AI Against Superbugs

At 34, Nadia works in an emergency department where minutes compress everything. A person arrives with fever, confusion, low blood pressure, or a fast-changing story that doesn’t wait for a microbiology schedule. Even when clinicians suspect bloodstream infection, the first actionable data often comes from vital signs, basic labs, and rapid bedside tests, not from the definitive organism ID that culture will eventually provide.

On a typical night, Nadia might see a cluster of cues: a patient with chills and a rising lactate, another with a line in place, a third with urinary symptoms that don’t fully explain how sick they look. The clinical question becomes: which patients are likely to have a serious bacterial infection that needs targeted antibiotics now, and which ones can be monitored while slower tests catch up? Traditional workflows rely on a sequence - collect specimens, incubate, run identification and susceptibility panels, then adjust therapy. Those steps are valuable, but their pace can lag behind the patient’s trajectory.

That’s where AI triage changes the tempo. Instead of waiting for the final microbiology answer, a model like the RAPID Lens Triage Model is designed to match early signals - the measurable “shape” of what’s happening - to patterns associated with infection categories. The “real time” part is not a marketing phrase; it’s about using signals that arrive during the first hours of evaluation, then producing a triage output that can guide where the next tests, consults, and antibiotic decisions concentrate. The triage is still provisional, but it is informed earlier than a culture-based workflow alone would allow.

The cultural reason this matters is straightforward: hospitals are built around delays. Labs batch samples; incubators run on fixed cycles; susceptibility testing has throughput limits. The scientific reason is more sobering: antibiotic resistance evolves in populations, but clinical decisions are made for individuals under time pressure. AI sits at the intersection - trying to translate complex lab-derived signals into faster, probabilistic guidance.

How Faster Triaging Became a Necessity in Infection Care

Long before AI entered the picture, clinicians learned that waiting for confirmation can be costly. Antibiotics are not neutral; they’re powerful interventions with collateral effects, including pressure that selects for resistant strains. That creates a constant tension: treat early enough to prevent harm, but avoid over-treating when the cause might be viral, inflammatory, or otherwise non-bacterial.

The standard response to that tension was never “wait for the lab.” It was to build clinical suspicion rules - patterns of symptoms, exam findings, and basic labs - that help clinicians estimate likelihood. These rules are imperfect, but they’re fast. The trouble is that they don’t incorporate the rich information embedded in microbiology workflows themselves: the behavior of specimens over time, the way growth signals change, the way certain test readouts separate likely bacterial categories from noise.

Historically, microbiology became slower and more reliable at the same time. Incubation-based methods require time for organisms to multiply to detectable levels. Identification and susceptibility testing add more time because they depend on specific biochemical reactions or automated panel readouts. Even when newer platforms shorten turnaround, the gap between “specimen collected” and “actionable organism details” remains.

AI triage tries to compress that gap without pretending the world is deterministic. It does not magically remove incubation time. Instead, it asks: what can be inferred earlier from partial information? If a specimen’s signals evolve in ways that correlate with certain infection types, then a model can output a triage score while cultures are still in progress. In other words, AI becomes a translator for early evidence, not a replacement for the eventual lab confirmation.

This is where the RAPID Lens Triage Model framing matters. “Lens” is a useful metaphor because it implies a focused view: the model doesn’t attempt to answer everything at once. It prioritizes what’s needed for triage - what’s likely, what’s urgent, and where additional testing would be most informative. “RAPID” signals the intent: fast enough to influence the early clinical window, not just fast enough to look good on a retrospective benchmark.

One concrete measurement that captures the practical difference is turnaround time, often expressed as time from specimen collection to result availability. Traditional culture workflows are measured in days; triage models aim to make decisions in hours by using signals that are already available earlier in the diagnostic pipeline.

Pattern-Matching Signals: From Incubation Clues to a Triage Decision

To understand why pattern-matching can be faster than traditional lab workflows, it helps to notice what labs actually do. A lab doesn’t just “wait and then read.” It monitors. Instruments track whether and how signals change - growth detection systems register changes in metabolic activity, automated blood culture readers track time-to-detection, and rapid assays capture early characteristics of samples.

Those signals are often treated as inputs to later steps, but they also contain information about what kind of organism is likely present and how aggressively an infection might behave. The challenge is that the signals are messy: they’re influenced by specimen volume, collection quality, patient factors, and instrument-specific noise. Humans can learn some patterns, but they cannot continuously scan the entire signal space across thousands of cases in real time.

AI’s advantage is pattern detection at scale. The model learns correlations between early signal features and eventual outcomes - such as whether an infection is likely bacterial, likely to be bloodstream-related, or associated with higher-risk patterns. Then, during a live workflow, it applies those learned correlations to new incoming signals and outputs a triage result.

This is not “magic,” but it is different from typical rule-based approaches. A ruleset might say, in effect, “If X and Y and Z, then suspect infection.” Pattern-matching can represent subtler relationships: not just whether a signal exists, but how it changes over time; not just which test is positive, but how the readout behaves relative to typical distributions. A model can treat a growth curve’s slope or early fluctuations as evidence - something that can be hard to operationalize manually without losing nuance.

The RAPID Lens Triage Model concept emphasizes that triage requires a specific kind of evidence. Triage is not identification. It is prioritization under uncertainty. That means the model’s job is to place a specimen somewhere useful in the decision space - marking it as likely higher-risk or likely lower-risk for certain categories - so that clinicians can align attention and next tests with urgency.

In practice, this can look like a decision-support layer that sits alongside other early indicators Nadia and her colleagues already use. If a patient’s early signals and lab readouts suggest a high likelihood of serious bacterial infection, triage outputs can heighten urgency for targeted steps. If signals suggest a lower likelihood, it can support more measured escalation while cultures and confirmatory tests mature. Either way, the model’s utility hinges on the same principle: earlier pattern interpretation can shift where time and resources go.

There is also a procedural nuance that matters in real hospitals: triage systems must be resilient to the fact that not every sample behaves ideally. Specimens can be contaminated; antibiotics may have already been given; timing may vary. Traditional lab workflows handle uncertainty by waiting longer and verifying more. AI triage handles uncertainty by scoring probability and updating as new signals arrive. It’s a different philosophy of evidence management - one that accepts incompleteness as the starting point.

A Counterintuitive Twist: “Less Certain” Can Still Be “More Useful”

Here’s the surprise: the most clinically useful AI outputs are sometimes the ones that admit the most uncertainty. A triage score that is not confident enough to “diagnose” anything can still be valuable because triage is about ranking urgency, not declaring truth.

That matters because antibiotic resistance isn’t only a microbiology problem; it’s a timing problem. If a patient needs prompt action, waiting for certainty can be worse than acting under partial evidence. A model that produces a probabilistic ordering - who is likely higher-risk right now - can improve decision quality even if its predictions are not definitive. In other words, the goal is not perfect answers; it’s better timing.

This reframing changes how we interpret performance metrics. It’s easy to equate accuracy with usefulness, but triage systems often trade certainty for speed. A system that is slightly wrong in absolute terms can still reduce harmful delays if its errors occur in a direction that doesn’t disrupt urgency. That’s why evaluation in this space often focuses on operational impact - how many high-risk cases get flagged early, how many low-risk cases avoid unnecessary escalation, and how the model’s outputs align with clinical action windows.

For Nadia’s environment, this is not abstract. When the patient’s physiology is changing quickly, “good enough to prioritize” can outperform “right but late.” The counterintuitive lesson is that the best early model might not be the one that sounds most confident; it might be the one that most reliably sorts the order of attention.

Nadia’s Night Shift: Where “Real Time” Meets Real Workflow

Consider what happens on a busy evening in a hospital that has to balance throughput and care. Nadia works with limited bandwidth: staff availability, beds, imaging queues, pharmacy schedules. Specimens are collected, but the lab’s pace is constrained by incubation and testing capacity. That means the diagnostic pipeline is not just a scientific process - it’s an operations process.

Nadia doesn’t need the model to know the exact organism yet. What she needs is help deciding whether to treat aggressively while waiting, and which additional investigations should happen sooner rather than later. If an AI triage layer flags a high-risk likelihood based on early signal patterns, it can support earlier prioritization - prompting closer monitoring, faster escalation of confirmatory tests, and more careful antibiotic selection. Conversely, lower-risk triage outputs can support restraint in cases where the clinical picture might be misleading and where unnecessary broad-spectrum antibiotics would add future resistance pressure.

In practical terms, the most visible part of AI triage is often its timing. A result that arrives in the first hours changes the decision environment. A result that arrives after culture confirmation changes it less, because the clinical team already has more definitive information by then. The RAPID Lens Triage Model is built around that difference: it aims to produce an output early enough to matter for triage decisions, before the lab finishes its full workflow.

Nadia’s real-world constraint is also human: even with decision support, clinicians must integrate the model’s output with what they see at the bedside. Symptoms, hemodynamics, comorbidities, and prior antibiotic exposure influence the interpretation of any test. AI triage doesn’t remove that work; it reshapes it. It can reduce cognitive load by surfacing pattern-based risk signals early, but it also forces the team to interpret a probabilistic message in the context of a living patient.

That’s why adoption isn’t just technical. It’s about how the model’s output fits into the existing chain of evidence. A triage score that arrives at the wrong moment, or that is difficult to reconcile with current workflows, won’t help. The same model that performs well in retrospective signal analysis might underperform operationally if it doesn’t align with when clinicians can act.

In Nadia’s world, the question becomes: does the output reduce dangerous uncertainty at the point where uncertainty hurts most? The only way to answer that is to watch what happens when triage scores intersect with time pressure, staffing limits, and the pace of lab confirmation.

What This Tells Us About Speed, Evidence, and Trust

AI triage by pattern-matching signals is a lesson in how evidence becomes usable. In microbiology, evidence often arrives as a final verdict - organism identified, susceptibility confirmed. In emergency medicine, evidence arrives in fragments, and clinicians must act before the verdict. A triage model like the RAPID Lens Triage Model tries to bridge the fragment-to-decision gap by translating early signals into a ranking of urgency.

There’s something telling about how this changes trust. People don’t trust numbers by themselves; they trust numbers that arrive at the right moment and connect to a decision they already understand. When AI outputs help Nadia prioritize without pretending to be omniscient, they become part of the evidence ecosystem rather than an external authority.

And that brings us to the broader wonder beyond this chapter: if early signals can be reinterpreted into actionable triage, then the bottleneck in infection care may not be only the lab’s ability to detect organisms - it may be society’s ability to treat uncertainty as information, not as a reason to wait. What else, in medicine and beyond, is currently delayed simply because we’ve insisted on waiting for the final answer?

End of chapter one. 4 more chapters in the full book.