Four views of artificial intelligence in pathology and laboratory medicine. That’s what panelists provided for attendees at the 7th Clinical Lab 2.0 workshop in late February in Chicago.
The University of Michigan’s Ulysses G. J. Balis, MD, spoke of AI’s use in laboratory operations and diagnostics. Tom Neufelder of Beckman Coulter spotlighted its use in instruments and postanalytically. Gaurav Sharma, MD, of Henry Ford Health is the skeptic, and the University of Pittsburgh’s Michael Becich, MD, PhD, is the rabid enthusiast who is working to unite pathology’s reports and calls himself a data plumber.
Clinical Lab 2.0 is a Project Santa Fe Foundation initiative established to provide thought leadership and help develop the evidence base for the valuation of clinical lab services in the next era of health care.
The panelists’ thoughts on the impact of AI follow.
May 2024—Ulysses G. J. Balis, MD, Inaugural A. James French, MD, professor of pathology informatics and director, Division of Informatics, Michigan Medicine, University of Michigan: I will talk about the opportunities that large language models offer us but have been largely untapped.
Pathologists and laboratorians spend a lot of time now looking for needles in haystacks. That’s a high-stress, time-consuming event, which is screening. If you can convert that screening to directed review, you will have elevated your level of practice to much more value-added work, where you’re letting AI or computational pipelines do some of that difficult work for you.
There is also classification and prediction. At one time, it was considered appropriate just to render a diagnosis, and we still need to do that. But in tandem with that, if we can elevate our practice to say a tumor or a lesion has the biologic potential, in one or two years, to metastasize with extremely high likelihood, or at the same time to say a patient is extremely likely or unlikely to respond to certain biologic agents, we can save the patient and the health system money by not selecting an expensive biologic agent that very likely may ultimately prove to be non-efficacious.
Evolution of disease processes ties into the same supporting concept: being able to predict the biologic potential of disease and get in front of it at the earliest possible time. In the Kidney Precision Medicine Project (kpmp.org) are examples of our using predictive analytics that identify up front those patients who will respond appropriately in acute kidney injury and resolve versus those who will go on to develop chronic kidney disease. Opportunities like that allow the lab to play a much more substantive and fundamental role in managing disease.
Finally, quantitation of disease prediction: A lot of what we do now in terms of time series data is qualitative or semiquantitative. AI is going to transform that into a much more quantitative approach, where we can say the patient in their current state is progressing compared with what they were a few months or years ago, based on numerical as well as language-based analytics.
Tools like this will be coming to a department like yours in the not-too-distant future, where pathologists will be informed not just by the histology but by an online computational tool with which the genes will have been pre-discovered by carrying out the needed molecular analysis. And then, by surrogate imputation to the underlying morphology, the genes can be colocalized to areas of histology where it is now possible to immediately generate a high-dimensional data set that can be visualized, with the projection showing where the gene products are expressed in a particular biopsy. This by itself, though, is raw data.
The large language models, when applied to this primary data, can assist a pathologist and laboratorian in making a narrative text, where you could have a computational pipeline that says, “Given these genes in this particular localization,” where the computer has segmented the genes for you, “please write a narrative report based on contemporary literature that supports the molecular diagnosis.” This is not a pipe dream—the fundamental tools to do this type of value-added segmentation, based on the tissue’s molecular underpinnings, are now in place.
The question is: How do we generate this into a narrative report that doesn’t take a laboratorian or a pathologist an hour or two to generate each report? It needs to be doable at scale. The inverse is possible as well, where you can start with the region of interest. For example, in diabetic nephropathy, a pathologist or histologist can identify areas of interest and ask which specific genes are involved in the disease process. And this inverse workflow allows an individual to generate a list of identified genes from the selected regions.
With large language models, the pathologist would spend time reviewing the report instead of constructing it. I mention this as food for thought of what is coming down the pike, and we will have to manage this in terms of validating it, making sure it hits all the quality metrics for safety and effectiveness, and then being able to identify the best possible ways to deploy it.
My second topic is in the financial realm: pull-through revenue. Pathology currently operates financially as a silo. This applies to the clinical laboratory itself, where much of what is generated, in terms of diagnostics, generates a definitive diagnosis that often can lead to a patient coming to your enterprise—whether it’s an affiliate or your direct hospital, if you’re a hospital-affiliated pathology department, but it could be a regional center as well—for definitive therapy.
You can show with AI, if you reverse federate the electronic health record in terms of the progression of diagnosis to clinical notes, that the patient came because of that consultation. That drives significant revenue to the enterprise at large, which is demonstrably tied to the original pathology or laboratory consultative efforts. Yet, at the enterprise level, that revenue is isolated in silos to the various departments that provide care. Surgical departments, oncology departments, radiation oncology if it’s palliative—all of those collectively generate revenue that is not recognized as contributory from the original pathology efforts.
It’s important to recognize that at an enterprise level, even though it may not be on the bottom line, pathology contributes incrementally to the margin of a large health system. Yet we don’t recognize that. If you were to take a step back, remove the dotted lines between the silos, and ask, at a systems level, if pathology contributes to the margin, the answer would be yes. Unfortunately, that value at this point is virtual. Large language models can be employed in a unique way, if you have reverse federation, for example, from the EHR, to ask: Given the original diagnosis and the clinical progression of notes, and given the following billing data, what can be said about pathology’s contribution to the overall margin? That becomes your evidence for making a stronger argument for why pathology should be valued differently than it is now. That’s yet another example of large language models potentially helping the margin value proposition of the laboratory. We’re leaving it now entirely on the table.
Tom Neufelder, senior vice president chief technology officer, Beckman Coulter Diagnostics: I would like to talk about how we see AI from the standpoint of instruments and operations in the laboratory and postanalytically in the clinical space.
The majority of AI machine learning applications are going to be embedded in instruments or the basis of a diagnostic application. When we think about instruments, there’s no question that AI, machine learning, and deep learning can be implemented to improve performance. It can increase serviceability and improve troubleshooting or provide a more natural way for an operator to interact with a machine through a large language chat function. You’ll see all of this over time.
When you think about application in the instrument, one important opportunity is in error reduction. If we think about leveraging AI to enhance computer vision, there are several facilities we can implement for basic functions such as barcode reading or tube identification and matching with the test that’s been ordered. All are easily done with machine learning in combination with computer vision.
At the same time, embedded within the instrument, we’ll see more AI applications that monitor the operation of the device itself, automatically detecting when there is an anomalous condition and then notifying the operator, or even through an Internet of Things type of connectivity back to service personnel signaling that intervention is required. From an operator standpoint, when we think about a large language model, things like troubleshooting can be challenging, and we don’t want to start with “reboot the machine” or “call your service representative.” A more natural way to interact with the system through large language models can make troubleshooting much more productive and effective and reduce the number of service calls required. We can also employ techniques like predictive analytics. When we monitor what is happening within the device, measures and metrics that we have in the system, we can see when the system is on a trajectory that will require service. Instead of relying on scheduled preventive maintenance, we can optimize that maintenance schedule so that devices are serviced when they need it, rather than try to manage the install base on a regular frequency, to minimize unplanned downtime.
There’s also no question that AI has facility in managing operations in the laboratory itself. With AI we have the ability in real time to monitor analytics more broadly than can be done manually. You’ll be able to quickly identify emerging situations in the lab, but also more subtle trends that are moving in a negative direction over a longer time frame, as well as situations that would’ve been difficult to analyze in the past and wouldn’t necessarily have been recognized until long after they had occurred. We’ll rely heavily on AI applications for monitoring such things as increasing turnaround time or increasing or unequal reagent consumption, either within a lab, across time shifts, or across multiple labs. You’ll be able to interact with the system to identify these developments without reverting to tracking through spreadsheets.
On the clinical side, per the FDA’s database on AI applications, as of the latest available data, which was end of July 2023, there were 692 AI-enabled clinical applications, and that was up from 521 at the end of 2022. The vast majority of those, between 75 and 80 percent, are in radiology, and they have traditionally operated on data from a single modality like x-ray, CT, or MRI image sets to perform functions such as perfusion quantification, bleed detection, tumor tracking, and lesion detection. What you’ll see going forward, and where the laboratory plays a much bigger part, are more multimodality AI applications. Clinical information from a variety of sources, and over time, will be combined to increase application performance.
One of the things we’re always concerned about, and the FDA is very concerned about, is generalization. When generalization is fully realized it eliminates bias typically due to overfitting on a small amount of data. The challenge is that to generalize an application, you need data from many sources and a relatively large amount of it. Interestingly, the applications we talk about are often trained on about 10,000 or so data points and then reach a plateau in the performance of the application. Contrast that with the large language models that are using millions of records for training. This is where it’s incumbent on the developer to obtain data from diverse institutions to ensure it can be deployed anywhere. We have to make sure we’re not going to see an anomaly within a particular location due to differences in practice or patient population. In health equity terms, gender and race as elements of bias are very important. If the application is properly trained with appropriately selected and curated data, we can ensure those biases are not present. However, many other more subtle aspects of bias can be introduced in an application. It’s important to remember that AI models have no inherent understanding of causality. AI models are based on correlation, and that correlation sometimes may be purely coincidental. In the engineering of the application, in the data science, we must monitor performance and ensure we are not introducing bias in an unintended way.
One of the key responsibilities incumbent upon the vendors of AI/ML applications is to demonstrate economic value, whether operational or clinical. Even in high-acuity situations in which a clinical application can more quickly identify an emergent issue with a patient, because of the cost to the institution to employ these devices, it’s essential to demonstrate an economic value. On the operational side, that’s usually relatively straightforward. You can assess impact to measures such as turnaround time or overall throughput. Alternatively, are you decreasing the use of consumables through intelligent applications? On the clinical side, it can be more challenging; we have to look at indirect or secondary effects, such as the opportunity to see more patients or perform more procedures by freeing up capacity. That will be critical in driving adoption of these applications. Over time, I believe you’ll see vendors offer risk-sharing–type agreements to deploy the application with a commitment to achieving performance metrics.
Gaurav Sharma, MD, system vice chair of clinical pathology, division head of regional laboratories, and medical director of the outreach laboratory, Henry Ford Health: Human story is all about a journey of ideas. Nothing more, nothing less. Throughout history, we have seen ideas duke it out with each other. There are good and bad ideas; often the bad ideas far outnumber the good. Here is where you have to wear the right hat, and the hat I choose is that of a skeptic.
People say that’s a negative word. It’s not. The negative word is cynic. A cynic is a person who has a negative mindset, who says, “This can’t be done.” A skeptic is a person who has an inquiring mindset: “Yes, I see this could be done, but I need to understand it.” I am a skeptic.
Much of what is said in the media is that AI will change the world for the positive. That’s the optimist crowd. The other view is doom and gloom: Kids will no longer be able to think. Who will write essays when ChatGPT can write it for them? It is the death of reasoning. Those are the pessimists.
I’m a realist: How can I use whatever is available to me, without too positive or too negative an expectation? I base that on history. In the 1990s the World Wide Web was supposed to bring us all on the same page. Information will flow easily and then truth and honesty will prevail. Nothing happened. We now have more misinformation than information. Y2K was supposed to end the world. We’re still here. And the Internet and social media were supposed to be great equalizers. They are anything but that. That’s why it’s important to have that dose of skepticism.
I come from Detroit, and in reading through its history, I came across an interesting story. When automobiles came onto the roads, people were upset. Carriage drivers went on strike because they were worried about their jobs and the automobiles frightening the horses. Town leaders said the automobiles were going to damage their rural roads. For a long time, automobiles were reserved for the rich. Vermont had a law that said if you want to drive an automobile, you must hire a person to walk in front with a red flag telling everyone that you are driving an automobile.
Now, 100 years later, we are seeing the same thing with artificial intelligence: It will take away jobs, and it is so dangerous that you have to flag everything you do with artificial intelligence. And it will make our kids dumb, and we will lose the ability to reason. This story has been repeated again and again. I promise you, if you don’t think AI is going to change your life, wait two years. AI will be as big a part of your life as a cell phone is now.
Why do we think in extremes? The answer is slippery slope fallacy. New things scare us. We think A will lead to B, B will lead to C, C will lead to D, and D will lead to E. That is a slippery slope, and it works both ways, positive and negative. We fall prey to our own imagination. Donald Rumsfeld said there are known knowns and there are known unknown dangers in the world, and there are unknown unknown dangers—“things we do not know we don’t know.”
But let me improve it. There are known known opportunities in the world, which were just described—how we can use AI to improve our operations, to do quality control checks. And how we can find the critical values that were missed. There are known unknowns: We want to do them but don’t know how. But for most inventions, the benefits are the unknown unknowns. They are not even known.
So if you asked a Detroiter who was picketing against automobiles in 1910, “What do you think of the interstate system and how will it change the American economy?” that person would have had no idea what you were talking about. That’s okay because it is an unknown unknown for them. The same is true for us: Artificial intelligence will change the world in a positive way, but we just don’t know it yet.
Let me give an example that I have observed with digital pathology. When its leaders came forward and advocated for whole slide imaging, there were cynics, skeptics, and enthusiasts. Whole slide imaging would be the foundation for digital pathology, and digital pathology would be the foundation for AI. But we cannot jump directly from the glass microscope to applying AI, so all change is gradual and builds on the change before it. If you have not even started the journey on ground level one, you can’t go to ground level four. The skeptic in me would say, Do I even have floors one and two? First let me build that.
As I see it, AI will be our third brain. The first brain is the one you are born with, the second brain is the one you carry in your pocket, and the third brain is the AI that will run in that cell phone. It will make our life easy by taking away what I call cognitive load. Rather than you having to remind yourself of things, AI will plan many predictable things for you. So our life will be easy. That’s the main benefit that will come out of AI.
But there is another lesson from digital pathology. The fight is not between a human pathologist and a computer pathologist. The struggle will be between a pathologist and a pathologist with the computer or a pathologist with AI. The last thing we want to do is think of AI as a mortal enemy, as a replacement. But we want to be sure we understand there are people out there who will use AI better than we do, so we might as well start exploring it and seeing if it works.
AI is artificial intelligence; NI is natural ignorance. AI will help you solve the data part, but AI does not help you solve the question part. It gets you the right answer, but it is premised on your asking the right question. Most of the time we don’t even know what our customer wants. Unless I have been a patient, I don’t know what patients want. AI will not help you jump that gap. You have to sit in their seat, you have to walk in their shoes, to understand the gaps or what expectations they have. And then use AI as a tool, not as a solution. You yourself are the solution. AI is only a tool to solve a problem.
Michael J. Becich, MD, PhD, chair and distinguished university professor, Department of Biomedical Informatics; professor of pathology, information/computing, and clinical/translational sciences; associate vice chancellor for informatics in the health sciences; University of Pittsburgh School of Medicine: Generative AI has been around for 30 years. What’s changed to bring it into the spotlight? Generative AI has changed based on a bunch of companies aggregating all the publicly available data around the world. They do this without giving you appropriate references, and by not disclaiming that their utility is largely gained from applying AI tools to what we’ve contributed on the public data/publication sources. I am a rabid enthusiast of AI in general—not so much for generative AI because of what these companies have done. Generative AI has created useful tools but without the proper attribution to those who generated this valuable “open” data.
They will not credit the massive contributions from things like PubMed, open-source publications, doctoral theses, and they’re sucking all that knowledge and claiming it as theirs to share back to the public, and they will not credit you or your innovative contributions. How many people have used ChatGPT to get a reference of a publication? When you do, what’s the word that comes to mind? “Hallucination” by the large language models these companies utilize. It’s kind of interesting that hallucinations are a focus of new science inquiries. Hallucinations “science” is because the people who have created this AI through large language models want to obfuscate you and your contributions to open data.
What do most think are the most important forms of AI? There are deep learning enthusiasts. There’s causal discovery—Pitt has a Center for Causal Discovery. There’s knowledge engineering through rule transfer learning. Guess what the public now thinks is most important? Generative AI.
So, in my rabid enthusiasm for AI, my first lesson to you is to please follow and carefully read the science and understand that generative AI is not the most useful machine learning/artificial intelligence tool for pathology practice and translational research.
Broadly, clinical pathology, especially molecular pathology, and the emergence of genomic diagnostics in clinical microbiology may be an increasingly critical part of pathology. I say that as an anatomic pathologist who is facing the integration of multiplexed imaging into cancer diagnostics. If you talk to a patient and most health care workers, if you are an anatomic pathologist you’re identified as pathology. If they do diagnostic testing, genomics, serum, or blood, we’re the laboratory, or lab medicine. It is a branding issue for us as clinicians that our patients and hospital caregivers distinguish us and recognize us as two separate diagnostic disciplines with different names. Yikes!
What should diagnostic pathologists be called? This meeting is called Clinical Lab 2.0. As an anatomic pathologist I always believed its focus was clinical pathology. The audience at this meeting is unbalanced, with few anatomic pathologists, and is emblematic of our societies in pathology, which are equally imbalanced. The enemy is us, as the chasm between AP and CP along with the fracturing of molecular pathology is a major problem. This is something that’s plagued me as an innovator in pathology. There’s this amazing promise that AI can bring anatomic and clinical pathology, including molecular pathology, together, if we share our data among ourselves and leverage that asset to the communities that care. Our patients are the most important community that cares that Lab 2.0 integrate effectively with our many fractured professional societies and annual meetings.
I care about people in my inner circle—my family, friends, associates, alumni from my many programs and initiatives. Anytime anybody calls me with a health care problem and asks to be connected to the right provider, I’m wicked fast and capable of doing that as an anatomic pathologist. I help to integrate their AP and CP and molecular pathology results to help guide their therapy and choice of clinicians. We as pathologists are a valuable resource for our inner circles and patients in general. I challenge Lab 2.0 to get us into the critical role we can play for more patients.
We got out of the business of pathology, of connecting patients and their needs to the right doctor. Getting off track is our problem. We’re dazzled by the new technology. I have a startup company; I’m an investor, and it’s my fourth. My first three companies were whole slide imaging companies. We’ve changed the world with whole slide imaging. We haven’t even touched it yet with what we could do if anatomic and clinical pathology were to become integrated and we could lead the implementation of precision—personalized—medicine. That is the goal of my fourth startup company.
If we create a diagnostic cockpit for our pathologists to be knowledge engineers, instead of pushing the report out and getting the billing right and making sure we maximize point-of-care testing billing, we can change the world with our own integrated pathology data. I still believe firmly that 70 percent of all the decisions made around patients start in the pathology laboratory with an abnormal test or a test that’s trending in the wrong direction, or a life-threatening diagnosis of cancer or HbA1c that’s off the charts. But we don’t leverage our own data—like crude oil—and refine it into knowledge, using AI, instead of a litany of individualized laboratory/pathology reports buried in some PDF or spreadsheet in the electronic health record systems to which we are all slaves.
I come from the South Side of Chicago. My dad was a foreman at what used to be called Standard Oil, now BP. I spent a summer, as a college student, working in the laboratory and saw them turn crude oil that stinks and pollutes into extremely valuable products. When people ask me what I am today as a biomedical informatician and revered cancer researcher and pathology informatics innovator, I say I’m a plumber—data refiner. That’s all I do: I pay attention to the valuable data in health care and make sure it gets to all the right people including clinicians and researchers, and I still need a solution to share it with patients in a digestible and informative manner. This is the place where generative AI for pathology may be useful. Perhaps Lab 2.0 can assist in this important patient-facing journey.
When I say integrate pathology (AP) and laboratory medicine (CP) data, I mean wrapping it with metadata, with the right kind of attribution, the right amount of descriptive data. What do you think is the biggest polluter in my environment in getting the right data to the right people, meaning the patient? EHRs, 100 percent. How many people have looked carefully at how our laboratory data is represented in the EHR? Yes, it’s depressing.
The reason I’m enthusiastic about AI comes down to three things. It will help unify anatomic and clinical pathology laboratories. We must be clear in our messaging who we are—and it’s not the laboratory or pathology. We must be clear to our health care colleagues and patients who we are. Get out of the basement and get out in front of providing the right care to the right patient at the right time, without duplicative testing.
Genomic germline testing is an exemplar. How many times does that test need to be done on patients in the United States? Once, and everybody knows that. We’re all pathologists—we get that. Let’s do that test once. Let’s then provide a way to share all of our data, especially germline testing, in one common pathology-driven repository, so that testing is done only once and at high fidelity and with sufficient metadata for integrated reporting. And when we do that test, maybe we should preserve the RNA and DNA—perhaps for the lifetime or two lifetimes—so as the technology evolves we can repeat it with the next generation of genomic testing. Tissue banking and RNA/DNA storage should be part of long-term strategic plans fueling the future of our discipline.
There’s a revolution going on in our microbiology labs—what are we doing to share the data of the microbiome and prepare for that evolution? Pathology is a tremendously valuable part of the solution space based on the establishment of molecular pathology as the newest “specialty” testing. Seventy percent of all the diagnoses that affect a patient’s life happen in our pathology laboratories. Seventy percent of the most valuable data that doctors use in EHRs is pathology/lab medicine data. And with all that value, we’re now up to only two percent of the hospital budget. Let’s leverage that return on investment with getting to disease-specific integrated AP/CP/molecular pathology reports instead of dozens of specialty laboratory reports from multiple laboratories.
We must get busy understanding that AI is the glue that can unify our pathology data and enable it to speak loudly for the role of pathology in medicine. This goal is why I’m a rabid AI enthusiast. This is also why I’m sitting in a basic science biomedical informatics chair in a medical school. That’s why I’m still involved in startup companies, solely to solve some of the faults standing in the way of the goal of an integrated pathology report. I’ve been busy creating industries, like whole slide imaging, that fuel the discipline of pathology. That’s what I think is my most important contribution: to try to glue together the many evolving laboratories of pathology. So far I have failed at accomplishing that worthy goal. The Association for Pathology Informatics is growing, we’re doing our thing, we have a member group—it’s great. But API has failed at changing our major pathology organizations and facilitating the integration of the CAP, USCAP, and the ASCP, much less AACC, AABB, and AMP. The list goes on.
If you go to a pathology meeting, you’re largely going to get a tiny slice of the diagnostic pathology pie. We need to learn from our colleagues in radiology. What’s the largest medical meeting in the world every year in the U.S.? Radiological Society of North America, or RSNA. There aren’t 70-70-70 numbers in radiology; there are 70-70-70 numbers in pathology. I’m a cancer tumor cell biologist and I’ve dedicated my life to cancer research. What is the other 70 number? Seventy percent of what we do in anatomic pathology is in service of cancer. With those daunting numbers, why can’t we get ourselves together to do what we need to do to transform the practice and enable the vision of precision pathology in the service of our patient and health care partners?
The most compelling thing we can do is to bring as many young people into our discipline as possible—and now. Look at Pittsburgh’s STEMM programs: science, technology, engineering, math, and medicine. We’re recruiting the next generation of people who are AI-informed, data-infused young people whom all of you can use in pathology administration and business and, more importantly, in practice, in AP, CP, and molecular pathology. We should all be doing that because the future is our young people and opening their eyes to the promise of pathology and laboratory medicine infused with AI-driven innovation insights.
I am confident we’ll solve the pathologist workforce problem with AI—intelligently designed diagnostic tools with the goal of integrated pathology reporting. This will reveal the fascinating and mission-critical lives we lead as pathologists. I am joining the Lab 2.0 team to make sure this happens.
See “CAP Today Recommends” at www.captodayonline.com for the Clinical Lab 2.0 proclamation (https://bit.ly/CT_Lab2-0).