Beyond the algorithm: trust, evidence and equity in AI-enabled cancer care

WEBVTT
00:00:00.359 --> 00:00:05.194
Hello everyone, and welcome to our weekly podcast brought to you by Vijay Oncology.
00:00:05.670 --> 00:00:09.908
This week, we will be covering the latest developments in artificial intelligence
00:00:10.100 --> 00:00:14.949
in the field of oncology so far in 2026, and what it means for clinicians who are.
00:00:37.047 --> 00:00:41.467
And colleagues, which really asked the question of what information are patients
00:00:41.467 --> 00:00:45.219
getting online when they ask about AI and cancer care?
00:00:45.451 --> 00:00:48.977
So not about using AI to, you know, give them information about how to treat
00:00:48.977 --> 00:00:52.437
their cancer, but what, if they want to know what can AI do in cancer care,
00:00:52.437 --> 00:00:53.659
what information are they getting?
00:00:54.071 --> 00:00:59.142
And they looked at both Google and YouTube, so Google web pages, YouTube videos.
00:00:59.375 --> 00:01:02.805
This was not chatbot generated content, but really indexed web content.
00:01:03.241 --> 00:01:10.665
And they found a really small number of resources, both on Google and YouTube, that address this.
00:01:11.124 --> 00:01:15.107
And most of the resources were at about a 12th grade reading level,
00:01:15.107 --> 00:01:18.165
which is higher than the average reading level in America.
00:01:18.699 --> 00:01:22.449
And most of them did not have good quality information.
00:01:22.745 --> 00:01:30.333
And notably, less than 20% addressed the safety risks of AI with hallucination and misinformation.
00:01:30.803 --> 00:01:34.727
And that's really significant because we need our patients to understand what
00:01:34.727 --> 00:01:37.199
are the limitations of AI.
00:01:37.847 --> 00:01:42.997
And this really speaks to the need for broader patient guidelines and content
00:01:42.997 --> 00:01:47.420
about what AI can do, but also what AI can't do and the potential dangers of AI.
00:02:05.206 --> 00:02:09.286
Cancer conversation. We have clinicians, we have patients, we have industry
00:02:09.286 --> 00:02:11.886
of all these people that are part of the AI conversation.
00:02:12.304 --> 00:02:18.266
But I think the concern that's really driven by this study is increasingly patients
00:02:18.266 --> 00:02:20.482
are using AI for health information.
00:02:20.743 --> 00:02:24.462
Patients are turning to online resources for information.
00:02:24.700 --> 00:02:29.516
And so it's important that they understand that when we find that information
00:02:29.516 --> 00:02:33.494
online, that there's risks that AI could hallucinate, that it could provide misinformation.
00:02:33.819 --> 00:02:38.994
And so without that education piece, you know, what happens if AI causes harm
00:02:39.313 --> 00:02:40.950
and who is responsible for that?
00:02:41.333 --> 00:02:47.156
And how do we handle potential conflict or discordance between what the clinician
00:02:47.156 --> 00:02:50.356
is recommending and what AI is recommending, right?
00:02:50.356 --> 00:02:54.546
And I think it comes down to there's still the arch of medicine where we take
00:02:54.546 --> 00:02:56.576
guidelines and we apply them to each patient.
00:02:56.576 --> 00:03:02.876
But I think it's this broader question of how does the content that our patients
00:03:02.876 --> 00:03:06.424
are consuming online impact outcomes?
00:03:06.987 --> 00:03:13.126
And we have to remember that what comes from generative AI is the prompt that
00:03:13.126 --> 00:03:14.795
is fed into generative AI.
00:03:15.108 --> 00:03:18.580
And so how you ask AI impacts what you get.
00:03:18.911 --> 00:03:24.512
And so if we don't educate the community on this, what you're inputting may
00:03:24.640 --> 00:03:27.784
impact the and then ultimately impact cancer outcomes.
00:03:28.499 --> 00:03:32.530
Next up, Gil Shamay from the Technian Israel Institute of Technology,
00:03:32.530 --> 00:03:37.820
PIFAR, explores the clinical utility of discordances between conventional biolicoid
00:03:38.140 --> 00:03:42.660
assays and deep learning-based histology predictions across six countertops.
00:03:42.820 --> 00:03:48.740
It's well known today that AI models can predict molecular information from,
00:03:49.230 --> 00:03:54.880
images, microscopic images of the tumor, meaning from H&E slide images.
00:03:55.588 --> 00:04:00.330
Many studies have already been done on this topic. In fact, we already started
00:04:00.330 --> 00:04:02.813
developing such models 10 years ago.
00:04:03.295 --> 00:04:07.342
And two months ago, one of our studies was published in Lancet Oncology,
00:04:07.557 --> 00:04:12.550
where we developed an AI model to predict the oncotype score from HND images
00:04:12.550 --> 00:04:15.231
and validated it on the Taylor X randomized trial.
00:04:15.873 --> 00:04:21.488
Now, although these studies showed better and better performance over the years,
00:04:21.859 --> 00:04:27.792
this performance reached a limit, meaning that there are always cases where the AI model is wrong,
00:04:28.460 --> 00:04:33.173
or that it doesn't match the molecular assay it intends to predict.
00:04:33.877 --> 00:04:38.990
For this reason, these models are often viewed today as low-cost,
00:04:38.990 --> 00:04:44.601
efficient, but inferior surrogates of the gold-standard molecular assays,
00:04:44.902 --> 00:04:47.369
which highly limits them clinically.
00:04:48.117 --> 00:04:55.255
What we show in our new study here at ASCO is that the mistakes of the AAI are
00:04:55.255 --> 00:04:57.202
not actually random errors.
00:04:57.718 --> 00:05:04.163
We show that AI learns new information that it was not directly trained to predict.
00:05:04.894 --> 00:05:09.005
And what happens is that the model doesn't really see the molecular information
00:05:09.005 --> 00:05:12.402
in the image. It doesn't see the receptors or the gene expression.
00:05:13.285 --> 00:05:18.886
Instead it sees the downstream effect of this molecular SAS reflected in the image.
00:05:19.286 --> 00:05:23.635
For example, a tumor can be ER positive by immunostochemistry,
00:05:23.635 --> 00:05:29.955
meaning that it has estrogen receptors, but still irresponsive to estrogen because
00:05:29.955 --> 00:05:32.320
its ER pathway is inactivated.
00:05:32.836 --> 00:05:37.439
For example, a tumor can be ER positive by immunostochemistry,
00:05:37.659 --> 00:05:44.095
which means it has estrogen receptors, but still irresponsive to estrogen because
00:05:44.095 --> 00:05:46.455
its ER pathway is inactivated.
00:05:46.797 --> 00:05:50.704
This means that the tumor may act like an ER-negative tumor,
00:05:51.406 --> 00:05:54.445
which means that it may look like an ER-negative tumor. And this is what the
00:05:54.445 --> 00:05:57.585
AI eventually sees in the HND image. So the AI predicts it as.
00:06:00.325 --> 00:06:03.485
ER-negative, meaning that the AI doesn't see the receptors,
00:06:04.328 --> 00:06:13.663
it sees if the tumor is responsive to estrogen, We also showed that these discordances between the AI,
00:06:14.305 --> 00:06:20.303
predictions from H&E images and the molecular essays are biologically meaningful
00:06:20.576 --> 00:06:25.211
and can be even useful for better guiding clinical decisions.
00:06:26.082 --> 00:06:29.345
Following on, Mafia Curran from Medicine Elizabeth Healthcare,
00:06:29.345 --> 00:06:34.325
Edgewood, Kentucky, discusses the impact of AI oncology, highlighting its potential
00:06:34.325 --> 00:06:37.123
to affect patients, trainings, and global accesses.
00:06:38.343 --> 00:06:43.399
AI is probably one of the fastest growing platforms in oncology and the proper
00:06:43.399 --> 00:06:48.063
use of it is something that will need to be discussed and implemented in the future.
00:06:48.493 --> 00:06:53.839
One of the things that, you know, my talk touches on is to understand exactly
00:06:53.839 --> 00:06:59.429
how as an oncologist, this will affect trainees, the next generation global
00:06:59.429 --> 00:07:02.149
practice in the world setting as well.
00:07:02.149 --> 00:07:06.019
And then also looking at how it affects patients most importantly.
00:07:06.019 --> 00:07:10.689
So I think if you look at how artificial intelligence can affect all three of these things.
00:07:10.689 --> 00:07:15.559
I think patients are probably the one that I think it has the largest impact on as well.
00:07:15.559 --> 00:07:20.959
And that can encompass many different things that it can range from tools that
00:07:20.959 --> 00:07:22.223
help with supportive care.
00:07:22.780 --> 00:07:28.369
Other things that touch on artificial intelligence scribes that help improve
00:07:28.369 --> 00:07:33.479
time with patients from a physician standpoint, and also improving time spent
00:07:33.479 --> 00:07:34.947
with patients as a result of that.
00:07:35.336 --> 00:07:40.929
Other things such as early screening, drug discovery, many potential applications
00:07:40.929 --> 00:07:43.310
that all affect patients over time.
00:07:43.635 --> 00:07:48.633
In terms of trainees, I think that artificial intelligence and the use of it
00:07:48.767 --> 00:07:50.642
in oncology specifically is at a
00:07:51.019 --> 00:07:55.919
critical point in time where we need to understand how artificial intelligence
00:07:55.919 --> 00:07:59.129
tools are being developed, what data is really going into that,
00:07:59.419 --> 00:08:02.200
and then how our trainees are utilizing that,
00:08:02.670 --> 00:08:07.836
in their day-to-day life and as they become future oncologists in the future.
00:08:08.181 --> 00:08:12.302
When we look at the global workforce, how artificial intelligence can affect.
00:08:13.306 --> 00:08:19.013
How care delivery is administered within low to middle income countries,
00:08:19.431 --> 00:08:22.874
you look at improving screening specifically as well, too.
00:08:22.874 --> 00:08:28.194
And there are many models and some tools that are now available that will help
00:08:28.194 --> 00:08:30.874
improve screening in areas that have,
00:08:31.654 --> 00:08:35.574
you know, limited resources overall, too, that will help improve detection,
00:08:35.574 --> 00:08:40.814
which will then improve the rates, you know, of cancer overall Dr.
00:08:40.814 --> 00:08:45.814
Kurian also comments on how AI can reduce disparities in care and the current
00:08:45.814 --> 00:08:48.286
state of FDA approved machine learning devices.
00:08:49.278 --> 00:08:52.127
AI has the potential to be able to
00:08:52.507 --> 00:08:58.367
improve disparities by early detection, and the use of those tools will then
00:08:58.367 --> 00:09:01.683
translate into improved outcomes and improved screening rates,
00:09:01.902 --> 00:09:05.584
which will then lead to reduced cancer rates for those low-to-middle-income countries.
00:09:05.745 --> 00:09:11.365
The challenge, I think, with this is that it requires infrastructure to be able to
00:09:11.708 --> 00:09:16.047
make those things happen, and it will also require a lot of upfront capital
00:09:16.047 --> 00:09:21.277
cost to be able to implement those things successfully in a long-term measure as well too.
00:09:21.643 --> 00:09:24.881
So I think those are probably the main barriers at this point in time,
00:09:25.142 --> 00:09:29.747
but I think ASCO and many other organizations are committed to being able to
00:09:30.167 --> 00:09:35.247
help invest those resources along with many other companies to help level the
00:09:35.247 --> 00:09:38.587
playing field for those experiencing cancer throughout the world.
00:09:38.587 --> 00:09:42.497
The things that I think are growing and will become more popular over time will
00:09:42.497 --> 00:09:45.913
be patient monitoring tools using artificial intelligence.
00:09:46.325 --> 00:09:50.377
And then also tools like open evidence, I think will become very,
00:09:50.377 --> 00:09:56.367
very, you know, crucial in terms of how our trainees, how our current workforce
00:09:56.367 --> 00:10:00.367
utilizes those things to help administer better care to patients.
00:10:00.367 --> 00:10:05.567
The way that I phrase it to individuals of that open evidence and other artificial
00:10:05.967 --> 00:10:10.340
platforms can elevate the level of care in areas that have.
00:10:10.943 --> 00:10:17.947
Struggled to have access to good clinical trials, good specialists as well too, that it can
00:10:18.347 --> 00:10:22.447
level the playing field and improve the care that's delivered within community
00:10:22.447 --> 00:10:26.750
practices to the level of those within academic practices as well.
00:10:27.301 --> 00:10:31.857
Next up, we have Jay Patel from the Atrium Health Wake Forest Baptist Comprehensive
00:10:31.857 --> 00:10:33.948
Cancer Center in North Carolina.
00:10:34.279 --> 00:10:38.897
It presents expert validation of the novel AI enabled system using large language
00:10:38.897 --> 00:10:42.847
model-based decision rules and electronic health records to match patients with
00:10:42.847 --> 00:10:44.322
cancer to clinical trials.
00:10:45.204 --> 00:10:49.151
During this co-development process, really the first phase of this is to validate
00:10:49.488 --> 00:10:53.168
that platform using real-world data from our own patient records,
00:10:53.488 --> 00:10:55.833
as well as real cancer trial protocols.
00:10:56.245 --> 00:10:59.952
And this has really been one of the gaps in prior literature that have mostly
00:10:59.952 --> 00:11:01.789
focused on synthetic data sets.
00:11:02.369 --> 00:11:06.602
So we took an opportunity to take thousands of patients within our own health
00:11:06.602 --> 00:11:11.722
system, develop the model, train the model, use machine learning plus large
00:11:11.722 --> 00:11:14.943
language models, kind of a hybrid architecture.
00:11:15.204 --> 00:11:19.976
To further train that model, and then test it in a separate independent test set of patients.
00:11:20.452 --> 00:11:23.842
So the test set of patients focused on individuals with breast cancer,
00:11:23.842 --> 00:11:25.549
lung cancer, and pancreatic cancer.
00:11:25.955 --> 00:11:30.164
So we were matching these patients to retrospective clinical trial protocols.
00:11:30.795 --> 00:11:35.950
In that study, we identified that there was actually 94% agreement or concordance
00:11:36.292 --> 00:11:41.000
between a clinical research coordinator reviewing the data versus the AI platform,
00:11:41.301 --> 00:11:44.420
so performing really just as well as the human review.
00:11:44.907 --> 00:11:49.119
So at different probability thresholds, we then determined that the platform
00:11:49.119 --> 00:11:54.389
can have upwards of 99% sensitivity and positive predictive performance really
00:11:54.389 --> 00:11:56.365
upwards of, again, 90% or so, meaning that.
00:11:56.870 --> 00:12:02.639
Every nine out of every 10 patients or so that the AI platform says the individual
00:12:02.639 --> 00:12:07.464
is eligible for the patient, they truly end up being eligible for that particular clinical trial.
00:12:07.714 --> 00:12:12.319
The biggest reasons that we're going through this process to advance responsible
00:12:12.319 --> 00:12:16.849
AI use in healthcare as it relates to clinical research, primarily due to the
00:12:16.849 --> 00:12:21.609
massive burden that this puts on research coordinators as well as our providers.
00:12:21.609 --> 00:12:25.879
So number one, our physicians that are seeing several patients throughout the
00:12:25.879 --> 00:12:30.489
day have limited time and capacity to remember every single clinical trial and
00:12:30.489 --> 00:12:32.221
the eligibility criteria for that trial.
00:12:32.483 --> 00:12:36.279
So this really prompts them in real time to identify patients who are matching
00:12:36.279 --> 00:12:38.502
to those trials during that visit.
00:12:38.961 --> 00:12:42.909
They can then send that information over to a research coordinator who can review
00:12:42.909 --> 00:12:48.559
the evidence from the AI platform that points them directly to the areas within
00:12:48.559 --> 00:12:51.603
the chart that that eligibility criteria came from.
00:12:51.980 --> 00:12:56.059
So with that, we've now seen with our preliminary data, this can save the research
00:12:56.059 --> 00:13:00.850
coordinators time by at least doubling the amount of patients they could prescreen.
00:13:01.118 --> 00:13:02.702
And that's really a conservative estimate.
00:13:03.438 --> 00:13:08.451
Atrium Health, Advocate Health, we tend to put on about 14,000 cancer patients
00:13:08.451 --> 00:13:11.205
on trials every year as of last year, those are our data.
00:13:11.641 --> 00:13:15.751
And so we expect we can probably double or triple that volume by leveraging
00:13:15.751 --> 00:13:19.167
kind of human in the loop AI processes to match patients to trials.
00:13:19.707 --> 00:13:25.111
Finally, Marie Hanaoka from the Institute of Science Tokyo, Japan explores how
00:13:25.111 --> 00:13:29.921
AI-derived surgical quality metrics can be incorporated into robotic surgery
00:13:29.921 --> 00:13:33.598
training, curricular and competency assessment frameworks.
00:13:34.829 --> 00:13:42.718
First, by reviewing video in which enhance the dissection plane or atomic nerves,
00:13:42.995 --> 00:13:48.452
the surgeon can understand why and where they entered the wrong place.
00:13:48.876 --> 00:13:57.424
So, I think it is more effectively by reviewing AI-enhanced video than compared with the original one.
00:13:57.906 --> 00:14:06.485
And I think it also makes it easier to compare the dissection with expert surgeons,
00:14:07.025 --> 00:14:13.017
and they can recognize how different the exposure of the correct plane can be.
00:14:13.371 --> 00:14:20.169
This is also my own impression from reviewing many, many of AI-enhanced surgical videos.
00:14:20.877 --> 00:14:24.856
Second, real-time AI analysis is now available.
00:14:25.744 --> 00:14:31.343
Although the clinical value still needed to be evaluated.
00:14:33.312 --> 00:14:39.392
In the future, AI could allow the surgeon when the dissection field is too close
00:14:39.392 --> 00:14:41.037
to the autonomic nerves.
00:14:41.420 --> 00:14:48.102
Such a system could encourage the surgeon to improve exposure and maintain a
00:14:48.102 --> 00:14:51.067
safer, more appropriate dissection plane.
00:14:51.659 --> 00:14:55.583
So, regarding the competency assessment,
00:14:55.920 --> 00:15:01.116
Japan currently has a technical certification system for surgeons.
00:15:01.592 --> 00:15:05.202
At present, evaluation is very.
00:15:06.748 --> 00:15:13.290
Subjectively assessed by reviewers. If AI could objectively evaluate the surgical
00:15:13.290 --> 00:15:19.924
mobilization and dissection quality, it could add more objectively to this system.
00:15:20.440 --> 00:15:27.386
In addition, AI may be useful in initial screening tool to technical assessment,
00:15:27.804 --> 00:15:32.748
probably to reduce the burden of human reviewers.
00:15:33.532 --> 00:15:37.410
She also discusses concerns around data ownership, patient's consent,
00:15:37.410 --> 00:15:42.152
and algorithmic transparency when considering integration of AI into the research.
00:15:43.052 --> 00:15:49.294
These issues are essential for the reasonable implementation for AI in surgery.
00:15:50.136 --> 00:15:56.486
First, patients should provide informed consent regarding the use of surgical video and data.
00:15:56.910 --> 00:16:04.677
Second, data must be securely stored and appropriately anonymized.
00:16:05.246 --> 00:16:11.240
Regarding algorithm transparency, I believe it is essential that AI systems
00:16:11.240 --> 00:16:13.880
are validated using data from
00:16:13.880 --> 00:16:19.525
multiple institutions before they are widely adapted in clinical practice.
00:16:19.919 --> 00:16:25.347
In addition, the safety of real-term AI assistance and its impact on surgeons',
00:16:25.800 --> 00:16:29.706
work should also be carefully evaluated.
00:16:30.258 --> 00:16:36.980
We are currently investigating this issue in an ongoing prospective clinical trial.
00:16:37.688 --> 00:16:43.535
AI should always support, not replace, surgeon's clinical decision-making, I think.
00:16:44.853 --> 00:16:47.269
Thank you to our speakers and to you for listening.