#APSA50: Artificial Intelligence

Published: 2019-06-14

20 Views

0 Likes

0 Shares

0 Comments

StayCurrentMD

View profile →

Inside this episode

Kai, the Library's AI content creator, listened to this episode and mapped who's speaking, the chapters, key claims, and cases. Every item links to the exact moment in the recording.

AI-enriched

Who's speaking

Alexander Gibbons — host
Alexander Gibbons — host
Michael Muley — guest

Chapters

0:00Introduction and Background — Introduction to Dr. Michael Muley, radiologist and Google Cloud product manager, discussing AI in healthcare at the 50th APSA meeting.
1:20Current AI Applications in Healthcare — Discussion of existing AI deployments in radiology, including head CT bleed prioritization, and why surgical specialties are less affected by current AI applications.
4:34Future Applications in Surgery and ICU — Exploration of potential AI applications in surgical specialties, including laparoscopic surgery, trauma triaging, and ICU sepsis alerts.
7:41Limitations of AI and Data Requirements — Discussion of deep learning limitations, particularly the need for large, high-quality datasets and challenges in healthcare data collection.
10:47Cloud Infrastructure and Ethical Considerations — Examination of cloud computing's role in enabling data sharing, ethical concerns around model bias, and AI's promise for providing consistent care.

Key claims

1:39Radiology has been thought of as one of the first fields where artificial intelligence will make an impact in healthcare — Michael Muley
1:46At the Radiological Society of North America annual meeting, there are about 100 to 150 startups presenting in the machine learning showcase — Michael Muley
2:24Yale has deployed AI for prioritization of head bleeds on head CTs, cutting down the time until a radiologist looks at studies with bleeds — Michael Muley
2:45Most AI applications in radiology are currently in the operational realm, augmenting workflow rather than in the diagnostic sphere — Michael Muley
3:54Evidence-based medicine has been discussed since the 70s, but most surgical training practices lack strong evidence behind them — Michael Muley
4:43Laparoscopic surgery provides a natural opportunity for AI because it captures data digitally with video — Michael Muley
4:50Surgical robots are not currently using AI tools primarily for regulatory reasons — Michael Muley
5:28Trauma triaging is currently done based on very simple rules and is not very data driven — Michael Muley
6:29Penn State rolled out a simple sepsis alert based on four criteria (elevated or low white count, temperature, heart rate) that did not work well because it went off too frequently — Michael Muley
7:06Emory has sepsis alerts deployed into clinical workflow using deep learning — Michael Muley
7:56Deep learning neural networks are very large and learn just by example, requiring many examples with good variety of the data space — Michael Muley
8:11The type of acquisition device (GE versus Siemens) may affect what an AI algorithm learns in radiology — Michael Muley
8:31There are approximately 7 to 8 billion people in the world, most without access to healthcare, and very few with medical conditions, making healthcare data availability very small — Michael Muley
9:11Tesla has 100,000 to 200,000 cars collecting data daily for training self-driving capability — Michael Muley
9:26Waymo has collected about 10 million miles of driving experience for self-driving car development — Michael Muley
10:26AI models learn whatever associations are present in training data, including spurious correlations like rulers in melanoma images — Michael Muley
11:33Cloud computing enables centralized data storage which makes it easier to create frameworks for data sharing between institutions for research — Michael Muley
12:26A model trained on US data may not work anywhere else due to population differences — Michael Muley
12:36If training data has racial or ethnic bias, the AI model will have that bias as well — Michael Muley
13:00Training a machine learning model is a very small part of what needs to be done for clinical deployment — Michael Muley
13:12How the FDA will decide what constitutes a safe medical device for AI applications remains an unsolved question — Michael Muley
13:28AI models are much more consistent than humans even if not better, which is their great promise — Michael Muley
13:38There is huge variance between how different radiologists read studies, sometimes incorrectly and sometimes due to difference of opinion — Michael Muley
13:53Similar variance exists among ophthalmologists in their clinical interpretations — Michael Muley

Open questions

How will the FDA decide what constitutes a safe medical device for AI applications in healthcare?
What are the actual outcome data from Emory's deployed sepsis alert system?
How can large, high-quality datasets be collected in pediatric surgery given the small patient populations and workflow constraints?

This episode was analyzed and enriched by Kai, the Library's AI content creator. Every item links to the moment it comes from — click a timestamp to listen in context.

Written for:

AI in Medicine: From Workflow Optimization to Evidence-Based Consistency

The episode's teaching points arranged as a structured lesson, building from the basics up to the finer points. Written by Kai from the episode transcript and reviewed before publishing.

For trainees · Teaching arc · AI-written, human-reviewed

Foundation: Data Availability Defines AI Feasibility

The fundamental constraint on medical AI is not algorithmic sophistication but data scarcity 8:31. While engineers working on consumer applications expect billions of data points daily, healthcare operates in a radically different regime: approximately 7 to 8 billion people exist, most without healthcare access, and among those with access, serious pathology remains uncommon 8:31. For comparison, Tesla deploys 100,000 to 200,000 cars collecting training data daily 9:11, and Waymo has accumulated 10 million miles of driving experience 9:26 — data volumes impossible to replicate in clinical medicine. This scarcity intensifies as you narrow the patient population: pediatric surgical conditions represent an even smaller dataset than adult pathology, making robust AI development correspondingly harder.

Deep Learning's Fundamental Limitation: Pure Pattern Matching

Deep learning neural networks learn exclusively by example and require extensive variety across the data space they will encounter 7:56. "Those models are very, very dumb, right? They learn whatever you show them" [q4]. This creates a critical vulnerability: the algorithms will learn any association present in training data, including spurious correlations 10:26. The classic example is melanoma detection models that learned to associate rulers with malignancy because training images of concerning lesions often included measurement tools. In radiology, models may learn that certain CT markers indicate which hospital acquired the image, then use hospital-specific disease prevalence rather than image features for diagnosis 10:26. Even acquisition device manufacturer — GE versus Siemens — can affect what the algorithm learns 8:11. The implication: you cannot assume a model has learned clinically meaningful features without rigorous validation.

Current Deployment: Operational Rather Than Diagnostic

Most AI applications in radiology currently augment workflow rather than replace diagnostic judgment 2:45. Yale's deployment of head CT prioritization exemplifies this approach: the system identifies studies with intracranial hemorrhage and moves them to the front of the reading queue, reducing time to radiologist review 2:24. This represents operational optimization, not diagnostic replacement. Similarly, ICU sepsis alerts have evolved from simple rule-based systems — which failed because they triggered too frequently 6:29 — to deep learning approaches now deployed at institutions like Emory 7:06. The pattern holds across specialties: AI handles triage, prioritization, and workflow management while clinicians retain diagnostic authority.

The Bias Problem: Models Inherit Data Inequities

A model trained on US patient data may fail when applied elsewhere due to population differences 12:26. More insidiously, if training data contains racial or ethnic bias, the resulting model will perpetuate that bias 12:36. Disease prevalence varies across populations, and if your training set underrepresents certain groups, the model's performance will degrade for those patients. This is not a problem you can engineer around after training — it must be addressed in dataset construction. The regulatory framework for ensuring AI safety in this context remains unsolved 13:12.

The Real Value Proposition: Consistency Over Superiority

AI models are much more consistent than humans even when not demonstrably better 13:28. There is huge variance in how different radiologists interpret studies, sometimes due to error and sometimes due to legitimate difference of opinion 13:38. Similar variance exists among ophthalmologists 13:53. This variance extends across all clinical decision-making. The promise of AI is not superhuman diagnostic accuracy but rather the elimination of practice variation and more reliable adherence to evidence-based standards. Training a machine learning model represents a small fraction of the work required for clinical deployment 13:00, but once validated and implemented, the model will apply its learned associations consistently across every case — something human clinicians, even with the best intentions, cannot achieve.

The discussants emphasized that evidence-based medicine has been discussed for decades, yet most surgical training practices lack strong supporting evidence 3:54. AI offers a path toward truly evidence-based practice by enabling systematic analysis of large datasets and consistent application of derived insights — not by replacing clinical judgment, but by reducing the variance that currently separates what we know from what we do.

Takeaways from this story

Deep learning models learn any association in training data, including spurious correlations like rulers indicating melanoma
Most deployed AI in radiology optimizes workflow (prioritizing head bleeds) rather than replacing diagnostic interpretation
Healthcare data scarcity is fundamental: 7-8 billion people, most without access, few with serious pathology
AI's primary value is consistency, not superiority—eliminating the huge variance seen among radiologists and other specialists
Models trained on biased data perpetuate that bias; a US-trained model may fail in other populations

Artificial Intelligence in Medicine: Workflow Optimization Now, Clinical Augmentation Later

The episode's main topic retold as a plain-language walkthrough — what it is, why it matters, and what the speakers concluded. Written by Kai from the episode transcript and reviewed before publishing.

For the care team · Explainer · AI-written, human-reviewed

Why AI in Medicine Exists as a Distinct Discipline

Artificial intelligence entered clinical medicine not because algorithms outperform physicians, but because healthcare generates vast amounts of digital data that humans cannot process efficiently. Radiology led adoption because imaging data is natively digital and standardized 1:39. At the annual Radiological Society of North America meeting, 100 to 150 AI startups now present products, some already deployed in clinical systems 1:46. The discipline exists to solve a throughput problem: too much data, too few eyes, and unacceptable variance in how clinicians interpret the same information.

The Core Clinical Problem

Clinical decision-making suffers from two fundamental weaknesses. First, humans cannot process the volume of data modern medicine generates—radiologists drown in imaging studies, intensivists miss early sepsis signals buried in vital sign trends, surgeons triage trauma patients using crude rule sets that lack strong data support 5:28. Second, clinician performance varies wildly. Radiologists show substantial variance in how they read studies, with disagreements stemming from both error and legitimate differences of opinion 13:38. Ophthalmologists show similar inconsistency 13:53. Most surgical training practices lack strong evidence behind them 3:54. AI promises not necessarily superior judgment, but consistent application of whatever standard the model learns.

How the Approach Works

Current AI deployment in healthcare focuses almost entirely on operational optimization rather than diagnostic replacement 2:45. Yale uses AI to prioritize head CTs with intracranial hemorrhage, cutting the time until a radiologist reviews critical studies 2:24. Emory has deployed deep learning sepsis alerts in ICU workflow 7:06—a marked improvement over earlier rule-based systems that triggered so frequently they became clinically useless 6:29. Trauma triaging remains an area where current approaches lack strong data-driven foundations 5:28, representing an obvious target for AI augmentation.

The underlying technology is deep learning: very large neural networks that learn just by example 7:56. These models require massive datasets with good variety across the problem space 7:56. In radiology, this means training on images from different scanner manufacturers, because acquisition device characteristics can affect what an AI algorithm learns 8:11. The models are, in the words of one radiologist working on these systems, very limited in their reasoning—they learn whatever associations exist in the training data, whether clinically meaningful or spurious 10:26. A melanoma classifier trained on images that happened to include rulers learned to associate rulers with malignancy. CT interpretation models can learn to recognize which hospital generated a scan based on external markers, then use local disease prevalence as a diagnostic shortcut.

Surgical specialties face steeper barriers to AI adoption than radiology, but laparoscopic surgery offers a natural entry point because it captures data digitally through video 4:43. Surgical robots do not currently incorporate AI tools, primarily for regulatory reasons 4:50. When that regulatory hurdle clears, AI-augmented robotic platforms will likely emerge quickly.

Where Practice Remains Contested

The central tension is data availability. Healthcare generates far less usable training data than other domains pursuing AI deployment. Global healthcare access remains limited, and medical conditions affect only a fraction of the population, making healthcare data availability relatively small 8:31. Compare this to autonomous vehicle development: Tesla has a large fleet collecting data daily 9:11, and Waymo has accumulated substantial driving experience 9:26. Pediatric surgery, with its small patient populations and rare conditions, faces particularly severe data scarcity.

Model bias remains an unsolved ethical problem. AI systems trained on US populations may not work elsewhere 12:26. If training data contains racial or ethnic bias, the model inherits that bias 12:36. Disease prevalence varies across populations, and models must account for this or risk systematic errors in underrepresented groups.

Regulatory frameworks lag behind technical capability. How the FDA will determine safety standards for AI medical devices remains an open question 13:12. Training a functional model represents only a small fraction of the work required for clinical deployment 13:00—the harder problems are validation, integration into clinical workflow, and ongoing performance monitoring.

When to Involve AI-Augmented Systems

For referring clinicians, the practical question is when AI-augmented care pathways exist and should influence referral decisions. Currently, this applies primarily in radiology (imaging prioritization) and critical care (sepsis alerts at select institutions). The discussion did not address specific referral criteria, but the operational focus of current deployments suggests AI tools function transparently within existing workflows rather than creating new referral triggers. A head CT ordered for clinical indications gets AI-assisted prioritization automatically; the ordering physician need not do anything differently.

The value proposition is not that AI replaces clinical judgment—"I'm not working at Google because I think my radiology job is going to go away, I think my radiology job is here to stay" [q1]—but that AI models offer greater consistency than human interpretation, even when not demonstrably superior in accuracy 13:28. Reducing practice variation and ensuring consistent application of evidence-based standards may prove more valuable than marginal improvements in diagnostic accuracy.

Cloud infrastructure enables the data sharing necessary for model development 11:33, though centralized storage does not automatically grant access—"just because the data is in the cloud doesn't mean that anybody has access to it outside of the rights holder" [q5]. The technical infrastructure now exists; the organizational and regulatory frameworks to use it safely remain under construction.

Takeaways from this story

Current AI in healthcare optimizes workflow (prioritizing studies, triaging patients) rather than replacing diagnostic judgment.
Deep learning models learn spurious correlations from training data; a melanoma classifier associated rulers with malignancy.
Healthcare data scarcity limits AI development—billions lack access to care, and rare conditions yield tiny datasets.
AI's primary value is consistency, not superiority: reducing the variance in how clinicians interpret identical cases.
Models trained on biased data inherit that bias; US-trained algorithms may fail in other populations with different disease prevalence.

AI in Medicine: From Workflow Optimization to Evidence-Based Consistency

The episode's teaching points arranged as a structured lesson, building from the basics up to the finer points. Written by Kai from the episode transcript and reviewed before publishing.

For specialists · Teaching arc · AI-written, human-reviewed

The Operational Deployment Reality

AI is augmenting workflow, not replacing diagnosis. Current clinical deployments focus almost entirely on operational improvements rather than diagnostic decision-making 2:45. Yale's head CT prioritization system exemplifies this: when the algorithm detects intracranial hemorrhage, it moves those studies to the front of the reading queue, cutting the time until a radiologist reviews them 2:24. Similarly, Emory has deployed deep learning-based sepsis alerts into ICU workflow 7:06 — a marked improvement over Penn State's failed rule-based system that triggered on four simple criteria (white count, temperature, heart rate) and generated so many false alarms it was shut down 6:29. The lesson: AI's current value lies in triaging, prioritizing, and streamlining, not in autonomous clinical judgment.

The Data Scarcity Problem

Healthcare data is orders of magnitude smaller than engineers expect. Deep learning neural networks learn purely by example and require massive datasets with good coverage of the problem space 7:56. Tesla has a large fleet of cars collecting driving data daily 9:11; Waymo has logged 10 million miles 9:26. Healthcare offers nothing comparable. Of the 7 to 8 billion people on earth, most lack healthcare access, and among those with access, few have any given condition 8:31. As you narrow to subspecialty populations — pediatric surgical patients, for instance — the available training data shrinks further. This scarcity is not just about volume: acquisition device matters (GE versus Siemens scanners may produce systematically different images that affect what the model learns 8:11), and you cannot simulate clinical encounters the way you can simulate driving scenarios.

The Spurious Correlation Trap

Models learn whatever associations exist in the training data, including the wrong ones. Neural networks are "very dumb" in this sense: they are excellent at detecting patterns but have no understanding of causality 10:26. The classic example is melanoma detection models that learned to associate rulers in photographs with malignancy, because dermatologists photograph concerning lesions with rulers for scale. In radiology, models can learn to recognize CT markers that indicate which hospital acquired the image, then use hospital-specific disease prevalence as a diagnostic shortcut rather than analyzing the anatomy. If your training set has racial, ethnic, or geographic bias, the model inherits that bias 12:36. A model trained on US data may fail entirely when deployed elsewhere 12:26.

The Consistency Proposition

AI's primary value is not superior performance but elimination of human variance. There is enormous variability in how radiologists interpret the same study — sometimes due to error, sometimes due to legitimate difference of opinion 13:38. The same holds for ophthalmologists 13:53 and, by extension, every specialty where humans make judgment calls. Most surgical training practices lack strong supporting evidence 3:54, yet trainees adopt whatever approach their attending uses and carry it forward. AI systems, once validated, deliver the same output every time for the same input 13:28. This consistency creates the opportunity to move toward genuinely evidence-based practice: if the model is trained on high-quality data reflecting best available evidence, it can reduce unwarranted practice variation and ensure that care aligns with what we actually know rather than what individual clinicians happen to believe.

The Deployment Gap

Training the model is a small fraction of the work required for clinical use. Building a performant neural network is the easy part 13:00. The hard parts are regulatory (the FDA has not yet established clear standards for what constitutes a safe AI medical device 13:12), organizational (creating data-sharing frameworks between institutions 11:33), technical (ensuring the model generalizes across acquisition devices, patient populations, and clinical contexts), and workflow-related (integrating the tool so it helps rather than disrupts care). Surgical robots do not currently incorporate AI tools, primarily because of regulatory constraints 4:50. Laparoscopic surgery offers a natural opportunity because it generates digital video data 4:43, but translating that opportunity into deployed systems requires solving all these non-technical problems first.

The discussants emphasized that AI will not replace physicians but will enable more consistent, evidence-driven care — a shift from the current state where most clinical decisions rest on weak evidence and inherited practice patterns rather than rigorous data.

Takeaways from this story

Current AI deployments prioritize workflow optimization over diagnosis—Yale's head CT triage cuts radiologist response time for bleeds
Healthcare data scarcity is severe: 7-8 billion people, most without access, versus Tesla's large fleet collecting daily driving data
Neural networks learn spurious correlations—melanoma models detected rulers, not lesions; CT models can learn hospital markers as shortcuts
AI's core value is consistency, not superiority—radiologists show huge interpretive variance even when incorrect or opinion-based
Model training is the easy part; regulatory approval, data sharing frameworks, and workflow integration remain largely unsolved

Topic overview

Interview with Dr. Michael Muley, a radiologist and product manager at Google Cloud's Healthcare and Life Sciences team, discussing artificial intelligence applications in medicine. The discussion covers current AI deployments in radiology (such as head CT bleed prioritization at Yale), limitations of deep learning methods (particularly the need for large, high-quality datasets), and the challenge of applying AI to surgical specialties where digital data capture is less established. Dr. Muley argues that AI's primary value lies in providing consistent, evidence-based care rather than replacing physicians, with near-term applications focused on workflow optimization and triaging rather than direct diagnostic replacement.

Key takeaways

AI in radiology currently augments workflow (e.g., prioritizing head CT bleeds) rather than replacing diagnostic interpretation. (2:24)
Deep learning requires large, diverse datasets; device variability (GE vs Siemens) can affect algorithm performance in radiology. (7:56)
AI models replicate training data biases; US-trained models may fail in other populations due to demographic differences. (12:26)
AI's clinical value lies in consistency rather than superiority; radiologists show significant inter-reader variability. (13:28)

Keywords

head bleeds sepsis melanoma head CT chest X-ray laparoscopic surgery Google Cloud Healthcare and Life Sciences Stanford University Yale Radiological Society of North America Penn State Emory Akron Children's Hospital

Hashtags

#ArtificialIntelligence #MachineLearning #PediatricSurgery #HealthcareAI

Transcript

Click "Show Transcript" to view the full text (15098 characters)

Welcome to the latest episode of Stay Current in Pediatric Surgery. Today's podcast was created and edited by Todd Ponsky, Alex Cassar, Ray Hanke, and me, Alexander Gibbons, in collaboration with the Behind the Knife team of Wu Do and Megan Akasha. This episode continues our coverage of the 50th ABSA meeting by interviewing the first speaker of the TED Talks series, Doctor Michael Muley of Google Cloud's Healthcare and Life Sciences team. He sat down with us to talk about artificial intelligence in medicine. Worried about being replaced by robot overlords. No need. Find out why in this episode. I'm Alexander Gibbons here at the 50th annual APSA meeting with Dr. Michael Muley, one of the guest speakers for this series of TED Talks. Dr. Muley's a product manager at Google's cloud's healthcare and Life Sciences team and is a radiology faculty member at Stanford University. He spent time researching machine learning and his co-founded companies dedicated to bringing radiology to the cloud. He's talking with us today about his personal views on artificial intelligence and how it pertains to healthcare. Doctor Muli, thank you for joining us. Hi, Alexandra, how are you? Doing well. So for our listeners who may not yet have seen how AI affects their day to day treatment of patients, what are some examples of how AI has already changed healthcare? Well, I think, uh, surgical specialties are, you know, somewhat naturally right now less affected by artificial intelligence applications, right? I think as a radiologist, right, and that is part of the reason why I actually went into radiology, um, there's a lot more data available and it's naturally digital. So radiology really has been thought of as one of the first fields where artificial intelligence will make an impact in healthcare, and it already is, right? So there's At our annual meeting, the Archaeological Society of North America annual meeting, there's about 100 to 150 startups presenting. In what they call the machine learning showcase, right, so all products, some of which you can buy right now, some of which are not yet available, somewhat promised, but you can actually see an impact on the field today with tools that are available from vendors and are deployed at certain healthcare systems. Do you have any examples that you know or have seen in action? So one thing that I know it's deployed at a couple of hospitals. Yale, for example, is prioritization for head bleeds on head CTs, right? So if there is a bleed in a head CT, those studies get prioritized over the rest, right? So basically cutting down the time until a radiologist looks at it. So most of them are really right now in the operational realm, right, really augmenting the workflow, not so much in the diagnostic sphere. Yeah, and I think that Sort of addresses a concern that some people have. There's the worry that, you know, robot overlords are going to be taking over and replacing doctors. What I'd like to point out is I'm not working at Google because I think my radiology job is going to go away, I think my radiology job is here to stay. I would love for the chest X-rays to go away. Even that I'm not sure about, really. It's. Yeah, how do you alleviate those concerns? I think it's education about what the capabilities of these methods are. Over the past 5 or so years, there certainly has been a significant hype, very strong worry in the radiology community a couple of years ago that radiologists. Going to get replaced, right? But if you actually learn about how these algorithms work and what the limitations are right now, that's much less of a worry, you know, but it's really, I think there is tremendous opportunity with artificial intelligence applications in healthcare to get to a, you know, truly evidence-based medical system. We've been talking about evidence-based medicine since the 70s, I think, right? And if you look at it, right, whatever you're taught during your surgical training is what you're going to do. Uh, is there evidence behind it? Most of the time not. There is, it's typically very, very weak, right? And so this opens up an avenue to collect these data sets, right, and gain new insights from those data sets that we previously were not able to do. And then you kind of alluded to how in the surgical specialties we're not really seeing much of this yet. Is there anything on the horizon that you can see or envision potentially being able to be applied to either surgery in general or specifically pediatric surgery? Yeah, I mean, I think pediatric surgery is even, you know, it's an even harder one compared to like general surgery or you know, some of the Let's say laparoscopic surgeries. Any time you do laparoscopic surgery, I think there is a natural opportunity? You capture data digitally with the video. There is obviously some surgical robots out there that are right now, primarily for regulatory reasons, not using any of these tools, once that hurdle is taken. will start to have artificial intelligence tools as part of that suite, right? I think there's a lot of areas where surgery will see this impact their work, but not directly in taking care of patients, right? So I think workflows, right, and prioritizing patients is something that's not just, you know, useful in radiology, but it's really useful in any specialty, especially surgery, right, where you, you, you know, Not really clear on how you prioritize patients, right? Uh, so let's say triaging, right, uh, in trauma, kind of done based on very, very simple rules and not very data driven, right? That's one good example of where, uh, I think there will be some impact in the. Next couple of years, right, if you talk about, you know, the actual operating room, you know, outside of the ancillary things, operational things, I think that that's gonna take, uh, a lot longer, if at all. Yeah, and I just, uh, listening to you talk right there, I can definitely see like AI being able to somehow. Incorporate all the slew of data that comes at us with uh the vital signs during a trauma or that's constantly monitored in the ICU and being able to kind of make sense of that huge amount of data that's coming our way and trying to, like you said, kind of triage and bring to our attention who we need to more specifically treat or treat sooner. And you know, actually ICU is a good example, right? Uh, so sepsis alert is something that people have been. Yeah, trying to do since the 70s. Uh, I remember when I was a surgical intern at Penn State, uh, we rolled out a simple sepsis alert based on four criteria of, uh, you know, elevated white count or low white count, temperature, heart rate. And we actually rolled it out into deployment at Penn State. Uh, it didn't work very well, right, because it was such a simple, dumb rule set. It just went off all the time, right? And obviously we had a process in place, right? The nurse pages, the intern. I kept getting many, many, many pages about that, right, of documents like, oh, it just went out for nothing. So we shut that down that experiment pretty quickly there, right? And today, you know, with the new types of artificial intelligence, particularly with deep learning, those alerts can be much, much more sophisticated, right? And so some places actually do have those deployed now, Emory being An example I'm aware of where they do have sepsis actually sepsis alerts actually deployed into the workflow nowadays, and you know at all how that's impacted their management of ICU patients or just that it is in place right now and they're still collecting that data? Yeah, I think they're still collecting data on actual outcomes. What are some of the limitations of AI at this time? And how do you think those limitations could potentially be overcome? Well, a big one is you need a lot of data, right? So this approach that we use today, deep learning, right, it's essentially very large neural networks. And they are really, really dumb, right? So they learn just by example. So you need to show it many, many examples of, you know, good variety of the data space you want to cover. Right. Uh, so, you know, different types of patients, different, uh, in radiology acquisition types, right? So whether you use a GE or a Siemens device actually may affect, uh, what the algorithm learns. So the limitation right now is really access to these very large data sets, especially in healthcare. So at Google, I like to tell our engineers who are used to getting billions of data points a day from search, and then they come and work on healthcare things. It's basically we have, you know, 7 to 8 billion people in the world right now, right? Most of them don't even have access to healthcare. You know, very few of them thankfully have anything wrong with them, right? The availability of data is very, very small. And then, Uh, you know, the more specific the disease gets, right, or your your patient population, right, let's take pediatric surgery, for example, the harder it gets, right? Collecting those big, big data sets with high quality, uh, labels in the end is really the big challenge right now. Yeah, because the AI, it's only going to be as good as the quality of the data that you're putting in. Right, so, um, self-driving cars are a good example. So Tesla, right, they have the opportunity to Collect all this data while those cars drive around. There's 100 200,000 cars driving around every day now collecting data for training the self-driving capability that they have promised. Similarly, VO, they have cars driving around, have collected about 10 million miles of driving experience now. It's very difficult to generate anything like that in healthcare, right? You're actually taking care of people. So first of all, it has to fit in the workflow when you collect the data. Then it's not like you can just simulate those things. How do you get large data sets is not just the availability of the data at all, right, but then there's technical reasons why you can't collect large data sets. There's organizational and regulatory reasons. So very, very challenging to actually create a data set to train these models. Even if they could perform all the tasks that a physician does. And then you mentioned like even type of machine being used can impact the data. I know another example is for image recognition for melanoma. They had a picture of a ruler for a lot of these, and so the AI ended up recognizing that if there's a ruler in the picture, then it's more likely to be malignant. Yeah, I mean, it's very, you know, those models are very, very dumb, right? They learn whatever you show them. They're really good at that, right? They are fantastic at learning those associations from images. If you give them enough data, right, but they learn whatever association is there that may be useful to learn that. That is a good example. On CTs, a good example we have is some uh marker that exists outside of the patient, right, that may tip the algorithm, uh, the model of that you're at one or another hospital where there's a different disease prevalence, things like that. And then with your work in bringing radiology to the cloud, you know, in a perfect world where we didn't have a lot of these regulatory issues, that would be a great way to generate these large data sets if everybody was able to upload it to the same kind of destination. Well, that is why I work on cloud right now. It's really, you know, I I think that this is for many, many reasons the way that data centers in general will have and healthcare specifically, which then opens up the opportunity of giving us access to larger data sets. Once the data is centralized, it's much easier to create the framework. That allows data sharing between institutions for research purposes. Just because the data is in the cloud doesn't mean that anybody has access to it outside of the rights holder. So you still have to solve these other issues. But once the data is in a format and in a place where you can easily solve the technical challenge, then it becomes much easier to solve the regulatory organizational challenge. I'm kind of in a similar vein. Um, what are some of the ethical concerns that you can potentially foresee as AI starts to play a larger role in medicine, and what standards do you think we can adopt now to kind of make sure that we are Preventing those issues from potentially getting larger. So I mean in healthcare, you're right, the ethical issues really resolve around bias of models. So a model that is trained on data in the US may not work anywhere else. Uh, it is very dependent on your patient population, right? So if you have any racial or ethnic bias in your data, the model is going to have that bias as well. And as you know, A lot of diseases have certain prevalences that differ across populations, so you have to take all that into account, really think about what the data is that you're using before you actually use it. I think training of a machine learning model is actually a very small part of what needs to be done. And then, I do think there's some questions about how do we regulate these things in healthcare. How is the FDA going to decide what is safe, what is a safe medical device, what is not? Those are currently still unsolved questions. I'm very excited to see these models actually make a real world impact again, I think. They are, you know, even if they're not better than humans, they're much, much more consistent. I think that's the great promise is consistency, or radiologists are a good example. Uh, an example where we have data, right, where there's there's huge variance between how different radiologists read studies, right. Uh, sometimes incorrectly, right? Uh, sometimes it's just a difference of opinion. Similarly, we have data for that with ophthalmologists, right, where there's also huge variance, and that really applies to, you know, anywhere where humans make decisions, right? There's there's always going to be variants, including pediatric surgery. So getting rid of that variance and actually providing consistent care according to what we know, I think is the great opportunity for these artificial intelligence applications. Well, thank you again very much for your time, Dr. Meli. I appreciate it and uh looking forward to hearing your talk at APSA. Thank you so much, Alan. This is Alexander Gibbons from Akron Children's Hospital, the contributing editor for this episode of Stay Current in Pediatric Surgery. We hope you enjoyed this interview with Doctor Muly about applications of the AI in medicine. His talk was one of the many great topics at this year's ABSA meeting. Please join us in the community section of the Stay Curtain Pediatric Surgery app to let us know what you think about our ABSA episodes and what you would like to hear from us in the future. The next episode will feature further coverage of the 50th ABSA meeting as we continue our collaboration with Behind the Knife. Thanks again for listening, and until then, remember, knowledge should be free.

Comments

Loading comments...