Valerie Neff Newitt
October 2026—Antibiotic resistance has changed dramatically. How laboratories detect antibiotic resistant infections and how patients are treated have not.
“We treat patients with our best-guess drugs, and we make those guesses based on the types of infections we’re seeing in the patients we take care of and what our local antibiotic resistance rates look like,” says Romney Humphries, PhD, D(ABMM), M(ASCP), director of the Division of Laboratory Medicine at Vanderbilt University Medical Center. “But increasingly it can be hard to guess,” she says, “because patients who have not had traditional risk factors for some of these resistant infections are showing up in our hospitals with them.”
Many innovators are coming up with creative ways to take on the antibiotic resistance problem, she says, noting they don’t rely on visible growth or no growth of the bacteria in the presence of the antibiotic but instead on other endpoints—with faster results. And though it’s “really, really early days” for artificial intelligence, she says, “the information has been promising to date.”
“I think in our future there’s a pretty big shift in our clinical microbiology labs from performing a lot of cultures to performing a whole lot of sequences,” she predicts. And the tools of the future will need to be cost-effective in all settings.
“We need to understand this is not something that isn’t going to cross borders,” Dr. Humphries says of the emergence of multidrug resistant bacteria.
The use of AI for antimicrobial resistance and susceptibility predictions is among the topics in the lineup next month at the AMP meeting in Seattle, where Dr. Humphries will be a plenary speaker. Quality monitoring in molecular microbiology is another, to be presented by Christina Wojewoda, MD, of the University of Vermont Medical Center, and Thao Truong, PhD, D(ABMM), M(ASCP)CM, of UW Medicine in Seattle. They’ll address practical compliance in smaller and larger laboratories.
“There are multiple ways to fulfill regulatory requirements [for quality monitoring] and to ensure the best patient care, so we’re going to give examples of what we’ve done in our laboratories,” Dr. Wojewoda says.
Several groups are working to develop neural networks and machine learning models, Dr. Humphries says, “to train the algorithms with the bacteria we have seen to recognize resistance from a genome perspective, but even more importantly, to be able to predict, when they see something not seen before, whether or not it is going to cause resistance.”
“In theory,” she adds, “you would not even need to grow the bacteria.” Instead, sequencing work could be done directly from the infected human sample.
At Vanderbilt, Dr. Humphries’ laboratory investigates resistance mechanisms as they emerge, sometimes beginning with what she calls “the old tried-and-true phenotype.” When an organism is known to be resistant but the laboratory doesn’t know why, sequencing can reveal suspicious features in the genome. “You can then clone those features in a wild-type or susceptible bacterium to see if you can recreate the phenotype,” she says.
For genomic susceptibility prediction, AI offers a different approach: Machine learning models look across bacterial genomes for features associated with susceptibility or resistance.
“You train your model with a whole bunch of bacteria for which you have the sequence and know the susceptibility results,” she explains. New organisms can then be used to challenge the model. AI can be particularly useful when a bacterial genome contains many mutations. “We can’t get our minds around what it means if we have bacteria with 25 mutations in it, but AI can,” she says.
AI scans the entire genome. As the model is trained, a feature might be found in the genome that predicts close association with the bacteria being resistant to a drug. “We’ve found you can then reverse engineer that to say, what is that feature it’s looking at and what does that part of the genome do? We’ve rediscovered [known] resistance mechanisms,” which in a sense validates the approach. “We’ve also found it may not be the resistance gene that predicts resistance but rather the regulator of that gene.”
Could such models eventually provide an early warning of emerging antimicrobial resistance? “Yes, absolutely,” Dr. Humphries says, provided the data set used to train the model is robust.
The work remains early, and for now AI use in clinical microbiology is limited (see “The path to AI-aided parasitology”). But broader use of AI in clinical microbiology is not a question of if but when, she says. For AI-based susceptibility prediction, she estimates routine use in five to 10 years. “It goes back to how simple it is for laboratories to implement, how widespread it is, and how comprehensive the tools are.”
Even with well-trained algorithms, she says, a person will always be needed for result interpretation, as they are for other laboratory testing. “We allow the instrumentation to auto-report some things and other things need to be reviewed by a human. I see this as similar to that.”
Improving antibiotic resistance testing has been difficult for several reasons, Dr. Humphries says. Technology is one: It’s hard, she says, to develop the tests that work as rapidly as needed. “We don’t fully know how all these bacteria become resistant to antibiotics. Typically, it’s a mixture of different mechanisms, and sometimes the bacteria develop resistance in ways we don’t predict.”
In addition, many bacteria types cause infection, and the number of antibiotics that need to be tested is large. “So you can’t do this cheaply; you need sophisticated methods.”
The regulatory framework, too, is a barrier, with the Food and Drug Administration taking a “narrow view” of the types of tests it will approve, Dr. Humphries says, “and that is only tests for drugs that are used within the drug labeling.” But antibiotics are regularly used off-label because of the number of different species of bacteria that can cause infection. “It is impossible to do a clinical trial to study every one of those species and develop FDA claims for a drug, and so physicians can get around this through the practice of medicine. They’re able to prescribe drugs as they see fit.”
Laboratories, however, are not able to test these drugs in whatever way they best see fit. “So that means as new technologies come to market, they are at a disadvantage compared with legacy devices that take two to three days to get the results out,” she says, “because those were cleared through the FDA in an era when the FDA did not have these strict rules.”
New technology also has to fit into the laboratory’s workflow (in addition to being accurate), and the tests have to be comprehensive, “which is where some of the regulatory world works against us.”
Then there is the timing of when the drug and the test come to market. “We need to have the diagnostic test available at the same time the drug comes to market,” so the need is for more early co-development, Dr. Humphries says, noting the two are regulated by different FDA branches. “That’s a complexity as well.”
Even if the barriers come down, there’s the reluctance to change clinical practice. “What we see, especially for more critically ill patients, is it’s a little harder for people to move from a broad-spectrum to a narrow therapy when the patient hasn’t yet turned the corner. It’s one of the things we need to figure out as a medical society,” she says—“how do we adapt our practices to optimize how we might be able to use these technologies.”
For now, rapid phenotypic antimicrobial susceptibility testing complements molecular diagnostics by providing definitive susceptibility data, Dr. Humphries and Vanderbilt colleague Janiece Glover, PhD, MS, wrote in their review of such platforms, published this year (Glover JS, et al. Clin Chem. 2026;72[7]:745–755). The platforms vary widely in methodology, they write, from real-time microscopy and optical detection to microfluidic culture systems and flow cytometry.
But they note the barriers to widespread clinical adoption, among them the need for direct integration with the electronic health record, the challenge of validation (especially for off-label uses or nonstandard organisms) in the absence of harmonized breakpoints across systems and regions, and high initial capital investment and recurring operational costs.
Once rapid testing is in place, Dr. Humphries says, the next step is to explore opportunities for other types of samples that may be seen more frequently in the laboratory than positive blood cultures “but are a little trickier to manage from a testing perspective.”
There’s a lot of work underway and movement in industry, she says, so laboratories need to be aware of what’s in the field and soon to come and develop a plan for what would work best in their setting.
Skill sets today will suit the technologies of tomorrow, she says. “A test is a test is a test. And if it’s a well-developed test, it should be functional in a variety of lab settings.”
While movement in a field is often slower than one might hope, she says, “it does move forward in the right direction.” She and Dr. Glover write in their review article that “the integration of AI, hybrid diagnostics, and full laboratory automation will likely define the next era of AST.”
In the AMP session on quality monitoring in molecular microbiology laboratories, Dr. Wojewoda and co-speaker Dr. Truong will discuss the range of activities that a laboratory can do to monitor quality.
“Different laboratories can approach it in different ways, and they’re trying to determine what the right fit is for their lab and the testing they do,” says Dr. Wojewoda, professor of pathology and laboratory medicine, University of Vermont Larner College of Medicine.
In the early years of molecular testing, she notes, with fewer closed systems, more quality monitoring was needed because the contamination risk was higher. Today, with more closed systems, the job is to figure out what’s best for quality, “but also what doesn’t break the bank in your laboratory. You can always do more, but that always costs more as well.” Routine quality monitoring is likely sufficient with FDA-cleared assays and closed systems, “but you’re still going to want to keep your eyes open for odd trends,” she says.
It’s a balance between doing too much, with its higher cost, and doing too little, she says, “and putting out incorrect patient results and not catching it until later.”
At the University of Vermont Medical Center, where she is director of clinical microbiology, the laboratory was doing wipe testing for one of its closed-system assays. None of the wipes were testing positive, so the laboratory stopped its wipe testing and decided to do weekly negative controls. “That saved us money and still will detect issues with that assay,” she says, noting another laboratory with a higher test volume or one with an open system runs a higher contamination risk and thus would need more monitoring.
The first step for any laboratory that wants to take a fresh look at its quality monitoring, she says, would be to consider what tests it’s doing and the risks for contamination and review all prior quality monitoring data. “Have they caught any issues? Has there ever been a problem with contamination for that assay? If not, maybe they can pull back on some of the monitors they’re performing.” If the laboratory has uncovered a problem in its prior monitoring, “they might want to scrutinize that test a little more carefully or more frequently to make sure they’re catching problems early.”
“It’s more of a risk assessment in terms of the assay,” she says, and, in looking at the prior data, whether there have been problems in the past.
Most pressing at this time, in Dr. Wojewoda’s view, are two issues: point-of-care testing and patient self-collection, both of which laboratories will see more of.
Molecular microbiology is only going to expand, and become more simplified, “and we will see it more in the point-of-care setting,” she says. “Laboratories will have to determine how comfortable they are with molecular microbiology at the point of care and how they can put a solid quality monitoring plan in place for it.”
Many of the point-of-care assays are easy to use and robust, “but there’s so many more of them, and because they’re so easy to use, they’re used in generalist labs. They’re kind of performed all over,” she says.
Educating the professionals who do the point-of-care testing is one job for the laboratory—what could go wrong, what to look for. The other is building that quality monitoring program to the laboratory’s size and scope and “setting the flags to make it easy to show there’s an issue.”
Self-collection of patient samples is becoming more common, and ensuring preanalytic quality will be important, “with a solid internal control [needed] for those samples,” she says.
In their presentation next month, Drs. Wojewoda and Truong will speak broadly in terms of CAP, New York State, and ISO requirements.
As with all things in health care, financial constraints are top of mind. “Because of financial struggles, we’ve looked at some of this and pared down what we’re doing,” she says of her institution, “so I will discuss our decision, how we made that decision, and the data we reviewed to get there.”
She expects the session she and Dr. Truong will lead to be broadly applicable: “Most people will be able to take the examples we give and apply them in their laboratories.”
Valerie Neff Newitt is a writer in Audubon, Pa.