Patients in the US frequently receive antibiotics that cannot kill the pathogens infecting them. Due to resource constraints, clinicians often prescribe empirically, rely on diagnostic tools that cannot reliably identify a broad gamut of pathogens, or fail to identify acquired resistance. This is often framed as a technological problem: how can new diagnostics improve on existing tools? However, the economic problems facing novel diagnostics are substantial and merit analysis. Existing diagnostics are cheap and cover most infections. Novel diagnostics can provide net savings when used appropriately, but siloed budgeting practices mean that the costs and the benefits land in different places. This mismatch stalls the deployment of novel diagnostics that would reduce healthcare costs and improve patient outcomes.
In this whitepaper, we outline the scale of discordant therapy and antibiotic resistance; the landscape of existing microbial diagnostics; how current detection technologies compare; who bears the cost when a patient's first antibiotic does not work; and where novel diagnostics, including those using AI, can create clinical impact.
Antimicrobial resistance (AMR) is a common problem that costs US health care $2.2–4.6 billion a year [1, 2]. AMR is technologically difficult to detect and can compound limitations of common microbial testing frameworks.
Discordant therapy is a broader problem than acquired AMR alone. About one in five US patients with a bloodstream infection are initially given a drug that has no effect on their pathogen, which is associated with 46% higher odds of death [3].
Hospital budgets often silo the costs of tests to a clinical lab while benefits are realized by pharmacies, wards, and insurers. Realizing the net impact of more expensive diagnostic tests requires cross-departmental coordination.
Who absorbs the cost of discordant therapy depends on the setting and the insurer. Typical Medicare and Medicaid programs use fixed payments, where the hospital absorbs excess costs. Under private insurance and in outpatient care, the insurer and the patient do.
Conventional culture is cheap and ubiquitous but has blind spots. Conventional clinical culture can grow a wide variety of pathogens but often fails to detect anaerobic bacteria, fungi, and fastidious pathogens.
Novel diagnostics have a high sticker price but can result in net savings. An expensive test can pay for itself when it reduces patient stays and avoidable transmission. At one hospital, genomic surveillance reported a modeled 3.2x return on investment [4].
Discordant therapy and acquired drug resistance
Antibiotic treatment usually begins before the cause of an infection is known. Laboratory results take days, and for a serious infection a delay in treatment is dangerous, so clinicians prescribe empirically: they choose a drug based on the likely pathogens and revise the choice once results arrive. In most cases the empiric choice works. When it does not, the patient spends the intervening days on a drug that is not treating the infection.
Resistance rates are the standard measure of a hospital's exposure to this risk. However, they describe only part of it, because acquired resistance is one of four reasons an antibiotic can fail.
Acquired resistance is what most people mean by AMR: a pathogen from a normally susceptible species stops responding to a drug because it has gained a mutation or a resistance gene. Surveillance systems are built to count these infections.
The clinical problem is broader: whether the antibiotic prescribed to the patient is effective against their infection. Treatment that does not match the pathogen is called discordant therapy; treatment that does is concordant.
Discordant therapy is antibiotic treatment that is not active against the pathogen causing the patient's infection. It is most often the first antibiotic a patient receives, chosen before test results are available, but any mismatch between drug and pathogen at any point in treatment is discordant. Acquired resistance is one cause; another is a drug chosen empirically that ends up being wrong.
Discordant therapy has four causes:
A fungal or viral infection, or a non-infectious inflammatory condition, treated with an antibacterial
A species that is never susceptible to the drug chosen
The pathogen is mistaken for a different species, often because it does not grow in culture or is part of a mixed infection, and the drug chosen does not act on it
A pathogen from a normally susceptible species gains a mutation or resistance gene and stops responding to the drug
Of note, three of the four causes of discordant therapy are resolved simply by knowing what pathogen is infecting the patient. Identifying the pathogen and testing it for resistance are components of the same diagnostic problem, and a hospital that monitors only resistance rates is measuring a fraction of the risk it carries.
The clinical and economic burden of discordant therapy
More than 2.8 million antibiotic-resistant infections occur in the US each year, causing more than 35,000 deaths [5] and adding an estimated $2.2–4.6 billion a year to health care costs [1, 2]. Discordant therapy is broader still: about one in five patients with a bloodstream infection starts on a drug to which their pathogen is not susceptible [3]. The remainder of this section describes which infections make up these totals, what the official counts leave out, what a single case costs, and whether better detection has been shown to reduce the burden.
Rates of drug-resistant infection
The US Centers for Disease Control and Prevention (CDC) estimates more than 2.8 million antibiotic-resistant infections and more than 35,000 deaths in the US each year, or 48,000 deaths when Clostridioides difficile is included [5]. These estimates rest mostly on 2017 data. CDC published a partial update in 2024 and has said it will release new estimates for at least 19 threats in 2026; until then, these remain the national figures.
Which infections matter most depends on the measure used:
By case count, the three leading infections are acquired in the community. By deaths, the list centers on the hospital.
In a CDC analysis of 890 US hospitals, 83% of infections with multidrug-resistant pathogens in hospitalized patients were community-onset, meaning they were present at or soon after admission [6]. A testing strategy aimed only at hospital-acquired infection therefore addresses a minority of cases.
The long tail of resistant infection
CDC's counts are concentrated in a small number of pathogens: by our arithmetic on its figures, the six most common threats account for over 90% of the 2.8 million cases [5]. However, a substantial amount of resistance sits outside this head, in four places.
Resistance that is not classified as a threat. CDC's total is, by design, a count of named threat categories: combinations of pathogen and drug that it ranks as urgent, serious or concerning, generally because few treatment options remain. Resistance to a first-line drug for which effective alternatives exist is not counted, however common it is. Ordinary urinary E. coli is the clearest example: among 1.5 million urinary E. coli isolates (pathogens grown from patient samples) from US female outpatients, 25% were non-susceptible to trimethoprim-sulfamethoxazole and 21% to fluoroquinolones [7]. An alternative drug exists for each of these infections, which is why they are not ranked as threats. Nonetheless, because the first prescription is often written without a culture, this uncounted resistance is a direct source of discordant therapy.
Pathogens that routine culture does not find. A resistance count can only include pathogens that were grown. In one study of 150 urogynecology patients, standard urine culture missed 67% of the uropathogens that an expanded culture method detected, and 50% in the women who reported urinary symptoms [8].
Resistance spread thinly across many pathogen–drug pairs. The largest global analysis modeled 88 pathogen–drug combinations across 23 pathogens. Six pathogens accounted for 73% of the deaths attributable to resistance in 2019 (929,000 of 1.27 million); the remaining quarter was distributed across the other 17 [9].
Unusual mechanisms within common pairs. Even for well-characterized pathogens, a share of resistant isolates carry rare or combined mechanisms that gene-based tests do not detect. We return to this in section 4.
of outpatient urinary E. coli non-susceptible to trimethoprim-sulfamethoxazole [7]
non-susceptible to fluoroquinolones [7]. CDC's 2.8 million does not count resistance to first-line drugs like these on its own [5].
of uropathogens missed by standard urine culture in one study [8]
Impact of discordant therapy
In a study of 21,608 patients with bloodstream infections at 131 US hospitals, 19% received initial therapy to which their pathogen was not susceptible, and discordant therapy was associated with 1.46 times the odds of death [3]. A second study found that appropriate therapy was delayed in 46% of resistant infections, and also in 33% of infections with susceptible pathogens [10]. Both studies are observational, so they show association rather than proof of cause. Nonetheless, taken together they suggest that the wrong first antibiotic is a routine event in hospital care, and not one confined to resistant pathogens.
of 21,608 patients with bloodstream infections at 131 US hospitals received discordant initial therapy [3]
the odds of death associated with discordant therapy [3]
Biotia's own data point in the same direction for urinary infection. In a retrospective study with SUNY Downstate, sequencing detected urogenital pathogens in 61.5% of culture-negative or mixed-culture specimens from symptomatic patients, and 85% of the antimicrobials prescribed empirically were ineffective against the pathogens detected, through off-target activity, partial coverage, or resistance [11]. These results come from a single center and have been reported as a conference abstract.
of culture-negative or mixed-culture specimens from symptomatic patients had urogenital pathogens detected by sequencing [11]
of empirically prescribed antimicrobials were ineffective (off-target, partial coverage, or resistance) [11]
The cost of using the wrong drug
The two best-documented national estimates measure different things by different methods.
a year [1]. What resistance adds to the cost of treating a bacterial infection (2014). Measured from national expenditure data.
a year [2]. Inpatient care for six multidrug-resistant pathogens, compared with uninfected patients (2017 dollars). Measured from a large hospital cohort.
In our view, the first figure is the most relevant to a diagnostic decision, because it isolates what resistance adds to an infection that would have been treated anyway.
Per case, a hospital-onset invasive infection adds the following hospital cost compared with an uninfected patient, in Veterans Affairs hospital data [2]:
These per-case figures, rather than the national totals, are the ones against which the cost of a diagnostic test should be weighed.
What the evidence says about paying for better detection
If the cost of discordant therapy is real, the next question is whether better detection demonstrably reduces it. The evidence supports paying for better detection where a team acts on the result and where testing is targeted. It does not show that speed alone improves outcomes.
Speed alone has not been shown to save lives. Two US randomized trials of rapid diagnostics in bloodstream infection found that antibiotics were changed sooner, but found no difference in mortality or length of stay; neither trial was powered to detect such differences [12, 13]. A meta-analysis of 31 studies found lower mortality with rapid testing, and in its subgroup analysis the difference was statistically significant only where a stewardship program acted on results [14]. This suggests that a faster test is better understood as an investment in a workflow than as an investment in an instrument.
Surveillance sequencing has paid for itself in at least one hospital. At UPMC Presbyterian, weekly whole-genome sequencing of healthcare-associated isolates over two years identified 172 outbreaks of 2 to 16 patients. The program estimated 62 infections and 4.8 deaths avoided, net savings of $695,706, and a 3.2-fold return on investment [4]. This is a single-hospital study and its savings are modeled from infections avoided. Nonetheless, it demonstrates that genomic data can function as operational infrastructure for infection control, and not only as a research tool.
Modeled return on investment
outbreaks identified over two years
infections and 4.8 deaths estimated avoided
net savings, modeled
Broad testing without targeting has had low yield. A review of 1,000 plasma metagenomic tests (the Karius test) at one large academic center judged that results had a positive clinical impact in 16% of cases, a negative impact in 2% and none in 82% [15]. Impact was higher for specific indications such as culture-negative endocarditis, which suggests that the value of metagenomic testing depends on where it is applied.
These results indicate that the limiting factor has not been the technology itself, but the alignment between a test, the team that acts on it, and the budget that pays for it.
Who pays when the first antibiotic is wrong
The cost of discordant therapy is real, but it does not fall where the testing decision is made. Who absorbs it depends on the care setting and the insurer, and within the hospital, the department that pays for a test is usually not the one that sees its benefit.
Medicare, Medicaid, insurance, inpatient and outpatient
| Setting and insurer | How care is paid | If the first drug is discordant | Who absorbs the cost |
|---|---|---|---|
| Inpatient, traditional Medicare and most state Medicaid programs | One fixed payment per admission, set by diagnosis | Payment does not change unless costs pass an outlier threshold | The hospital |
| Inpatient, private insurance | Often per day, per case or as a percentage of charges | Payment rises with each extra day and drug | Largely the insurer |
| Outpatient, any insurer | Each visit, test and prescription billed separately | Each extra visit, drug or admission is billed | The insurer, and the patient's share each time |
For Medicare, the outlier threshold in fiscal 2027 is about $49,000 above the standard payment; Medicare then pays 80% of costs beyond it [16, 17]. Medicaid payment rules are set by each state. Private contracts vary, and the description above is a generalization.
Value and payment sit in different places
Inside the hospital, each department sees part of the picture:
| When | Who | Sees | Does not see |
|---|---|---|---|
| Day 0 | Laboratory | The cost of each test and, for outpatients, the payer's reimbursement for it | What a right or wrong answer costs downstream |
| Days 1–5 | Pharmacy and antimicrobial stewardship | The reserve antibiotic (a newer drug held back for resistant infections), at $8,000–19,000 a course for the drug alone (7–14 days at $1,200–1,400 a day) | The test that could have avoided or confirmed it |
| Day 1 onward | Wards and finance | Extra bed-days and blocked beds | Which test result caused them |
| At discharge | Payer | A fixed payment, or the extra days | Which department spent or saved |
The laboratory's own accounts sharpen the problem. For hospitalized patients a test is pure cost, because it is bundled into the fixed payment. For outpatients the laboratory bills the payer and sees that reimbursement directly, and where a hospital expects its laboratory to break even or contribute a margin, each test is judged on whether payment covers its cost. A novel test that payers do not yet cover, or cover below cost, is therefore an uphill sell inside the laboratory even when it would save the hospital money elsewhere.
“A test is therefore usually bought on its cost, because its value appears in someone else's budget.”
This is important because it means the economic case for a better test cannot be made, or even observed, from within the laboratory alone; seeing the value requires analysis across departments.
Three illustrative scenarios
The scenarios below show how the payment rules play out for individual patients. The dollar amounts and lengths of stay in the first two are illustrative, not drawn from a study.
| Concordant first drug | Discordant first drug | |
|---|---|---|
| What happens | Stay of about 5–7 days | Stay of about 8–12 days, possibly a second operation |
| Payment | One fixed Medicare payment | The same payment, unless the outlier threshold is passed |
| Who pays the difference | — | The hospital, at up to about $3,300 per extra day (the 2024 US average expense per inpatient day) [18] |
| Patient outcome | Recovery on the expected course | Longer stay, more post-acute care |
| Concordant first drug | Discordant first drug | |
|---|---|---|
| What happens | One visit, a culture, a generic oral drug | Emergency visit days later, then 3–4 days in hospital on an intravenous drug |
| Payment | A few hundred dollars | About $10,000–25,000 |
| Who pays the difference | — | The insurer, and the patient up to the out-of-pocket limit |
| Patient outcome | Symptoms resolve in days | Days untreated, risk of kidney or bloodstream infection |
| Concordant first drug | Discordant first drug | |
|---|---|---|
| What happens | Appropriate therapy on day 0: median infection-related stay of 8 days | Appropriate therapy after 1–4 days: 9–10 days; after 5 or more days: 13 days [19] |
| Payment | One fixed payment for publicly insured patients; per day or per case for privately insured ones | The same fixed payment for publicly insured patients, unless the outlier threshold is passed; rises with each extra day for privately insured ones |
| Who pays the difference | — | The hospital for publicly insured patients; the insurer for privately insured ones |
| Patient outcome | Recovery on the expected course | Longer stay, with more days on a drug that is not treating the infection |
Rapid identification of the resistance mechanism may also affect survival. In a separate observational study of 137 patients with carbapenem-resistant Enterobacterales bloodstream infection at eight New York and New Jersey hospitals, 30-day mortality was 24% where the laboratory ran a rapid test for a common carbapenemase gene (KPC) on positive blood cultures and 47% where it did not [20]. The study was not randomized, and we cannot rule out that hospitals running the test differed in other ways.
The insurer also changes the hospital's financial result. In a 78-hospital, Merck-funded study of 1,488 patients with carbapenem-nonsusceptible Gram-negative respiratory infections, hospitals broke even on average (a mean loss of $139 per patient). However, the modeled result was a loss of about $14,000 per Medicare case and $48,000 per Medicaid case, against a gain of about $15,000 per case for private and other payers [21].
What a test is paid
Inpatient tests are bundled into the fixed payment and earn the laboratory nothing extra. For outpatients, Medicare pays by the test [22]:
These rates bear little relation to the value of the answer. Of note, a rate also does not guarantee coverage: some payers restrict large panels and metagenomic tests.
Reserve antibiotics put a price on a laboratory result
Reserve antibiotics are last-resort drugs. The World Health Organization's classification defines its Reserve group as antibiotics that should be kept for confirmed or suspected infections caused by multidrug-resistant pathogens [23]. Hospitals restrict them for two reasons: using them sparingly slows the emergence of resistance to the few options that remain, and they are expensive.
In US hospitals they are used mainly for infections resistant to carbapenems, a class of broad-spectrum antibiotics that are themselves kept for serious hospital infections, where standard drugs no longer work. A prescriber typically cannot order one directly; the antimicrobial stewardship team releases it, usually once laboratory results confirm that it is needed and indicate which one will work or patients fail to improve on other drugs. The choice among them depends on the resistance mechanism the laboratory reports [24], and the mix of mechanisms is shifting: one family of carbapenemase (an enzyme that breaks down carbapenems), NDM, rose from 5% to 40% of carbapenemase-producing Enterobacterales in CDC surveillance between 2016 and 2023 [25].
The mix is shifting: NDM rose from 5% to 40% of carbapenemase-producing Enterobacterales in CDC surveillance between 2016 and 2023 [25].
At one health system the newer reserve agents cost about $1,200–1,400 a day in 2026, against about $10 for meropenem, a standard carbapenem, so a course runs about $8,400–9,600 for 7 days and $17,000–19,000 for 14 days [26]. Wholesale prices in 2019 put a 14-day course at $13,000–15,000 [26]. Under fixed payment the hospital absorbs that cost. A laboratory result that confirms or rules out the need for a reserve antibiotic is therefore one of the few with an obvious dollar value attached.
The two programs that act on results
Two programs act on laboratory results: infection prevention and control, and antimicrobial stewardship. Both are required of hospitals that participate in Medicare; the stewardship requirement took effect in March 2020 [27]. Infection control handles surveillance, isolation and outbreak investigation, with a median of one preventionist per 121 beds [28]. Stewardship approves restricted drugs and reviews prescriptions against laboratory results. The regulations require that both programs exist but leave their design to the hospital, so their staffing, authority and day-to-day practice vary widely from one health system to another.
Neither program owns the laboratory budget, and the financial penalties attached to them are narrow: among resistant pathogens, Medicare's hospital-acquired condition penalties cover only MRSA bloodstream infection and C. difficile [29]. The groups best placed to act on a diagnostic result are therefore not the ones that purchase it, and most resistant pathogens carry no direct financial penalty that would prompt a purchase.
What this has meant for diagnostic vendors
The same mismatch has shaped the diagnostics market. Between 2023 and 2026, two of the best-known rapid-detection companies failed: Accelerate Diagnostics filed for bankruptcy in May 2025 and its assets were later sold to Bruker; T2 Biosystems announced liquidation in February 2026, having reported $7.2 million in 2023 revenue against a $584 million accumulated deficit [30]. Two prominent genomic-prediction companies, Ares Genetics and Day Zero Diagnostics, ended as asset sales to bioMérieux [30]. For the two rapid-detection companies, a common thread is a product priced in the hundreds of dollars competing with an $8.65 test inside a bundled payment.
| Company | Category | Outcome |
|---|---|---|
| Accelerate Diagnostics | Rapid detection | Chapter 11, May 2025; assets later sold to Bruker |
| T2 Biosystems | Rapid detection | Liquidation announced February 2026 |
| Ares Genetics | Genomic prediction | Asset sale to bioMérieux, April 2024 |
| Day Zero Diagnostics | Genomic prediction | Asset sale to bioMérieux, June 2025 |
The technological landscape
Hospital laboratories layer five kinds of detection technology. Although each newer method was introduced to overcome a limitation of culture, none has replaced it; each is added on top. This section describes the existing technologies, the gaps they leave, and the emerging approaches, many of them AI-based, that are aimed at those gaps.
Existing detection technologies
Two terms in the table need explanation. Phenotypic susceptibility testing exposes the pathogen to a drug and measures directly whether it still grows. Metagenomic sequencing reads all the genetic material in a sample without first growing the pathogen; it is usually a send-out, meaning the sample goes to an outside reference laboratory.
Turnaround times run from sample collection and include the time the pathogen needs to grow. They are the authors' estimates (time for the culture to turn positive plus each method's run time), except for sequencing; a study of 13 US hospitals measured a median of 2.7 days from blood collection to a susceptibility result [32], and conventional susceptibility testing is commonly put at 48–96 hours after a blood culture turns positive [33].
| Technology | Sample to result | Cost | Identifies the pathogen | Detects resistance | Regulatory status |
|---|---|---|---|---|---|
| Culture and phenotypic susceptibility testing | 2–5 days | Medicare pays $8.65 [22] | Yes, after growth for species grown in clinical culture | Yes, measured directly | Cleared by the US Food and Drug Administration (FDA) |
| Mass spectrometry identification | 1–2 days | Medicare pays $8.08 [22] | Yes, minutes after growth | No | FDA cleared for identification |
| PCR panels and lateral-flow tests | About 1 day | Medicare pays $218 [22] | Only pathogens on the panel | Only known genes on the panel | FDA cleared |
| Rapid phenotypic testing | 1–1.5 days | About $250 per test (2018 list price) | Limited panel | Yes, measured directly | Several more cleared since 2024 |
| Sequencing | 2–5 days | Typically $200–300 for Biotia's send-out test | Any pathogen sequenced | Genes, and predicted susceptibility | LDT or RUO |
Three patterns matter for a purchasing decision:
1. Speed trades against breadth. Mass spectrometry names the pathogen quickly but says nothing about resistance, whereas a PCR panel finds only the pathogens and genes it was designed to find.
2. Most methods still wait for growth. For bloodstream infection, every routine method except direct-from-blood tests and metagenomic sequencing starts from a positive culture.
3. Sequencing sits outside FDA clearance for susceptibility. It is offered as a laboratory-developed test, which shifts the burden of validation to the laboratory running it.
The practical consequence is that no single technology currently provides a fast, broad and inexpensive answer, and a laboratory's testing strategy is necessarily a combination of methods chosen for particular clinical situations.
Where existing technology falls short
Existing tests handle resistance caused by a single gene well. However, they fall short when resistance has several combined causes, when nothing grows in culture, and when the pathogen is identified late or not at all.
Single-gene resistance is largely solved. MRSA, vancomycin-resistant Enterococcus, the common carbapenemases and rifampicin-resistant tuberculosis are each explained by one gene or a small set of mutations, and PCR detects them reliably.
Combinatorial resistance is not. Of 847 carbapenem-resistant Pseudomonas aeruginosa isolates tested in CDC surveillance, six carried a carbapenemase gene [34]. In the rest, resistance arises from combinations of changes, such as loss of the channel through which the drug enters the cell, and a gene panel has nothing to report.
carbapenem-resistant Pseudomonas isolates carried a carbapenemase gene that a panel would detect [34]
Many infections never produce a culture result. About 47% of severe sepsis admissions are culture-negative [35], and no pathogen is identified in 62% of adults hospitalized with community-acquired pneumonia [36].
Some pathogens are not routinely sought or tested. Anaerobic bacteria, which grow only without oxygen, are the clearest case. Standard urine culture is incubated in air, so anaerobes in urine are not looked for at all [8]. Where anaerobes are grown, they are rarely tested for susceptibility: in a 2008 survey, only 21% of hospital laboratories performed anaerobe susceptibility testing in-house [37]. We did not find a more recent US survey.
of severe sepsis admissions are culture-negative [35]
of adults hospitalized with community-acquired pneumonia have no pathogen identified [36]
of hospital laboratories tested anaerobe susceptibility in-house (2008 survey) [37]
Identification itself arrives late or not at all. Culture needs growth, so often the first antibiotics are chosen empirically before the pathogen is known. Faster or broader alternatives exist, but they cost the laboratory far more to run: Medicare's outpatient rates, a rough proxy for cost, are 25 to 250 times the rate of a culture for tests that work directly from the sample or the blood-culture bottle [22]. For hospitalized patients, where most serious infections are treated, the laboratory receives no separate payment for any test (section 3), so that higher cost comes out of its own budget. These tests are also not available in every laboratory, and add transport time when sent out.
These gaps share a structure: in each, the information needed to choose the right drug is either not captured by current tests or arrives after the initial prescribing decision has been made.
Emerging approaches
AI is being applied to each of these gaps. The central challenge is the same in every case: converting incomplete or indirect evidence into a prediction that can support a treatment decision.
Predicting susceptibility from the genome. Machine-learning models take a sequenced genome and predict, drug by drug, whether the pathogen is susceptible. In principle, they are strongest where resistance is combinatorial and gene panels fail. As of October 2026, no sequencing-based susceptibility test has FDA clearance, to our knowledge, and the public data available to train such models is limited: the largest public collection holds 170,750 isolates, three species make up over half of it, and resistant isolates are over-represented [38].
Reading the patient instead of the pathogen. Host-response tests measure the patient's immune signals from a blood draw to judge whether an infection is bacterial at all and how severe it is. This addresses the first cause of discordant therapy (the wrong class of pathogen) and the culture-negative case. At least six such tests have been cleared or authorized by FDA since 2021, and one, TriVerity, also received a Medicare add-on payment for fiscal 2026 [39].
Predicting from the health record. Models trained on a patient's history and local resistance patterns can recommend an antibiotic before any test result exists. In retrospective studies, past cultures and antibiotic purchases predicted resistance in 700,000 community urinary infections [40]; an algorithm reduced second-line antibiotic use by 67% relative to clinicians while reducing inappropriate therapy by 18% [41]; and "personalized antibiograms" (patient-specific versions of a hospital's table of local susceptibility rates) matched clinicians' coverage with narrower drugs [42]. However, we are not aware of a prospective randomized evaluation of any of them.
How AI can be incorporated into clinical tools
There are two ways an AI tool can reach clinical use and be paid for.
As a standalone product. The algorithm is itself the product: software that takes existing data, such as the health record, and returns a prediction, with its own regulatory clearance and its own payment. This route is rare and new. The first example in this field arrived in October 2026, when Medicare began a temporary add-on payment of up to $61.84 per inpatient case for a software-only sepsis prediction tool [17].
Maximum temporary Medicare add-on payment per inpatient case for a software-only sepsis prediction tool, beginning October 2026 [17]. The first separate payment for an algorithm in this field.
As product-plus. The algorithm is built into something that is already bought and paid for: a test with a billing code, an instrument, or a laboratory workflow where it saves labor. Until this year, every AI tool in this field was paid for this way, and most still are. Genome-based susceptibility prediction and host-response scoring both follow this route, because the algorithm has no output without the assay that feeds it.
In either case, a hospital should evaluate AI on the same evidence as any other test or workflow.
AI-enabled results are most likely to earn their cost where they trigger a decision that someone already owns:
| Decision | Owner | What is at stake |
|---|---|---|
| Releasing a reserve antibiotic | Stewardship | An $8,000–19,000 course |
| Lifting isolation | Infection control | Bed-days and staff time |
| Confirming or ruling out an outbreak | Infection control | Further transmission, and the cost of investigating |
The practical importance of this framing is that it moves the question from whether an algorithm is accurate to whether its output changes a decision that already carries a cost and an accountable owner.
Implications for hospital leaders and laboratory directors
The following points are the authors' reading of the evidence above. We propose them as a framework for evaluating diagnostic investments, not as findings in themselves.
The share of patients whose first antibiotic did not match their pathogen is a more complete measure of exposure than a resistance rate, and it can be computed from routine pharmacy and microbiology records [3].
A test's cost falls on the laboratory and its benefit on pharmacy, the wards and, depending on the payer, the hospital's margin. A business case built inside one department will understate the value.
The financial case for faster, better-targeted therapy is strongest where payment is fixed, which for most hospitals means traditional Medicare and most Medicaid inpatients.
In randomized trials, faster results changed antibiotics sooner but did not change mortality or length of stay, and the trials were not powered to detect such differences. In pooled studies, a significant mortality benefit was seen only where a stewardship program acted on results.
Sequencing and metagenomic tests have shown value in outbreak surveillance and in specific culture-negative syndromes, and are the logical candidates where gene panels fail. Untargeted use has shown low yield.
For any new test, AI-enabled or not, the useful questions are which decision it informs, who owns that decision, and what that decision costs today.
More broadly, these points suggest that AMR detection should not be treated as an isolated laboratory function, but as a component of a hospital's wider infrastructure for prescribing, infection control and financial planning. When diagnostic results are coupled to the teams and budgets that depend on them, their value becomes both visible and measurable.
References
- Thorpe KE, Joski P, Johnston KJ. Antibiotic-resistant infection treatment costs have doubled since 2002, now exceeding $2 billion annually. Health Aff (Millwood) 2018;37(4):662–669. doi:10.1377/hlthaff.2017.1153
- Nelson RE, Hatfield KM, Wolford H, et al. National estimates of healthcare costs associated with multidrug-resistant bacterial infections among hospitalized patients in the United States. Clin Infect Dis 2021;72(Suppl 1):S17–S26. doi:10.1093/cid/ciaa1581
- Kadri SS, Lai YL, Warner S, et al. Inappropriate empirical antibiotic therapy for bloodstream infections based on discordant in-vitro susceptibilities: a retrospective cohort analysis of prevalence, predictors, and mortality risk in US hospitals. Lancet Infect Dis 2021;21(2):241–251. doi:10.1016/S1473-3099(20)30477-1
- Sundermann AJ, Kumar P, Griffith MP, et al. Real-time genomic surveillance for enhanced healthcare outbreak detection and control: clinical and economic impact. Clin Infect Dis 2026;82(1):134–141. doi:10.1093/cid/ciaf216
- Centers for Disease Control and Prevention. Antibiotic Resistance Threats in the United States, 2019. Atlanta: US Department of Health and Human Services, CDC; 2019. doi:10.15620/cdc:82532
- Jernigan JA, Hatfield KM, Wolford H, et al. Multidrug-resistant bacterial infections in U.S. hospitalized patients, 2012–2017. N Engl J Med 2020;382(14):1309–1319. doi:10.1056/NEJMoa1914433
- Kaye KS, Gupta V, Mulgirigama A, et al. Antimicrobial resistance trends in urine Escherichia coli isolates from adult and adolescent females in the United States from 2011 to 2019: rising ESBL strains and impact on patient management. Clin Infect Dis 2021;73(11):1992–1999. doi:10.1093/cid/ciab560
- Price TK, Dune T, Hilt EE, et al. The clinical urine culture: enhanced techniques improve detection of clinically relevant microorganisms. J Clin Microbiol 2016;54(5):1216–1222. doi:10.1128/JCM.00044-16
- Antimicrobial Resistance Collaborators (Murray CJL, Ikuta KS, Sharara F, et al.). Global burden of bacterial antimicrobial resistance in 2019: a systematic analysis. Lancet 2022;399(10325):629–655. doi:10.1016/S0140-6736(21)02724-0
- Bonine NG, Berger A, Altincatal A, et al. Impact of delayed appropriate antibiotic therapy on patient outcomes by antibiotic resistance status from serious Gram-negative bacterial infections. Am J Med Sci 2019;357(2):103–110. doi:10.1016/j.amjms.2018.11.009
- Morfa Romero EJ, Augenbraun M, Rivera T, et al. 2835. Utilization of clinical-grade metagenomics in urinary tract infection diagnostics: improving high-risk patients' outcomes with a genomic-based assay. Open Forum Infect Dis 2023;10(Suppl 2):ofad500.2445. doi:10.1093/ofid/ofad500.2445
- Banerjee R, Teng CB, Cunningham SA, et al. Randomized trial of rapid multiplex polymerase chain reaction–based blood culture identification and susceptibility testing. Clin Infect Dis 2015;61(7):1071–1080. doi:10.1093/cid/civ447
- Banerjee R, Komarow L, Virk A, et al. Randomized trial evaluating clinical impact of RAPid IDentification and Susceptibility testing for Gram-negative bacteremia: RAPIDS-GN. Clin Infect Dis 2021;73(1):e39–e46. doi:10.1093/cid/ciaa528
- Timbrook TT, Morton JB, McConeghy KW, Caffrey AR, Mylonakis E, LaPlante KL. The effect of molecular rapid diagnostic testing on clinical outcomes in bloodstream infections: a systematic review and meta-analysis. Clin Infect Dis 2017;64(1):15–23. doi:10.1093/cid/ciw649
- Kaur I, Shaw B, Multani A, et al. Real-world clinical impact of plasma cell-free DNA metagenomic next-generation sequencing assay. Infect Control Hosp Epidemiol 2025;46(5):504–511. doi:10.1017/ice.2024.242
- Medicare Payment Advisory Commission. Hospital Acute Inpatient Services Payment System. Payment Basics, revised November 2025. https://www.medpac.gov/document-type/payment-basic/ (accessed 9 October 2026).
- Centers for Medicare & Medicaid Services. Medicare Program; Hospital Inpatient Prospective Payment Systems for Acute Care Hospitals and the Long-Term Care Hospital Prospective Payment System and Policy Changes and Fiscal Year 2027 Rates; final rule. 91 FR 49570, 4 August 2026 (document 2026-15833).
- KFF State Health Facts. Hospital Adjusted Expenses per Inpatient Day (AHA Annual Survey, 1999–2024). https://www.kff.org/health-costs/state-indicator/expenses-per-inpatient-day/ (accessed 9 October 2026).
- Lodise TP, Kanakamedala H, Hsu WC, Cai B. Impact of incremental delays in appropriate therapy on the outcomes of hospitalized adult patients with Gram-negative bloodstream infections: "Every day matters". Pharmacotherapy 2020;40(9):889–901. doi:10.1002/phar.2446
- Satlin MJ, Chen L, Gomez-Simmonds A, et al. Impact of a rapid molecular test for Klebsiella pneumoniae carbapenemase and ceftazidime-avibactam use on outcomes after bacteremia caused by carbapenem-resistant Enterobacterales. Clin Infect Dis 2022;75(12):2066–2075. doi:10.1093/cid/ciac354
- McCann E, Sung AH, Ye G, Vankeepuram L, Tabak YP. Contributing factors to the clinical and economic burden of patients with laboratory-confirmed carbapenem-nonsusceptible Gram-negative respiratory infections. Infect Drug Resist 2020;13:761–771. doi:10.2147/IDR.S236026
- Centers for Medicare & Medicaid Services. Clinical Laboratory Fee Schedule, calendar year 2026, fourth-quarter file (codes 87077, 87186, 87154, 87633, 0152U). https://www.cms.gov/medicare/payment/fee-schedules/clinical-laboratory-fee-schedule-clfs/files (accessed 9 October 2026).
- World Health Organization. AWaRe classification of antibiotics for evaluation and monitoring of use, 2023. Geneva: WHO; 2023.
- Infectious Diseases Society of America. IDSA 2026 Guidance on the Treatment of Antimicrobial-Resistant Gram-Negative Infections. Published 30 July 2026. https://www.idsociety.org/practice-guideline/amr-guidance/ (accessed 9 October 2026).
- Ansari UA, Campbell D, Brandenburg JM, et al. Increase in blaNDM among carbapenemase-producing, carbapenem-resistant Enterobacterales, United States, 2016–2023. Emerg Infect Dis 2026;32(6):1028–1030. doi:10.3201/eid3206.251404
- (a) Allina Health Laboratories. Comparative Daily Costs for Common IV Antibiotics; costs as of February 2026. https://labs.allinahealth.org/Image/ViewDocument?uid=1f9f2afe-9f3f-442c-b2a7-26ad1bd7748b (accessed 9 October 2026). (b) Clancy CJ, Potoski BA, Buehrle D, Nguyen MH. Estimating the treatment of carbapenem-resistant Enterobacteriaceae infections in the United States using antibiotic prescription data. Open Forum Infect Dis 2019;6(8):ofz344. doi:10.1093/ofid/ofz344
- 42 CFR 482.42(b). Medicare and Medicaid Programs; Regulatory Provisions To Promote Program Efficiency, Transparency, and Burden Reduction; final rule. 84 FR 51732, 30 September 2019 (document 2019-20736).
- Bartles R, Reese S, Gumbar A. Closing the gap on infection prevention staffing recommendations: results from the beta version of the APIC staffing calculator. Am J Infect Control 2024;52(12):1345–1350. doi:10.1016/j.ajic.2024.09.004
- Centers for Medicare & Medicaid Services. Hospital-Acquired Condition Reduction Program. https://www.cms.gov/medicare/payment/prospective-payment-systems/acute-inpatient-pps/hospital-acquired-condition-reduction-program-hacrp (accessed 9 October 2026).
- Company filings and announcements: Accelerate Diagnostics, press release, 8 May 2025 (Chapter 11); Indaba Capital, press release, 24 November 2025 (sale to Bruker); T2 Biosystems, Form 10-K for fiscal year 2023 (revenue $7.2 million; accumulated deficit $584.3 million) and press release, 3 February 2026; OpGen, Form 10-K for fiscal year 2023, filed June 2024 (Ares Genetics insolvency, sale of assets to bioMérieux, April 2024); bioMérieux, press release, 16 June 2025 (Day Zero Diagnostics assets).
- Waggle KD, Griffith MP, Rokes AB, et al. Methods for cost-efficient, whole genome sequencing surveillance for enhanced detection of outbreaks in a hospital setting. PLoS One 2026;21(9):e0355550. doi:10.1371/journal.pone.0355550
- Tabak YP, Vankeepuram L, Ye G, Jeffers K, Gupta V, Murray PR. Blood culture turnaround time in U.S. acute care hospitals and implications for laboratory process optimization. J Clin Microbiol 2018;56(12):e00500-18. doi:10.1128/JCM.00500-18
- Banerjee R, Humphries R. Rapid antimicrobial susceptibility testing methods for blood cultures and their clinical impact. Front Med (Lausanne) 2021;8:635831. doi:10.3389/fmed.2021.635831
- Grass JE, Bulens SN, Bamberg WM, et al. 486. Epidemiology of carbapenem-resistant Pseudomonas aeruginosa identified through the Emerging Infections Program (EIP), United States, 2016–2018. Open Forum Infect Dis 2019;6(Suppl 2):S238 (conference abstract). doi:10.1093/ofid/ofz360.559
- Gupta S, Sakhuja A, Kumar G, McGrath E, Nanchal RS, Kashani KB. Culture-negative severe sepsis: nationwide trends and outcomes. Chest 2016;150(6):1251–1259. doi:10.1016/j.chest.2016.08.1460
- Jain S, Self WH, Wunderink RG, et al. Community-acquired pneumonia requiring hospitalization among U.S. adults. N Engl J Med 2015;373(5):415–427. doi:10.1056/NEJMoa1500245
- Brook I, Wexler HM, Goldstein EJC. Antianaerobic antimicrobials: spectrum and susceptibility testing. Clin Microbiol Rev 2013;26(3):526–546. doi:10.1128/CMR.00086-12
- Dickens E, Derelle R, Beardmore R, et al. A comprehensive AMR genotype-phenotype database (CABBAGE). Nucleic Acids Res 2026;54(16):gkag780. doi:10.1093/nar/gkag780
- US Food and Drug Administration, 510(k) and De Novo databases: K210254, K203748, K220991, K230917, DEN230036, K241676. Centers for Medicare & Medicaid Services. Fiscal Year 2026 Hospital Inpatient Prospective Payment System final rule. 90 FR 36536, 4 August 2025 (document 2025-14681).
- Yelin I, Snitser O, Novich G, et al. Personal clinical history predicts antibiotic resistance of urinary tract infections. Nat Med 2019;25(7):1143–1152. doi:10.1038/s41591-019-0503-6
- Kanjilal S, Oberst M, Boominathan S, Zhou H, Hooper DC, Sontag D. A decision algorithm to promote outpatient antimicrobial stewardship for uncomplicated urinary tract infection. Sci Transl Med 2020;12(568):eaay5067. doi:10.1126/scitranslmed.aay5067
- Corbin CK, Sung L, Chattopadhyay A, et al. Personalized antibiograms for machine learning driven antibiotic selection. Commun Med (Lond) 2022;2:38. doi:10.1038/s43856-022-00094-8
- Couto-Rodriguez M, Danko DC, Wells HL, et al. Analytical validation of a highly accurate and reliable next-generation sequencing-based urine assay. Microbiol Spectr 2026;14(6):e0202625. doi:10.1128/spectrum.02026-25