Sepsis: when the response to an infection becomes the thing that kills you

Start here: what sepsis is

Sepsis is not an infection. It is what can happen because of one.

Most infections stay local and the immune response that clears them is proportionate. Occasionally that response goes wrong: it becomes systemic, self-amplifying and self-damaging, and it starts to injure the body's own organs — the kidneys stop filtering, blood pressure falls, the lungs stiffen, consciousness clouds, clotting goes haywire. The current international definition captures exactly this: sepsis is life-threatening organ dysfunction caused by a dysregulated host response to infection ^[1].

That definition was a deliberate break from the past. The Third International Consensus task force (Sepsis-3) named the limitations of the previous framework explicitly: an excessive focus on inflammation, the misleading model that sepsis follows a continuum through "severe sepsis" to shock, and the inadequate sensitivity and specificity of the older systemic inflammatory response syndrome (SIRS) criteria ^[1]. "Severe sepsis" was retired as a category. Septic shock was redefined as a subset with circulatory and cellular-metabolic abnormalities severe enough to substantially increase mortality — in practice, a patient needing vasopressors to maintain blood pressure and with a raised lactate despite adequate fluid.

The scale is larger than almost anyone outside medicine realises. Using multiple cause-of-death data from 109 million individual death records, the Global Burden of Disease analysis estimated 48.9 million incident cases of sepsis worldwide in 2017 and 11.0 million sepsis-related deaths — 19.7% of all global deaths ^[2]. Roughly one death in five, worldwide, involves sepsis somewhere in the causal chain. Age-standardised incidence and mortality both fell substantially between 1990 and 2017, and the burden is heaviest in sub-Saharan Africa, Oceania and south and east Asia ^[2]. The paediatric and neonatal burden has been estimated separately ^[3], and paediatric criteria were revised in 2024 ^[4].

Why speed is the whole story, and why that is hard. Sepsis is a time-critical emergency in which the treatment — antibiotics, fluid, source control, organ support — is not exotic. What is difficult is recognising it soon enough. The early signs are entirely non-specific: fever or low temperature, fast breathing, fast heart rate, confusion, feeling terrible. Those are also the signs of an ordinary infection that will resolve on its own. There is no test that says "this one is sepsis." A US study of sepsis-associated mortality in acute care hospitals examined its prevalence, underlying causes and — crucially — its preventability ^[5].

Its counterpart in this series. The kidney is often the first organ to fail, and sepsis-associated acute kidney injury is common enough to have its own consensus report ^[6]. The chronic kidney disease review covers what happens when kidney function is lost slowly; this one covers what happens when it is lost in hours.

Three pillars follow — measurements, management, and progress — with a model between the first two about the single quantity that clinicians can actually control: time.

Pillar 1: measurements and diagnosis

A syndrome, not a test

This point governs everything else. There is no blood test for sepsis. There is a test for infection — culture the blood, the urine, the sputum, the wound — and there are tests for organ dysfunction — creatinine, bilirubin, platelet count, oxygenation, blood pressure, mental status. Sepsis is the clinical judgement that the second is being caused by the first.

That is why the definition is operationalised through a score rather than a threshold on any single measurement.

SOFA and qSOFA, and which one belongs where

The Sequential Organ Failure Assessment (SOFA) score grades dysfunction across six organ systems — respiratory, coagulation, liver, cardiovascular, central nervous system and renal. Under Sepsis-3, sepsis is operationalised as suspected infection plus an acute rise of 2 or more SOFA points ^[1].

SOFA needs laboratory results, so the same task force derived a bedside prompt. qSOFA scores one point each for systolic blood pressure ≤100 mmHg, respiratory rate ≥22/min, and altered mentation — three things anyone can assess in seconds without a laboratory ^[7].

The validation study is worth reading carefully, because it is routinely misquoted in both directions. Across 1.3 million electronic health record encounters at 12 hospitals, with confirmatory analysis in a further 706,399 encounters at 165 hospitals:

• Outside the ICU, qSOFA had the best predictive validity for in-hospital mortality (AUROC 0.81), statistically greater than SOFA (0.79) and SIRS (0.76) — supporting its use as a prompt to consider sepsis.
• Inside the ICU, the ordering reversed: SOFA performed best (AUROC 0.74) and qSOFA was distinctly worse (0.66), barely better than SIRS (0.64) ^[7].

So qSOFA is a screening prompt for the ward and the emergency department, not a diagnostic criterion and not an ICU severity score. Relative to a score below 2, patients with qSOFA ≥2 had a 3- to 14-fold increase in hospital mortality across baseline risk deciles ^[7]. An AUROC of 0.81 is useful; it is not a test, and a low qSOFA does not exclude sepsis.

Lactate

Lactate is the closest thing sepsis has to a severity dial. It rises when tissue oxygen delivery fails to meet demand and when metabolic handling of lactate is impaired — so an elevated serum lactate is a marker of the circulatory and cellular derangement that defines shock, and it is written into the Sepsis-3 septic shock criteria alongside the vasopressor requirement ^[1]. Its prognostic accuracy has been compared head-to-head against SOFA and qSOFA for mortality in adults with sepsis ^[8], and lactate clearance is one of the resuscitation targets in the campaign bundles ^[9].

Cultures, source, and procalcitonin

Two things must happen alongside the scoring. Blood cultures should be drawn before antibiotics where this does not delay them, because knowing the organism allows the initial broad-spectrum regimen to be narrowed later. And the source must be found — pneumonia, urinary tract, abdomen, skin, line, joint — because some sources need a physical intervention rather than a drug. Antibiotics do not drain an abscess.

Procalcitonin is a blood marker that rises more in bacterial infection than in viral illness or sterile inflammation. Its established value is less in making the diagnosis than in stopping treatment: procalcitonin-guided protocols have been used to initiate or discontinue antibiotics in respiratory infection ^[10], shortening courses in a way that matters both for resistance and for the patient ^[9].

Centerpiece: a simple simulatable model of time to antibiotic

Everything in Pillar 1 exists to buy time, so the model is about what time is worth.

The founding observation comes from a retrospective cohort of 2,731 adults with septic shock across 14 ICUs and 10 hospitals in Canada and the United States between 1989 and 2004. Among patients who received effective antimicrobial therapy only after hypotension began, there was a strong relationship between delay and in-hospital mortality: an adjusted odds ratio of 1.119 per hour of delay (95% CI 1.103–1.136, p<0.0001). Treatment within the first hour of documented hypotension was associated with a survival rate of 79.9%. In multivariate analysis, time to effective antimicrobial therapy was the single strongest predictor of outcome ^[11].

Model that as a constant per-hour odds ratio acting on the odds of death:

odds(t) = odds(1) · OR^(t−1), and survival(t) = 1 / (1 + odds(t))

anchored at the grounded 79.9% survival at hour 1. That leaves exactly one contested parameter — the odds ratio — and the interesting thing is that the literature supplies three different values for it.

What the model explains. Four things.

First, why "every hour counts" is the right instinct. Every published estimate has the same sign. Delay is associated with death, in a retrospective cohort from the 1990s and in a 49,331-patient prospective mandated-protocol cohort two decades later ^{[11] [12]}. A separate cohort likewise reported that empiric antibiotic treatment from the first hour reduced mortality in severe sepsis and septic shock ^[13]. Whatever the magnitude, the direction has never been in doubt, and there is no plausible mechanism by which waiting helps.

Second, why you should be careful with the number you have heard. The widely-repeated "7.6% per hour" comes from the same paper as the adjusted odds ratio of 1.119, and the two are not the same claim. Plotted on one axis they diverge by 33 percentage points by hour 7. The adjusted odds ratio is the more defensible of the two — it comes from the multivariate model — and it implies a real but much gentler decline. When a striking number circulates without its companion, the effect looks larger than its own source's primary analysis supports.

Third, why the mandated-care estimate is shallower, and why that is not a contradiction. In the New York cohort the median time to antibiotics was under an hour and 82.5% completed the 3-hour bundle on time ^[12]. When almost everyone is treated fast, the remaining delays are smaller and the patients who experience them differ in ways adjustment may not fully capture. A shallower slope in a well-run system is what you would expect if the steep part of the curve had already been removed by the system itself.

Fourth, the honest confounder. Both studies are observational. Sicker patients may be recognised faster because they look worse, which biases toward less apparent benefit of speed; or patients with atypical presentations may be both slower to treat and sicker for unrelated reasons, which biases the other way. Randomising patients to delayed antibiotics is not ethically possible, so this question will never have the trial that would settle it.

What the model deliberately does not do. It applies one odds ratio across twelve hours when the true relationship is almost certainly not constant. It describes septic shock — the sickest subset — and does not generalise to milder sepsis, where the time-dependence is weaker and the case for immediate broad-spectrum antibiotics has to be balanced against overtreatment. It says nothing about whether the antibiotic was the right one, which is what "effective" is doing in the original phrasing. And it isolates one component of a bundle: the same New York study that found an antibiotic-timing association found none for time to fluid bolus ^[12].

Pillar 2: management — a bundle, not a drug

There is no sepsis drug. Outcome depends on doing several ordinary things quickly and well, and the field's central instrument is therefore a bundle: a small set of actions to be completed within a defined window. The Surviving Sepsis Campaign has issued and revised these guidelines repeatedly ^{[14] [15] [16] [9] [17]}, with a 2018 update that compressed the earlier 3-hour and 6-hour bundles into an hour-1 bundle ^[18].

The components

Early effective antibiotics, broad-spectrum initially, narrowed once cultures return — the subject of the model above ^{[11] [12]}.

Source control. Drainage, debridement, removal of an infected line or device — a bundle element present in the campaign guidelines since their first edition ^{[19] [20]}. This is often the decisive intervention and it is not pharmacological.

Fluid resuscitation. Intravenous crystalloid to restore circulating volume. The composition matters somewhat — balanced crystalloids versus saline has been trialled in critically ill patients ^[21] — and the amount matters a great deal, because too much fluid causes its own harm. The most sobering evidence here is the FEAST trial, in which fluid bolus resuscitation increased mortality in African children with severe febrile illness ^[22]. That result is a permanent caution against assuming more fluid is better.

Vasopressors when fluid alone cannot maintain perfusion — noradrenaline first-line, with vasopressin as an adjunct studied in its own trials ^[23], and the vasopressor requirement forming part of the septic shock definition itself ^[1].

Organ support — mechanical ventilation, renal replacement therapy — buys time while the infection is treated. Sepsis-associated acute kidney injury has its own consensus framework ^[6].

Corticosteroids remain genuinely contested. Low-dose hydrocortisone with fludrocortisone reduced mortality in one septic shock trial ^[24]; a later trial of hydrocortisone in septic shock did not confirm a survival benefit ^[25]. The guidelines reflect that unresolved disagreement rather than settling it ^[9].

What the bundles displaced, and the argument they started

The modern era began with early goal-directed therapy, a protocolised resuscitation strategy that reported a large mortality reduction in a single-centre trial ^[26] and reshaped practice worldwide. It was then tested in large multicentre trials — ProCESS randomised patients to protocol-based care ^[27] and ARISE to goal-directed resuscitation ^[28] — which found no benefit over usual care. The usual reading is not that the original was wrong but that the control arm had improved: the general standard of early recognition and prompt treatment had risen so much that the protocol added nothing on top.

That history is why the hour-1 bundle is contested. The disagreement is not about whether speed helps. It is about whether a rigid, auditable time target is the right instrument, given that mandating antibiotics within an hour for everyone with suspected sepsis pushes clinicians toward broad-spectrum treatment of patients who turn out not to have a bacterial infection at all — with consequences for that patient and for antimicrobial resistance ^[29]. Both concerns are legitimate, and no bundle design has yet resolved the tension between them.

An honest note on failed drugs

For a decade, activated protein C was the exception: a recombinant biologic that reduced mortality in severe sepsis in a large trial ^[30], was approved, and was later withdrawn after subsequent trials failed to confirm the benefit. Sepsis has an unusually long list of therapies that worked in trials and did not survive contact with wider practice — one reason the field turned toward phenotyping, described below, rather than searching for another single agent.

Pillar 3: progress

Recognising sepsis earlier, by machine

If recognition is the bottleneck, then automated surveillance of the electronic health record is the obvious lever. Prediction models using minimal EHR data have been developed for the ICU ^[31], gradient-boosted models have been trained to predict mortality in Sepsis-3 cohorts ^[32], and the general reporting standards for such models have been formalised ^[33]. These systems sit inside the same trade-off as qSOFA: raise the sensitivity and you generate alerts on patients who do not have sepsis, and alert fatigue is itself a patient-safety problem.

Precision sepsis: the syndrome is several diseases

The most conceptually important development is the recognition that "sepsis" aggregates patients whose biology differs. Using consensus clustering on 29 clinical variables in 20,189 patients meeting Sepsis-3 criteria, and validating in a further 43,086 patients, a prospective pneumonia cohort and three randomised trials, investigators derived four clinical phenotypes (α, β, γ, δ) that differed in host-response biomarkers and in mortality — and, critically, in simulated response to trial interventions ^[34].

That last point is the reason this matters. If a trial enrols all four phenotypes and a therapy helps one and harms another, the trial reads as neutral. Some of sepsis's long list of failed drugs may be treatments that worked in a subgroup nobody had identified. Transcriptomic and host-response subclassification is being pursued along the same lines, and the underlying immunology supports it: sepsis involves not only hyperinflammation but a later immunosuppressed state, sometimes called immunoparalysis, in which cellular dysfunction leaves patients vulnerable to secondary infection ^{[35] [36] [37]}. The most striking demonstration is that patients with sepsis reactivate multiple latent viruses, at rates comparable to transplant recipients on immunosuppressive drugs ^[38].

Surviving sepsis is not the end

The outcome literature has shifted from hospital mortality toward what happens afterwards. Survivors carry elevated risks of death, readmission, new physical disability and cognitive impairment for months to years — a cluster often called post-sepsis syndrome, and one in which persisting immune derangement is implicated ^[39]. For a condition historically measured by whether patients left hospital alive, taking the years after discharge seriously is a substantive change in what counts as success.

Antibiotics themselves are a finite resource

Finally, the treatment at the centre of this review is under threat from its own use. The global burden of bacterial antimicrobial resistance has been quantified with forecasts to 2050 ^[29], and mortality associated with 33 bacterial pathogens estimated for 2019 ^[40]. Sepsis care depends on broad-spectrum antibiotics working; the pressure to give them immediately to everyone who might have sepsis is in direct tension with preserving them. Procalcitonin-guided de-escalation and prompt narrowing once cultures return are the practical responses, and they are the reason "start fast, then stop early" is better advice than either half alone ^[9].