Summary
Despite advancements in algorithm development, implementing AI in clinical practice faces challenges. While data acquisition is no longer a bottleneck, infrastructure, reimbursement, and clinical workflow integration remain obstacles. Validation of AI models is complex, requiring careful consideration of preanalytical variables and establishing appropriate comparison metrics.
Charna Albert
January 2026—For Jansen Seheult, MD, and others deeply committed to bringing artificial intelligence to the laboratory, it’s impossible to deny that algorithm development has advanced by leaps and bounds.
Take Mayo Clinic in Rochester, where Dr. Seheult is medical director of digital pathology and artificial intelligence in hematopathology. More than 20 artificial intelligence algorithms are deployed across the clinical practice in pathology and laboratory medicine. Yet that’s a small fraction of all the algorithms that have been developed at the clinic.
“We have 10 times more algorithms in development or that have been developed, compared to what is in production,” says Dr. Seheult, assistant professor of laboratory medicine and pathology. That number is emblematic, in fact, of what Dr. Seheult and his colleagues are finding as it comes time to implement in clinical practice algorithms developed in a research or translational setting. “There is a bottleneck,” he says. “Operationalizing an AI model goes beyond just developing the model. The challenge is fitting that model or pipeline into an existing clinical workflow.”
Not long ago, the sticking point in the process was data acquisition and curation. “We have the resources now to develop the algorithms,” says Carolyn Glass, MD, PhD, codirector of the Division of Artificial Intelligence and Computational Pathology and associate professor of pathology at Duke University School of Medicine. “The bottleneck now is the infrastructure required to clinically implement and monitor the safety of those algorithms, as well as the reimbursement structure.”
On the research side, investigators may not appreciate the complexity of clinical workflows. “Even pathologist investigators don’t necessarily know what goes into bringing one of these algorithms live in the lab,” says William Jeck, MD, PhD, associate professor of pathology and Dr. Glass’ colleague at Duke. “Algorithms created and developed in a research context have to go through a kind of vetting system to become ‘clinical grade.’”
Moreover, model performance claims have to be established or verified before deployment and clinical use. “When we go to validate one of these algorithms, it becomes more complicated,” Dr. Jeck says. Formal requirements for turning an algorithm into a laboratory-developed test have yet to be issued. “And it’s not entirely obvious what the relevant comparison is. Is the relevant comparison the result you get reading a slide with glass versus what the algorithm says? Is it what you get reading a slide digitally versus reading a slide digitally with the algorithm? That’s the sort of thinking that has to go into what’s the proper comparison, the proper validation schema, and all of that is difficult.”
Many of the familiar regulatory and quality frameworks apply to AI, Dr. Seheult says. “CLIA is broadly extensible to AI and the machine learning components of a clinical test, as we have extended it in the past with other new and emerging technologies like mass spectrometry and next-generation sequencing.” CLIA does not call out the specific details or protocol on how to establish or verify performance claims, however. “For accuracy or precision studies, for example, it doesn’t necessarily say what sample size you should use or how you should design that study. That’s mostly left up to interpretation.” In this regard, Dr. Seheult believes laboratories can rely on the familiar guidance frameworks used for method validation and verification studies of in vitro diagnostics and laboratory-developed tests, with minor modifications to account for the unique aspects of AI and machine learning. (Recommendations for evaluating machine learning systems in the clinical practice of pathology were proposed in a white paper published in 2024: Hanna MG, et al. Arch Pathol Lab Med. 2024;148[10]:e335–e361.)
One task is to anticipate the preanalytical variables and other failure modes that might affect algorithm performance. “We need to outline what those preanalytical variables might be for AI,” Dr. Seheult says.