September 2, 2026 | Artificial intelligence (AI) can help reengineer pathology toward the point of care and provide guidance to first-on-the-scene responders, including those in military battlefields with no internet connectivity and limited medical training. The technological possibilities are driven largely by the data foundation that AI models learn from and analyze as well as institutions finding the “Goldilocks” speed at which to move on adoption that is neither reckless nor paralyzed by fear.
These were among the key points made in the opening keynote address at last week’s Next Generation Dx Summit by Nam Tran, Ph.D., associate dean for biobanking and multimodal data at the University of Pittsburgh as well as medical director for point-of-care testing (POCT) at the University of Pittsburgh Medical Center (UPMC) composed of 45 hospitals and 800-plus clinics. In addition to being trained by one of the founding fathers of the POCT field, Tran is co-inventor of three AI platforms with a zero-trust, edge computing design built to operate entirely locally.
The model of pathology and laboratory medicine is changing, necessitating a move away from centralized excellence in ivory tower settings toward what he terms “diagnostic power projection” where capabilities and technical expertise expand beyond the confines of buildings and geographic borders. This future state is marked by smart decisions being made on local devices, near-patient guidance, remote review and consultation, and the ability to easily scale.
Laboratory workforce shortages “remain persistent [and] there is no clear end in sight,” says Tran. Meanwhile, the emergence of new diagnostic technologies is outpacing the foundational capability of many institutions to bring them online. It takes some institutions two years to introduce a point-of-care test, even as diagnostic demand is increasing due to hospital consolidation.
UPMC itself is “very expansionistic,” he says, adding five hospitals in the last year alone. The catalysts for smaller hospitals to join larger organizations are lowering reimbursements that are tightening the purse strings. “Ultimately, pathology [and] lab medicine ... must learn how to project [its] diagnostic power more broadly, because that is going to help us capture more customers ... [and] expand with these growing health systems.”
Two-thirds (67%) of hospitals nationwide were affiliated with a health system as of 2022, up from 56% in 2010, reports Tran.
The shortage of medical lab technicians and scientists existed prior to the pandemic, but COVID-19 made it painfully plain, says Tran, who at the time was clinical pathology director at the University of California, Davis. Nearly three quarters of labs were already reporting being understaffed, about a third of them significantly so. In addition to rising demand, and fewer people entering the profession, this workforce is also “aging faster by 78% than the entire U.S. labor workforce.”
Burnout is a significant factor, affecting 85% of lab professionals at some point and 50% currently, with 69% feeling that a career change is forthcoming, he says. The culprits include a lack of adequate staffing, heavy workload, and pressure in completing the ever-growing volume of post-pandemic tests on the market—almost all of them in vitro diagnostics that get obsolete quickly.
Reimbursement is also going down, continues Tran. Payment reductions of up to 15% are on tap for next year on the heels of three prior rounds of 10% cuts, and more of the same is coming in both 2028 and 2029. “The government shutdowns aren’t helping, those are now becoming more frequent, and that level of uncertainty doesn’t help our workforce in any way.”
The “easy answer is we need to work smarter and do more with less,” but that’s easier said than done, he points out. “We need to figure out how to glean savings through efficiency improvements ... [and] force multiply” so individual pathologists and laboratory scientists can accomplish more without exhausting themselves.
This necessitates improving clinical and operational excellence using business intelligence to know where inefficiencies lie, so they can be fixed, and then developing strategies to move diagnostic capabilities closer to the point of care, says Tran.
Point-of-care testing in the world of lab medicine happens at or near the site of patient care, with the goal of accelerating the therapeutic turnaround time from test order to treatment, says Tran, which is what ultimately matters. The POCT arsenal is expanding beyond disposable, transportable devices to include smart devices like the continuous glucose monitor available over the counter that he has been wearing for the past 18 months, as a non-diabetic, simply to learn how what he eats impacts his body. That sort of wellness data is starting to show up in electronic heath records (EHRs) and potentially mineable for future technologies, he adds.
Tran proposes that point of care be redefined as “diagnostic expertise at the time and place of patient care,” which could be data from a device that gets interpreted by a pathologist at a facility 100 miles away from an individual. POCT is already available in chemistry, blood gases, cardiac biomarkers, and infectious diseases, including molecular tests, he says. Some of these are waived under the Clinical Laboratory Improvement Amendments as simple, low-risk diagnostic tools.
That list includes pregnancy, diabetes, coagulation, and hematology tests that have helped in “projecting diagnostic power to the world” in the context of field response (e.g., ambulance) as well as at community hospitals, remote/austere environments, and in emergency departments and intensive care units (ICUs). “That has made a difference because they’re generating data, providing clinicians with insight, helping them make a decision, and generating some sort of impact,” says Tran.
POCT as well as core labs and molecular diagnostics are increasingly being digitized, and the idea is starting to gain traction in surgical pathology with whole slide imaging. All those raw signals being generated can be structured to create actionable intelligence, he says.
A data foundation is the prerequisite to capturing high-quality digital images and then putting them in a standardized format and pulling in metadata for clinical context from EHRs, adds Tran. This is what allows for remote reviewing, near-patient guidance, and network expertise at a scale that keeps pace with the growth of health systems.
Digital pathology brings together whole slide images, digitized histology, and remote pathology review into a single interconnected workflow. But digitalization isn’t enough, says Tran, since it doesn’t address how pathologists are to directly interface with partner facilities and patients and aggregate data into meaningful and actionable knowledge.
This is where AI comes in to enable predictive analytics, real-time monitoring, risk stratification, and diagnostic decision support, he continues. Both generative and non-generative AI have a role to play.
Non-generative AI typically means machine learning systems that analyze existing data to help make predictions, classify diseases, and make decisions through pattern recognition, Tran explains. This describes most of what is currently being used in pathology and laboratory medicine.
One example is image classification where a whole slide image gets processed by various neural networks to label patterns representing some sort of disease, he says, citing the “no-code, end-to-end” Whole Slide Imaging (WSI) Genie platform developed at the University of Pittsburgh. WSI Genie unbiasedly identifies features, pre-processes images, and then reveals patterns of potential interest.
Many AI platforms exist that can detect patterns, says Tran, but the output does not eliminate the need for the pathologist to have human-to-human or human-to-machine interaction. A slew of questions can arise at consultation: “What does this mean in context? How confident is this output? What should I do next? Can I ask an expert? How does this fit the case?”
This is where generative AI enters the story—specifically, the more sophisticated, multi-modal variety that can bring together different forms of data (text, audio, images, code, and synthetic) to produce meaningful output, he says. These platforms employ multiple AI agents that specialize in different domains.
In terms of digital pathology, a chatbot layer adds the ability to ask questions in natural language and summarize findings for human-in-the-loop review, says Tran. Based on report generator work being done with AI at the University of Pittsburgh, the time savings for pathologists is substantial and the report summaries “provide guidance for next steps, knowledge to users, and ultimately extends the expert’s reach.”
Concerns about health information privacy and computing cost can be remedied by bringing chatbots onto an institution’s own servers and network firewalls, although it remains to be seen how the capabilities compare to public, cloud-based models such as OpenAI’s ChatGPT, Google’s Gemini, and Meta AI. Locally installed GPTs are “as private as you want them to be,” says Tran, noting that even those confined to a laptop can have remarkable offline capabilities.
Tran is co-inventor of an AI platform, developed at the University of Pittsburgh, called Pitt-GPT-Plus. It’s a multi-agentic local chatbot that can be customized for any task—not just to look at lab procedures and data, he says.
The opportunity to stress-test the AI model arose after Tran and his team were approached by an Eastern European nation that wanted soldiers in the field to be able to quickly access battlefield medicine knowledge, with a goal of 90% accuracy (the “alternative of having nothing” being much worse), he shares. The soldiers were receiving only three hours of medical training and operating in a “communication-limited, emission-controlled, and electronically contested environment” as a matter of survival.
Pitt-GPT-Plus was adapted to become Field Care GPT, in collaboration with the University of Pittsburgh Center for Military Medicine Research (CMMR) and runs on a high-performance laptop with no internet connectivity. It is trained solely on battlefield medicine information—specifically, 29 military documents, including the 75th Ranger Regiment Medic Handbook, says Tran.
Field Care GPT has voice-to-text capability and knows multiple languages, he adds. It has also been evaluated with emergency medicine and ICU physicians and trauma surgeons, as well as an Army Special Forces Medical Sergeant.
In the real world, the conversational AI chatbot delivers an accurate response in 30 seconds. In the case presented, a soldier had been shot in the chest and having increased difficulty breathing, and Field Care GPT directed the medic to check the person’s airway and if the individual was having trouble breathing to put a chest seal on the hole. If the problem worsened, an emergency medical procedure (needle decompression) would need to be performed to treat a complication called tension pneumothorax, and the chatbot would then walk the medic through that procedure.
It appropriately cites all its responses back to the documents it was trained on, Tran notes. The chatbot also has a memory, enabling a conversation to continue as needed.
In formal validation studies where Field Care GPT was tested with 200 prompts, most for acute injury and prolonged healthcare, it got 98% correct, he says. “The 2% it didn’t get right was because of bad prompts,” such as telling the chatbot “My arm is burning” without the context of it referring to an arm being on fire or a rash.
When fielding questions for properly treating military working dogs, the chatbot scored 100%. This included an appropriate “I don’t know, I wasn’t trained on that” response when purposefully asked out-of-scope questions about the weather and sports figures. In quality verification testing, 87% of the time the structure and syntax of responses matched the ideal human response—excellent performance, Tran says, given that in the English language there are many ways to say the same thing.
The average response time with Field Care GPT is 32.4 seconds, says Tran, adding that it has been used in real-world observational trials with law enforcement working dog classes. When studied in parallel with SWAT-related medical training in rural California where there is zero cell phone reception, it got 100% of the scenarios correct within 60 seconds and left first responders asking for a copy of it.
Only a few weeks ago in Michigan, the CMMR participated in Northern Strike, one of the largest multi-agency, multi-branch, multi-national military exercises. Field Care GPT was tested alongside drones and other cutting-edge medical evacuation technologies.
At the University of Pittsburgh, Field Care GPT is now being translated to the world of pathology where it could find utility in report writing, customer support, and simply aggregating data, says Tran. “We just have to train it on exactly what we want it to do.”
Realizing these ambitions is going to take more real-world healthcare data to train on than is available for the job because, for good reason, it’s hard to access other institutions’ data, he continues. Fortunately, generative AI not only powers chatbots but can be used to make synthetic data. By 2030, much of the data being used to train and develop AI is expected to be “not real ... [but] based on real data.”
There are many ways to generate synthetic clinical chemistry data using 1s and 0s, says Tran, as well as platforms to generate synthetic images. AutoPix AI is the technology developed by the Computational Pathology and AI Center of Excellence at the University of Pittsburgh, with a non-generative component to help do segmentation and pick out patterns within images and a generative component to generate synthetic images—no coding experience required—and tell human users if the output of the synthetic data is good or bad. A demonstration of this involved a low-quality picture of an acute myelogenous leukemia pulled in from Google, used to generate four images graded as “poor” by AutoPix AI and prompting researchers to input more pictures to get better synthetic images.
“The end state is medical sensor fusion,” says Tran, referring to a sought-after means of combining data from multiple health and environmental sensors into a single, highly accurate picture of a patient's physical state. The concept is not new, and the same logic applies to self-driving cars using cameras, radar, LiDAR (light detection and ranging), and ultrasonic sensors to make meaningful decisions. In the healthcare world, raw inputs include demographics, patient history/physical exam data, genomics, point-of-care tests, and wearable devices.
The roadmap to diagnostic power projection, the primary objective, follows the stage of “AI-ready diagnostics” that likely describes few facilities today. Many institutions are at the digital foundation step where they are still standardizing varied medical data into a format that AI models can utilize, and in California some remain stuck in the early digitalization step because a state law only recently allowed remote sign-out.
Changes to institutional training programs are needed and, at the University of Pittsburgh, this began August 1 with the launch of an interactive online AI education course for the entire medical residency program that has “AI in the background teaching people AI,” Tran reports. Educating people further will require a culture shift that is more forward-thinking and positive about AI, as is certainly the case at his home institution. “It takes time and building the right team to help push that forward.”