix can support it.
12:33
The other thing we have done more in advanced technology is like,
12:38
you know, quite often, well, especially during COVID-19, you have these different
12:43
strains that have evolved, right?
12:47
Like you have the alpha strain, the beta strain, the gamma, and so forth.
12:51
What happens is if you design an algorithm, like when we designed it,
12:57
The algorithm was looking at certain strains, but a new strain might have other
13:05
features that need to be embedded in the algorithm.
13:08
So we have started looking at the continual learning algorithms that could help
13:15
in predicting this kind of changes without forgetting what it has learned in the past.
13:20
Because a lot of algorithms, they're very good on solving the current task,
13:26
but they forget what they have learned in the past.
13:29
And this is called catastrophic forgetting or catastrophic interference problem
13:33
in machine learning community.
13:35
So we're looking at algorithms that would solve this,
13:39
and this is a great case study for it in how you can build dynamic architectures
13:44
so that when the strains or when these viruses are evolving,
13:49
you could use more modern architectures to solve that problem without losing
13:55
the information you have learned in the past, similar to how humans do it.
14:01
And if we look at other ways, if you look at this cartoon, I think there are
14:07
different sources of information from which you can.
14:13
Learn and try to get a full picture of a patient's health.
14:19
So here you see the diagnosis codes from the doctors or you could have textual
14:27
notes that come in from the clinical data or there is imaging data of different modalities.
14:35
There is cell imaging data and there is also these nodes that are written,
14:42
or we would call these unstructured data.
14:45
And you can think of other, you know, EG, ECG, you can have claims data,
14:51
you can have genomics data for this patient.
14:55
So these are what we call multimodal data, as in like, you know,
15:00
the data is in different formats.
15:02
It's coming from different sources, and you use this information and fuse it
15:09
with multiple algorithms to understand or diagnose the patient's health.
15:15
And for example, what I'm going to show in the next slides or what you might
15:20
be using, something called ChatGPT, is using one modality or two modalities
15:26
of information. I guess it's using three right now.
15:29
It's using text, voice, and image data.
15:32
So, if you have this multi-modal input, the algorithms become a lot more challenging
15:39
in order to get the right output.
15:43
The algorithms don't get challenging. It gets challenging to design the algorithms.
15:50
It's also important to understand where the data is coming from and if the data
15:54
is being fused in the right format to make the right decision.
15:57
So the sources of all these data formats becomes very important.
16:03
So we kind of did this single modality one in the trauma project that we are
16:09
working with the state, which is TRC4 funding for this.
16:13
So Dr. Khutub from Matrix is leading this project.
16:17
So you can see that. Let me play this video. I hope you're able to see it.
16:24
So a trauma assistant was built where it gives some more nuanced information
16:29
of what action should be done in an emergency situation.
16:33
And it runs in like, you know, a matter of milliseconds.
16:38
So here is an information. A 28-year-old male falls from a height of 10 feet
16:43
and presents with numbness and weakness in both legs. and he has normal sensation
16:48
in his upper limbs but is unable to move his legs and has absent patalar reflexes.
16:53
So what action should you take and does this patient require a level 1 trauma center?
16:59
So when we run that, I hope the videos...
17:05
So it is giving guidance on the actions that could be taken,
17:10
like, you know, immediate actions, whether there's immobilization,
17:13
assessment, vital signs, monitoring, establishing IV access,
17:17
and then referral through the trauma centers and the level of care that is needed.
17:24
So here in summary, ensured rapid transport to a level one trauma center is
17:30
initiated while monitoring and stabilizing the patient as per guidelines.
17:34
But then you can think and imagine further where it gives like dynamic information
17:40
on what is the trauma center that could be,
17:45
I mean, though the, I think, ground level crew would know the nearest trauma
17:50
center, I think you could have access to the usage,
17:55
the wait times, everything dynamically in a visual so that you,
18:01
and the traffic that you would have all these kinds of information in a single
18:08
flow where you could be guided to go to the right trauma center and get the care immediately. So.
18:16
This kind of application is possible where you have here a singular modality
18:22
of information. This is not multimodal, but multiple other inputs can be given to this assistant.
18:29
And that's what this was a pilot study.
18:31
And we are advancing this with now collecting information of all the data from the state of Texas.
18:38
And Amina is looking at advancing this chatbot to do a lot of these additional features.
18:46
That would support care in these situations.
18:51
The other tool I thought you should know that got released literally last week
18:57
or 10 days back, it is called MedGamma.
19:03
It is again a language model but it's particularly trained with medical data
19:11
to support diagnosis or a lot of pre-analysis and care.
19:17
So I wanted to show some of the demos for the MetChemma because I think if you
19:23
want to explore or just, you know, even play with and see what it is capable of,
19:29
this would be a good tool to explore. Sorry.
19:35
So if you have a use case, you can define the use case.
19:39
Let's say you want to generate a text report and you want to get an evaluation,
19:43
but you want to get performance at a certain level.
19:47
Like, you know, do you want, you can set the performance that you would want,
19:52
whether you want a high performance or a low performance.
19:55
And I'll tell you in a second why.
19:58
That would change the outcomes in some cases. But so you could look at the data.
20:05
You have the input in here with multiple images.
20:09
And then you're looking at certain cases and trying to see what,
20:13
based on your setting of what the performance is, you get certain response.
20:20
So the reason you're tuning that performance is that you're able to process
20:26
it faster on a lighter model.
20:28
And this is particularly important when you're thinking about embedding these
20:36
kinds of tools in your workflows because they consume a lot of energy and which means a lot of money.
20:44
It costs a lot because of the compute resources they need. So you would want to reduce that burden.
20:51
And one way is to kind of use lighter models.
20:55
And if, let's say, 99% accuracy is not needed in a certain situation,
21:02
maybe 90% gives you the confidence to go in a certain route,
21:07
I think you could change those or adjust those performance metrics in this model.
21:13
So it made a big splash in the last one week.
21:16
It's a 4B model is what they call it. So I would encourage you to explore it.
21:22
And they kind of show it how, like, you know, they demonstrate it,
21:27
like, you know, how it looks like for an intake optimization with this tool.
21:34
So there is a built-in agent that is going to ask questions to gather information about the patient.
21:42
So it kind of generates an intelligent pre-visit report that would help you
21:46
as a provider or a caregiver and give you a detailed evaluation of the pre-visit report.
21:55
So this is just a simulation of how it runs with what it would ask the end user,
22:02
so what they're visiting for and based on that,
22:08
I think what they feed in and the questions or what the queries it gives to
22:14
the patient, it kind of gives you a generic summary.
22:19
And you can see that some details that are not mentioned and all of this additional
22:27
information that would help you to diagnose or help you to provide the right care for this patient.
22:37
So obviously, I'm sure you have lots of questions on whether the output that
22:43
this tool is providing is reliable, but I think this is where collaboration with humans,
22:51
experts, is very critical.
22:55
Because based on the data, the generalized databases that it is looking from,
23:02
the outputs that it could view sometimes may have mismatches,
23:07
may have incorrect information.
23:09
And that is where the expert is able to quickly say, well, that,
23:13
for example, those triggers do not fit with this portfolio.
23:18
And you would call it out and the model is able to adapt.
23:22
So I would say this can be used as a secondary or a surrogate still because
23:29
of some of these concerns on how the data is being processed, the source of the data.
23:35
But I think compared to several other models that are out there,
23:40
this model seemed to be doing really well with medical data.
23:43
Some other examples where you can think of where AI could support is,
23:49
let's say, a lot of models that are built today or a lot of analysis that is
23:54
done is done on adult medical data,
23:59
not as much, let's say, pediatric data sets are not there for a lot of diseases.
24:06
So what I mean of course
24:09
it would help to evaluate on pediatric
24:12
data sets and get the right diagnosis but
24:16
there are also AI models that could transfer the knowledge that is there from
24:20
the adult data sets and try to make diagnosis with pediatric data so this is
24:28
again something that is still in the works and a lot of models,
24:34
for example, what we have seen just now with MedGamma, one of the models,
24:38
the large language models, there is a lot of personalization and,
24:43
But there is also opportunity to translate complex information with a language
24:48
that patients can easily process or understand.
24:53
So the tone, the terms, the jargon, all could be reduced when you're communicating with kids.
25:03
And also there are tools, including ones we are building on,
25:08
that provide real-time monitoring and decision-making,
25:13
And also there are researchers in Matrix who are looking at how do we ensure
25:17
that what we are building is secure and sensitive information is protected.
25:27
So, for example, in the case of trying to predict the variations,
25:33
you could think of AI that is processing vast amounts of data, basically.
25:39
And it's pinpointing subtle signs of a disease that might otherwise go unnoticed.
25:44
So this will help in early intervention and all of that.
25:49
But well, what would you do if you didn't have the data for the specific disease
25:55
you're looking at or if it is really scarce and difficult to collect data for
26:00
that specific situation?
26:02
So then it's really challenging for the model to make the prediction.
26:07
So there are newer techniques like transfer learning and the continual learning models.
26:12
One of the features is transfer learning that could leverage existing data sets
26:18
and adapt this knowledge to the unique situation or characteristics that you
26:24
were looking at and try to give a better diagnosis or diagnosis or so forth.
26:32
And similarly, the large language models where I said you could have an opportunity
26:38
to think about how to generate information that is more digestible to the end audience.
26:45
In this case, you could have the model, let's say,
26:51
if you look at the example on the right here, you see that there is a specific
26:58
information about this condition that was provided by the doctor.
27:05
And you could translate that in a different way to an adult and a different way to the kid.
27:11
And I think you do that on real time on a daily basis.
27:14
And you could have the models also do this, not in real time, perhaps before,
27:22
during the entry-free intake time or post-visit care.
27:30
In different situations, you could think of using this type of individualized
27:36
and age-appropriate or customized responses.
27:43
And the other example I was saying was where you could do the real-time monitoring and prediction.
27:49
There might be situations, I think, there are smart health homes is like a classic
27:54
example where we can think of using these kinds of sensors that monitor vital
28:00
signs and use that information to make a certain prediction if there is an anomaly in the signal.
28:09
And you can think of using this in different
28:12
contexts not just within a smart health home
28:15
and on a regular basis like you know you have all of us have some kind of smart
28:19
device now monitoring our critical signals but then fusing them in trying to
28:24
make the predictions i think you need really a light device that is able to
28:30
do in again a matter of less than i think 10 milliseconds.
28:35
So the neuromorphic computing technologies, we are building help in this kind
28:41
of analysis and prediction.
28:44
And the last one I want to highlight is about privacy.
28:49
So a lot of times these data sets, whether those are proprietary data sets that.
28:55
Are designed in your clinical setting or something that you're using out of a general database,
29:05
you might want to secure this data,
29:09
And I know there are certain standards that you follow with the HIPAA compliance and all of that.
29:15
But in general, instead of having a single cloud infrastructure or a server
29:22
where you store all of this data,
29:24
you would want to think about something called a federated model in how this
29:32
data is being saved and how you learn from this data sets that are stored in this cloud.
29:40
So what a federated learning infrastructure does is there is distribution of
29:46
information across multiple clouds, many clouds, fog.
29:51
There are a lot of these terms that are used in the compute community.
29:54
So you could have that distributed and you're just not learning from a single
30:00
data source, but then you're learning from these,
30:03
different data sources and making the decision, whether it is diagnosis or treatment or so forth.
30:11
And this is something that is commonly used nowadays when you're trying to chat
30:16
or messaging or in any of those tools, federated learning is naturally embedded
30:21
in your smartphones or devices.
30:24
And these could be extended to more of the medical data to ensure that the data is more secure.
30:31
So the end user only can see data that is relevant to them.
30:35
They do not have the visibility of the generic or other data.
30:41
So it is more secure in this using this federated learning approaches.
30:49
Sorry, I think this is, again, an extension of another example of MedGammer
30:54
that I wanted to show in diagnostic support.
30:57
So you can see the prompts that are given here.
31:01
Again, if you're an expert radiologist, please describe the findings for the bow chest X-ray.
31:08
And you see the radiologist impression on the left.
31:11
And you could also see on the prompt on the right, what do you think is the
31:16
most likely cause in this image?
31:18
And what would be the next steps that you would
31:21
suggest and this is like you know more a quick
31:25
response to see like well consult a dermatologist for
31:28
this additional steps so you can think
31:31
of using this in different ways in your workflows it could you know integrate
31:39
multiple images again here both 2D and 3D scans it can generate comprehensive
31:45
and reports and there is visual question answering is what you see on the right.
31:50
It's called the VQA technology where you're giving that image and asking it
31:56
to analyze it and you can query about the image.
32:02
And you can also think about using it in clinical decision support. Again, this is a caveat.
32:08
It comes with a big caveat, right? You want to really think about the data sources
32:13
and inbuilt models and the level of accuracy this model is giving.
32:18
But since it has just been recently released, I would look at some of the example
32:26
scenarios and try to work with it and see how it does without giving any proprietary data.
32:34
And similarly, I think I've already shown this, so I'm going to skip that.
32:37
So as you are thinking about adopting or integrating more actively any of these
32:45
models in your workflows, there are multiple things to think about.
32:50
Well, the model does great in terms of performance, but does it work in all
32:57
the situations? Is the model fair?
32:59
And why would this happen? Why would the model be not fair?
33:03
The reason is it might be trained only on a subset of the data.
33:08
It might be trained only in certain geographical regions.
33:14
Maybe it does not translate well across the country in different scenarios,
33:19
different geographical regions.
33:20
So it's very important to be aware what was the training data for the model
33:26
and what are its limitations before using a model.
33:30
So that it doesn't perpetuate incorrect information or it might not amplify
33:39
unfair predictions in some scenarios.
33:43
Similarly, with the output that is generated from this tool,
33:49
if the data set it is relying on is not a robust one, it can generate incorrect information.
33:57
And this is where human AI teaming and collaboration is very important to correct
34:03
the outputs of this model.
34:05
As an expert, you're able to do that.
34:08
So it's important to not fully rely on the outputs that are viewed from the models.
34:18
And I think this is probably more of a concern in human-rich environments, and this is a.
34:29
Or in mission-critical environments where if you're using AI,
34:33
you have to think about the accountability.
34:36
And I'm sure you are a lot more cognizant and trying to embed this in your workflows
34:42
more actively as to how to reduce any kind of compliance issues.
34:50
So when you're using these types of models, what if the model has predicted
34:56
something wrong and you have relied on that information and has given the treatment based on that?
35:02
So how do you ensure that this kind of impact is reduced? use.
35:10
So it's very important, again, to tie to the earlier things,
35:15
whether the model is fair, whether it is giving the right information,
35:19
whether it is trained on the right data sets.
35:21
Is the model matured enough to make accurate predictions?
35:25
So this kind of information is helpful in deciding whether to use that model
35:34
because of these kinds of issues on accountability and so forth.
35:38
And the other point to think about is whether the model is even trained on open
35:45
datasets or are there going to be some IP ownership issues on where the model
35:50
is getting its data from or where it was originally trained on.
35:54
There are models which are originally trained on and then you just deploy them in your scenario.
36:00
They don't need that additional training, but there are other models that are
36:04
continuously training and updating based on the information that you're given.
36:09
So in either of these scenarios, it's important for you to think about where
36:14
that data is coming from.
36:17
So I guess you see that underlying theme for all of these is to be knowledgeable
36:22
about the data that the model is trained on and the source of the data and whether
36:30
the reliability of the data so forth.
36:33
So if you're more interested in trying to understand some of these,
36:37
we do have sessions in tinkering with an AI scientist or we do bootcamps for
36:43
MD students in MDMS and AI program. We were doing bootcamps earlier.
36:48
So if you're interested, we would keep you in the loop when the next training session happens.
36:54
So thank you so much for your time. And I'm open to questions.
36:59
Thank you, Dr. Kuditipudi, for that wonderful presentation on transforming healthcare with AI.
37:05
Let's see if anybody has any questions. So let me start with me.
37:09
So most of us are sort of scared of AI. We have not used it.
37:13
Where do we start as a novice in AI in the medical field?
37:17
Is the MedGamma a good place to start?
37:22
It's a good place to play. What I would say is, I think no matter which profession
37:29
we are in, everybody should be an AI generalist in this day and age.
37:34
So what does that mean is to be aware of some of the tools and actually try to,
37:42
it takes very minimal effort in playing with these tools because the user interfaces
37:47
have improved significantly.
37:50
Definitely, I think large language models is something I think that is,
37:56
if not today, pretty soon going to be embedded in your workflows and thinking.
38:01
So it's a great tool for you to play with. And Medgamer is, yes,
38:06
definitely something I would play with.
38:08
And I think you can immediately at least see, oh, well, it does really well
38:13
in the situation, but it does terrible here.
38:15
But you're giving the guidance to have vector models that would help you.
38:20
But also, I think in general, trying to understand the basics of how these models work.
38:27
You don't have to understand the full intricacy of the exact models.
38:31
But I think, as I said, being an AI gendronist, trying to understand,
38:36
well, what does the pipeline look like?
38:39
As in how the data is being taken into the model, as in what kind of pre-processing
38:46
is done, how it is fed into the model.
38:49
And then once it is fed into the model, what kind of outputs the model is giving
38:54
and how reliable are they?
38:57
So I think from the AI developers itself, there are things that we should be
39:01
doing in reporting these statistics on the situations in which the model should
39:08
be used and should not be used.
39:10
But I think to answer your question,
39:15
I think it's really important for everyone to be an AI generalist.
39:18
And this is a good tool to start with. I was actually going to compile a list
39:23
of tools that would help you.
39:26
And I can share this after the talk and if anybody is interested in playing with a series of tools.
39:34
So there's a question in the comment box by Dr.
39:37
Lip said, very interesting, thank you. Will AI eventually completely replace the physician?
39:46
I think this is where the hype is, I would say, right?
39:51
I mean, yes, AI is able to do some of the tasks that are very streamlined or,
39:58
you know, where the data is streamlined, where the data is structured,
40:02
where the environment is structured.
40:05
It doesn't do very well in unstructured environment still. It doesn't learn
40:10
continually or adapt as well as humans do in different situations.
40:15
We are far from that.
40:17
There are a lot of claims that, you know, today we have an AGI is going to come,
40:22
which is artificial general intelligence, or you might hear super intelligence.
40:27
You hear all these terms that are thrown out there that it's going to happen
40:32
in the next couple of years.
40:34
And a lot of researchers who are in the field do not think it's there.
40:37
Because of several underlying fundamental problems that need to be solved with these models.
40:45
And again, I think we humans, we are 3D structures, right?
40:50
So we move in environments and we do a lot of things beyond just processing
40:55
this textual or information from the images.
40:59
So if you think of robotics or other tools that are going to support the clinicians or physicians,
41:08
they are also they also need quite a bit of evolution to even think of working
41:14
in these unstructured environments they work very well in factory settings where
41:19
everything is streamlined.
41:22
So I personally do not think that they are going to replace the physician but
41:28
there is going to be a lot of teaming up that's going to happen.
41:33
So this is why I would encourage everyone to be an AI generalist.
41:39
Thank you. This is Dr. Hedgin Beller. Do you have a question?
41:43
I see you have raised your hand. Yes.
41:46
Sorry, Dr. Kumar. This is Dr. Asanas, and I'm here with my residents on the
41:49
inpatient service. Dr. Hedgin Beller was actually the one that logged in for us.
41:54
So again, my name is Dr. Achachwan Asanas, and I'm a pediatric hematologist
41:58
oncologist, and I also work for the physician's practice as an administrator.
42:04
You mentioned the issues with With regard to HIPAA and utilizing the federated model,
42:10
there are currently AI platforms in use for physicians kind of collecting data
42:17
with regard to the interview process.
42:19
I think I can mention a bridge here.
42:22
They have mentioned that they erase the data after a period of time.
42:28
As AI generalists, what should we be looking for? really the data model of how
42:33
to make sure that we're using secure mechanisms?
42:37
Or should we feel fairly comfortable when people say that they're erasing the
42:42
information before anybody outside may be able to access it?
42:49
Very good question. So, I mean, when someone is saying they're erasing data,
42:56
well, I think perhaps, I mean, the way they designed it, perhaps it is like
43:01
a three-month time frame, six-month time frame, the data is erased.
43:04
But what happens is it really depends on the underlying architecture.
43:09
It happened with some of the, several of you might be using messaging platforms like WhatsApp.
43:16
I don't know if you're familiar with it.
43:19
There are some of these messaging platforms where the user was thinking or informed
43:24
that data is being erased.
43:27
But in actuality, because of how architectures are built behind where the models are deployed.
43:36
The data could be residing in some or some artifacts of this could be residing
43:42
in some form in one of these architectural building blocks.
43:48
So how do we know as an end user? Sometimes this is really only experts can see that.
43:56
And sometimes these things come out, you know, much later than you would like to know about.
44:02
So what, I mean, I'm not trying to scare you here, but I think it's very important
44:08
to ask questions about the sources of data.
44:12
And the architecture they're using and the confidence rate at which they think
44:17
this data is going to be, data is being erased and ask them to show you the
44:21
flow, how data is being erased.
44:24
I mean, probe on it further to kind of see what is the process.
44:30
And what I would also ask is like, consult with an AI specialist,
44:36
not just with a vendor you are working with, but consult with an AI specialist.
44:40
I mean, there are several that we could refer to from the matrix consortium,
44:46
for example, who may not be expert in the exact technology you're using,
44:50
but they could help you ask the right questions also.
44:54
So I would use that layer approach, trust, but I think be cautious.
45:01
Thank you, Dirish. And then there are comments by Dr. Jones. One HIPAA, let's see.
45:07
I can go ahead and comment, Dr. Kamad. I think, you know, interesting.
45:11
I presume you guys know about open evidence, again, out by Mayo and in collaboration
45:17
with New England Journal of Medicine,
45:20
and now actually JAMA Network as a tool.
45:23
It's free and available to clinicians. You have to have an MPI number to use.
45:28
And it's HIPAA compliant now as well. And I actually just ran your question
45:33
through on the trauma patient that you had.
45:36
And he actually gave a rather remarkable
45:39
answer, very thorough with the references that you can then tap on.
45:43
So open evidence is actually used. Now, again, you talk about a place to start.
45:48
I warn the residents routinely when they use these tools, prompt engineering
45:52
is still very, very important, and particularly in pediatrics.
45:57
If you miss a key element, and age is one of them, to put into these tools,
46:02
you will get very convincing wrong answers.
46:05
And so that's still one of the biggest concerns I have as we begin to introduce
46:09
this to our residents is if you don't have that knowledge and that background to see or to,
46:16
when you get that wrong answer to at least turn your head and say,
46:19
man, that doesn't sound quite right.
46:21
You're going to miss it because it, again, it's so confidently wrong that you will make errors.
46:27
And so prompt engineering, I think is, is such a key element to begin with when
46:32
you train people how to use these, these tools. Right.
46:36
Right. There are bots that are being generated to help with the prompt engineering as well.
46:42
Hopefully those are made more available for different platforms pretty soon
46:48
because I completely agree with you.
46:51
Missing keywords can give you completely different answers or inaccurate information
46:57
that shouldn't be relied on.
47:00
And the other thing about references, sorry, to just add to that,
47:04
earlier tools, like at least one year or six months back, it would give you
47:09
references. So you would think that, well, this information is reliable.
47:13
A lot of times those references are not existing publications.
47:19
It would make up references with high confidence.
47:22
So it's very important to check the references that are being pointed to.
47:28
I agree. I ran into it as an editor, actually. I always check the references
47:33
because those are not right.
47:35
Anyway, there's a question by Dr. Rizvi. How good is the news of AI hospital in China?
47:42
How good is the news of AI hospital in China?
47:48
I don't know if I read this specific news article or information,
47:56
But I know there is a lot more active adoption of tools at all levels in some countries.
48:07
So I think there is perhaps some significant amount of training data that you
48:13
would have these models being trained on.
48:15
And over time, they would become stronger and more accurate.
48:20
But I think there is a lot more active adoption of tools.
48:22
Yeah, I think there are several aspects within a hospital that you could use
48:27
smart technologies and which I could see happening in the U.S. as well.
48:33
But it wouldn't be a complete AI hospital, as I said. I think the human element
48:38
is still quite significant in here.
48:41
Thank you. Any other questions, comments for Dr. Kuritipati?
48:45
I have one other. Again, I think one of our biggest challenges is I've been meeting with Dr.
48:51
Sankri and also our CIO here or vice president for information technology here
48:58
at UTL San Antonio is a lack of a ready access to an enterprise or private level LLM.
49:04
And we've been trying to use collators, as some people will label them,
49:09
like u.com and other ways to improve efficiency.
49:12
But we're still limited, even though we have a very powerful list of 23-plus LLMs that we can use.
49:20
They're not enterprise. They're not private. And so it really limits what we
49:24
can load up into our sources or what people call a RAG or interpretation or
49:30
what we can ask as far as questions and data we can share.
49:33
I mean, do you see things at your level coming out of the UT system or in our partnership with UTSA,
49:40
which will are merging with UTSA, which will move us towards having enterprise
49:45
or private level LLMs for doing some of our work?
49:49
Because it is limiting our office in the office of DME with moving forward.
49:54
And I'm watching Stanford and other people I talk to all the time.
49:57
I tell them what we're doing. They take and run it, and they're already ahead
50:01
of us within months because we just seem to be limited on the resources side
50:06
of this equation with AI.
50:08
Yeah, I think there are two very good points to think about here.
50:13
Regarding the private instance of LLM, I'm actually having a meeting with the
50:18
CIO next week, with your CIO.
50:20
So we are trying to think about that structure and facilitating tools for physicians
50:27
and researchers in UT Health.
50:29
So I'm sure this is of prime importance, and the VPR's office is already thinking about it.
50:36
So let's just build this team to think about how do we help researchers.
50:42
The other part is think about the data that you have or the assets that you
50:49
have within your groups.
50:53
I personally believe in open sourcing, you know, whatever models we are building and everything.
51:00
But there are valuable assets that you have within your units,
51:05
whether it is data that is collected from the patients or even in the intake forms.
51:11
There's several things, right, that are really valuable assets.
51:14
So if you can think of what are the digital assets that you would have within
51:19
your units, and I would not give it away for free. to any entity.
51:24
I would really think about it.
51:27
I know Stanford, I think probably you're talking about their research center on health.
51:34
They're being very smart and pretty aggressive in collecting data from multiple
51:40
states and trying to deploy their models.
51:43
I know we could do better and we could move faster.
51:47
But independent of which group is using
51:50
these assets i would think for you
51:54
as a clinician in your or if you're
51:57
a chair of a department what are the assets that you
52:00
have within your unit it may
52:03
not always be a cost associated with you
52:06
giving those assets but there could be other negotiations that
52:09
could happen we're not at least
52:12
in academic settings from the engineering side we are
52:15
not generally trained to think that way but i think it's
52:17
very important to add value i mean
52:20
get value out of those assets that you would have thank you one last question
52:25
a question can we get an ai tool to automate billing so billing for the insurance
52:33
companies or the patients can this be automated i think there are multiple tools.
52:39
I don't know how it fits within your workflow, but I think this aspect of the
52:46
problem is a lot more, I wouldn't say easier,
52:49
but I think it is being adopted in multiple industries already.
52:54
So there are tools to do it.
52:56
OCR technology has matured so much, mature recognition technology,
53:00
even if it is all like, you know, legacy paperwork that you have,
53:03
it could all be screened and fed into models to automate the process.
53:07
So I think there are tools to look at.
53:14
Thanks for listening to that informative Grand Rounds talk.
53:17
I'm Holly Wehmint, and thanks for listening to Pediatrics Now, our way.
53:22
Music.