Anne Paxton
March 2025—For clinical pathology practice, generative artificial intelligence can open new efficiencies and opportunities, and the authors of an article published in Archives of Pathology & Laboratory Medicine set out how it can be used and its risks.
“Clinical pathology is a team sport, and it lends itself to a lot of quantitative analyses and assessment of data,” says Rama R. Gullapalli, MD, PhD, associate professor of pathology and of chemical and biological engineering at the University of New Mexico and a coauthor of the article (McCaffrey P, et al. Arch Pathol Lab Med. 2025;149[2]:130–141). Understanding medicine at scale or producing a population-level interpretation of tests is clinical pathology’s domain, he says.
“The core data for AI was mostly one way for a long time—it was human generated. Now with generative AI, it is bidirectional,” Dr. Gullapalli says. Generative AI can look not only at structured databases but also unstructured data such as handwritten notes. “Clinical pathology is such a large repository of patient data, it’s a very attractive proposition to AI scientists to take all of this data and build these models to predict patient outcomes, population health, and so on.”
Even for a given patient, he says, the volume of data can become so large, and the need for data interpretation so central to how care is managed, that “robust tools like generative AI, like large language models that can survey the data and make sense of it,” become essential. It’s what generative AI can do, he says: “Pull meaning out of a large amount of data in a way that is accessible and interpretable.”
On the level of a single laboratory, committing to memory all of clinical pathology’s many tasks, standards, guidelines, and so on is impossible. “That’s where a technology like generative AI can help put everything together,” he says, “so that everyone on the team becomes some degree of an expert. They have access to all this information and can retrieve it in real time.”
Numerous use cases await clinical laboratories that decide to take advantage of generative AI, in particular large language models, Dr. Gullapalli says, citing the need to improve workflow, manage inventory efficiently, and select instruments. “Those are pan-clinical pathology issues, whether it’s in clinical chemistry, microbiology, hematopathology, or molecular diagnostics,” and solutions differ by laboratory. Generative AI makes possible more systematic, broadly applicable solutions, he says.
Dr. Gullapalli and his coauthors suggest clinical laboratories begin with generative AI tasks that don’t have a direct impact on clinical decision-making and are easily audited by human experts. They cite as an example using a large language model to spot areas within an existing standard operating procedure that are unclear, contradictory, redundant, or inaccurate.
Here are a few of the more than 30 workflow enhancement opportunities they cite for clinical chemistry, microbiology, hematopathology, molecular pathology, transfusion medicine, and clinical informatics:
- Automated interpretive drafts and resulting of chemistry values.
- Tailored results with dual patient-clinician interpretive reports.
- Summarization of microbial epidemiologic and resistance databases.
- Integrative report creation using multimodal data (flow cytometry, molecular, histopathology).
- Search and pattern identification of patient molecular data at scale across databases.
- Real-time error correction modules for safe transfusion practices.
- Automated SOP generation and dynamic text protocol templates for AP/CP.
Dr. Gullapalli advises choosing use cases wisely, starting small and expanding from there. Small projects might be editing and updating standard operating procedures, using large language models to assess whether laboratory policies and procedures align with accreditation requirements, and streamlining human resource functions. “We need pilot studies to implement these at small scale,” he says.
In his own subspecialty of molecular pathology, information management is difficult. “We are talking petabytes of data,” Dr. Gullapalli says. “A data analysis tool like a large language model that is trained on all of these databases could potentially give more focused answers to ease my workflow and could be incredibly helpful.”
Using generative AI can be seen as a variation on delegating tasks to others, says coauthor Peter McCaffrey, MD, MS, chief AI officer for the University of Texas Medical Branch at Galveston health system.
He is director of laboratory information services, pathology informatics, and the Division of Bioinformatics and Artificial Intelligence, as well as faculty in the pathology, radiology, and cardiology departments. “So I have an interesting perspective on the kind of outsized power we have in pathology to find and navigate things that are new, both to our internal operations and, more broadly, to the health system leadership and enterprise.”
Of any job as a leader in pathology, he says: “You’re not reading everything yourself, even today. You’re delegating it.” Using generative AI “doesn’t mean you’re free from having to vet the content, but it does mean you have an infinite account, if you will, to delegate. There’s infinite bandwidth you can delegate to. It’s imperfect, but it’s pretty good and it’s not that different from your current labor pool. It spares everyone from having to do things that are not very meaningful to them.”
ChatGPT can do the core tasks that AI is good at for searching, retrieving, and synthesizing data, he notes. “Even if you don’t say, ‘Write my urine toxicology interpretation,’ you can say, ‘Just surface for me what meds they are on, what their urine drug screen results are, the relevant details about their past.’ This kind of thing is where generative AI has tremendous value. Its assistive ability, to let me focus only on the interpretive part, is where the biggest value is and where it can liberate a lot of what we do.”
Selecting instruments is another use case, he says. “Say I’m going to consider a new chemistry vendor, for example. I could go to ChatGPT, turn on the web search console, and ask what vendors are out there, what models they have, how I find their product documentation, what the big points of difference are between Roche Cobas versus Siemens versus Ortho.” Having a senior resident do this could require weeks, “but it can be generated in seconds now.”
Going further, one could ask for an analysis of the papers the vendors cite and for critiques of each paper.
Finding contradictions between standard operating procedures would “be highly impractical to accomplish through paying 20 technologists to do it,” Dr. McCaffrey says. “You could do this today by asking ChatGPT to do this kind of comparison.”
Generative AI could provide decision support in clinical chemistry and microbiology, he says. “One example could be asking, ‘Would the results of an assay be influenced by anything else that’s in the patient’s clinical record already?’” such as a medication, an active disease process, or a recent transfusion. “Now, in most laboratory information systems, we just have the standard cutoffs like the normal reference ranges. But how personalized are those?”
Another example is contextualizing to the patient chart. “Say a person is not yet abnormal in their CBC but they’re trending that way, if you looked at their last history of CBCs. Theoretically, the ability to catch that earlier would reflect in the rapidity of care and potentially the outcomes for things like colon cancer.” He calls it “opportunistic screening” and says the clinical laboratory is ideally placed to achieve it. “We are where we have the most data the earliest in care. We’re in the right place to do that, and it renders a new perspective for decision-making.”
An efficiency for a future time: being able to anticipate the laboratory results a specialist will want when a patient is referred, such that appointments are not compromised and repeat visits are unnecessary. “Can AI basically say, ‘Can we get everything needed done so it’s ready at the first appointment?’ We could potentially move people’s therapeutic schedules up by months.”
In laboratory test ordering, AI could provide physicians with direction. “It’s not doing diagnosis,” Dr. McCaffrey says, “but it says, ‘You should consider something you missed’ or ‘Based on what is going on, you should add this to the order set.’”
“A lot of clinicians I talk to aren’t defensive about that gap,” he continues. “A primary care provider might say, ‘I don’t enjoy the position of getting a test menu of 1,000 things. Could the test menu be more intelligent and give me suggestions?’ That would be a classically CP thing to do. We manage the test inventory. We at least validate the order sets. We can also provide soft guidance on how they’re used.”
Dr. McCaffrey predicts generative AI will prove to be as useful in clinical pathology as in anatomic pathology, though “its manifestation might be a little different” and the potential is greater.
“The volume of things you can interpret in CP is so much higher than the volume of things you do interpret. In theory, what would a world look like where clinical pathology gave some level of interpretation to every result? That’s a world achievable only through the use of AI, if and only if AI does most of the lift and we focus our practice on adjudicating worthwhile interpretations.” To Dr. McCaffrey, that is likely the most interesting of all use cases. “It lets us change essentially the service model we have behind laboratory test interpretation.”
For those starting to use generative AI or thinking about it, Dr. Gullapalli says “literally everyone in this space is starting out at the same level now. There are no experts and no naive people.” It takes only curiosity and a willingness to learn. “You have to experiment on these interfaces. You have to ask questions that are focused, precise, and content relevant.” It is called prompt engineering.
“If you distill it,” Dr. McCaffrey says, “what is the job of pathologists? Our job has always been to interact with, embrace, and develop rigor around new technologies. We’ve been doing it ever since the microscope was invented.”
He advises against over-mystifying or under-mystifying generative AI. “It’s not this magical thing that is going to nuke everyone’s job. It’s not this crazily dangerous thing that we should get rid of. It’s just a tool that has good and bad but can be really helpful if used.”
“Let’s say we’re in a field that does a lot more inference, that has a lot more cognitive deliverables,” he says. “What do you think you need to be prepared for? I would say how to prompt, how the tools work, and understanding how the data is retrieved and acted upon.” Pathology residents have learned far harder material in the past—genomics, for example, he says. “So this is not something to be afraid of.”
“We’ve literally handled way more dangerous stuff,” Dr. McCaffrey says.
In his view of the future: “Pathology becomes more important because what people will want out of us is inference all the time, everywhere, on everything constantly.”
The same technology that offers “immense” potential also introduces novel failure modes, Drs. Gullapalli and McCaffrey and coauthors write, one of which is confabulations, known as hallucinations, “exceptionally drastic in the health care setting.”
But Dr. McCaffrey sees the hallucination as not quite a failure mode. “The large language model is sort of like your brain. It has a linguistic and memory center. Its job is to hallucinate. You say something and it riffs on it. So we shouldn’t expect to have hallucination-free experiences. We should consider how to constrain hallucination and how to catch it if it arises.”
The biggest pitfall, in Dr. Gullapalli’s view, is automation bias, which he says is “being deferential to these algorithms and models to the point of ignoring your own training, instinct, and knowledge.”
“Large language models are not a substitute for human thought and judgment, especially in the medical context,” he says.
In their article, they point to the pitfalls that are most important to clinical pathology. One is that generative language models often struggle with numerical operations and comparisons, they write, “which may impair the reliability of assessing trends and shifts in laboratory values in CP.” Another is that the models may hallucinate details in a clinical interpretation such as patient comorbidities, which would not only compromise the integrity of the interpretation but also force pathologists to check every cited comorbidity.
If generative AI doesn’t achieve a performance level that allows it to operate largely unsupervised, the authors write, “clinical pathologist activities may just shift, being responsible for many more model supervision and output verification tasks in exchange for any decrease in interpretive thought work.”
While these and other risks exist, generative AI’s future is “exciting and incredibly bright,” they write, “as we are witnessing the beginnings of scalable and programmable cognitive work.” And clinical pathology is “well positioned to become the core ‘copilot’ within the health care workflow at large.”
Dr. McCaffrey envisions a future in which pathology is considered more of a cognitive utility rather than a service. “Like electricity, we infuse the EHR everywhere and therefore we’re asked to infer things more often. But maybe we rethink the model of value so that we have to be more of an infrastructural part of the hospital and the EHR, not just doing the data but generating some level of opining about the data.”
He thinks of lab testing for primary care, “where you’re catching that subtle anemia that will become indicative of a colon mass no one caught yet. This is a gap we’re uniquely suited to fill.
“We sit on the data channels,” he says, “and we have always been a cognitive utility, even before AI.”
Anne Paxton is a writer and attorney in Seattle.