Is ARDS 10-15 years ahead of ACLF?

The question I keep coming back to is whether acute respiratory distress syndrome (ARDS) is 10-15 years ahead of acute-on-chronic liver failure (ACLF).

It is not as strange a comparison as it might first sound.

In ARDS, one patient may have focal bacterial pneumonia with dependent consolidation. Another may have diffuse viral pneumonitis, endothelial injury and a hyperinflammatory phenotype. Both fulfil the same bedside criteria, but their biology and their response to treatment may be quite different.

We see the same problem in ACLF. A patient with alcohol-related hepatitis, marked systemic inflammation and predominantly hepatic failure may share a diagnostic label with a patient who has bacterial sepsis, advanced cirrhosis, vasoplegia, kidney failure and immune dysfunction. The label helps us recognise risk and communicate clearly. It does not mean that these patients have the same illness or need the same treatment.

ARDS and ACLF are both high-mortality syndromes reached through several different routes. Both are defined using accessible clinical variables rather than a single diagnostic test. Most importantly, both risk turning a heterogeneous population into one apparently uniform disease.

ARDS is not a perfect analogy, and it certainly has not solved precision medicine. However, it has already moved through several stages that ACLF is now confronting: agreeing a usable definition, improving supportive care, running repeated negative trials, identifying subphenotypes and beginning to test whether those phenotypes can guide treatment.

In that sense, I think the 10-15 year comparison is broadly right. The useful question is whether ACLF can travel the same road faster.

Definitions gave ARDS a common language

ARDS definitions did not reveal a single disease. Their achievement was more practical: they created a common language.

The 1994 American-European Consensus Conference definition allowed clinicians and researchers to identify a recognisable population. The Berlin definition improved the timing, imaging, origin-of-oedema and oxygenation criteria, while stratifying severity.[1] The newer global definition has widened the frame again, recognising non-intubated ARDS, high-flow nasal oxygen and the realities of resource-limited settings.[2]

None of these is a direct measure of alveolar biology. Even the Berlin categories have only modest discrimination for mortality. LUNG SAFE showed that ARDS was still frequently under-recognised and that lung-protective ventilation was inconsistently applied.[3] A definition can make research possible without guaranteeing recognition, good care or biological precision.

ACLF is still negotiating this stage. The EASL-CLIF definition emerging from CANONIC centres the syndrome on acute decompensation, organ failure and high short-term mortality.[4] APASL starts with acute hepatic insult, jaundice and coagulopathy. NACSELD gives greater weight to extrahepatic organ failures. The recently developed A-TANGO organ-failure score attempts to refine the thresholds and capture additional patients.[5]

They overlap, but they do not identify the same population. Geography, aetiology and clinical setting matter. A hepatitis B-dominant cohort in Asia is not interchangeable with an alcohol- and metabolic-disease-dominant European ICU cohort. Even within one region, changing a creatinine, bilirubin or vasopressor threshold changes who is labelled as having ACLF.

This definitional debate matters. We need to agree who has the syndrome if we are going to study it. My concern is what we then ask the definition to do.

A case definition is a tool for capture. It is not necessarily a treatment phenotype.

One definition does not need to do every job. The best criteria for epidemiology may not be the best criteria for transplant selection. A strong mortality score may still tell us very little about responsiveness to an anti-inflammatory treatment. Greater sensitivity for organ failure does not automatically produce a more treatment-responsive trial population.

The biggest gains came from better supportive care

ARDS improved outcomes most convincingly by reducing iatrogenic injury and delivering supportive care more consistently.

Low tidal-volume ventilation reduced mortality compared with traditional tidal volumes in the ARDS Network trial.[6] A conservative fluid strategy increased ventilator-free days without increasing shock or dialysis in FACTT.[7] Early, prolonged prone positioning produced a striking mortality benefit in selected patients with severe ARDS in PROSEVA.[8]

None of these interventions required us to identify the molecular cause of ARDS at the bedside. They protected the injured lung, reduced avoidable harm or matched a practical intervention to a clinically recognisable group.

For me, this is the most immediately useful lesson for ACLF. The search for a disease-modifying treatment is important, but so is the care around it: early recognition and control of precipitants, appropriate antimicrobials and source control, careful haemodynamic assessment, avoidance of nephrotoxins and fluid overload, organ-protective ventilation, renal support, nutrition, early transplant evaluation and repeated review of reversibility.

There may never be a neat ACLF equivalent of six millilitres per kilogram of predicted body weight. It is a multiorgan syndrome superimposed on a chronically diseased liver. Even so, an agreed minimum care bundle could improve practice and reduce noise in trials. It is difficult to interpret an investigational therapy when co-interventions, transplant referral and organ-support practices vary substantially between centres.

ARDS also tells us that evidence does not implement itself. LUNG SAFE found important gaps in recognition and delivery long after low tidal-volume ventilation had become standard.[3] ACLF trials should report not only the allocated treatment, but how reliably the surrounding standard of care was delivered.

Negative trials forced ARDS to confront heterogeneity

ARDS has tested many plausible pharmacological treatments with largely disappointing results. Some simply did not work. However, a broad syndromic trial can also combine patients in whom a treatment helps, patients in whom it is irrelevant and patients in whom it causes harm.

Latent class analyses of ARDS trials repeatedly identified broadly hyperinflammatory and hypoinflammatory subphenotypes, differing in biomarkers, vasopressor use, acidosis and outcome.[9] In secondary analyses, these groups appeared to respond differently to fluid strategy and simvastatin.[10,11] This is not yet a licence to prescribe by phenotype. The interactions were identified retrospectively and need prospective confirmation. This work did, however, change the question from “Does this drug work in ARDS?” to “Is there a recognisable subgroup in which it works?”

The LIVE trial provided an equally important caution. A ventilation strategy personalised to focal or non-focal lung morphology did not improve 90-day survival overall. Misclassification was common, and patients whose morphology was incorrectly classified appeared to fare worse.[12]

The lesson is uncomfortable but important. If a classifier determines treatment, classification error becomes part of that treatment's risk.

ACLF may be even more heterogeneous. The underlying hepatic substrate, regenerative reserve, precipitant, pattern of organ failures, balance between inflammation and immune dysfunction, direction of change, frailty and transplant eligibility may all matter. A binary ACLF label captures very little of that.

Patients are still commonly enrolled into studies because they cross a syndromic or severity threshold. That may be necessary for recruitment. It may not be enough for treatment enrichment.

The history of ACLF intervention studies is already instructive. MARS improved several biochemical and physiological measures in the RELIEF trial but did not demonstrate a survival benefit at the scheduled dose.[13] European multicentre testing of G-CSF in the GRAFT trial did not improve survival or key clinical outcomes, despite earlier signals from smaller studies and a plausible regenerative rationale.[14] Targeted albumin administration in hospitalised decompensated cirrhosis did not improve the composite clinical outcome in ATTIRE and caused more serious adverse events.[15] DIALIVE, by contrast, has shown mechanistic effects and faster ACLF resolution in a small randomised first-in-human study, but it was not powered to establish a mortality benefit.[16]

These results do not prove that there is a hidden group of responders. A negative intervention may simply be ineffective. They do show why future trials must do more than recruit a population with a high event rate. High risk improves statistical efficiency. It does not guarantee a shared, modifiable mechanism.

This distinction is central. Prognostic enrichment finds patients likely to have the outcome. Predictive enrichment finds patients more likely to respond to the treatment. ACLF needs both, and should not confuse them.

ACLF should take time seriously

ACLF is a dynamic syndrome. Clinically, we already think in those terms. Is the patient improving, worsening or simply accumulating further organ failure? Our research classifications are often much more static than our bedside decisions.

ACLF grade can improve, resolve or worsen over a few days. In CANONIC, reassessment between days 3 and 7 was more informative than initial grade alone and identified patients with very different short-term outcomes.[17] PREDICT subsequently described distinct courses of acute decompensation, including a pre-ACLF group with intense systemic inflammation and a high risk of progression.[18] The route into ACLF contains information too.

Most classifications still flatten this moving process into a single admission value.

A bilirubin of 300 µmol/L is not a phenotype. A bilirubin of 300 that has risen rapidly for three days may represent something quite different from the same value falling after control of the precipitant. The same is true of vasopressor requirement, creatinine, lactate, encephalopathy and oxygenation.

This is one area in which ACLF could move faster than ARDS. Trials should collect a minimum serial dataset from the outset, for example at admission, 24 hours, 72 hours and a later landmark such as day 5 or 7. Trajectory should be designed into the study, not drawn afterwards.

This does not mean waiting several days to treat a deteriorating patient. Some treatments will need to start before a trajectory is clear. It means designing the study around the time window in which the treatment is biologically plausible. Prevention, early rescue and established multiorgan failure are not the same clinical state. They should not automatically share the same entry criteria.

There are obvious methodological problems. Patients must survive and remain observable long enough to be classified. Treatment changes the subsequent trajectory. Missing measurements are rarely random. Studies need prespecified landmarks, explicit handling of transplant and death, and honest reporting of classification uncertainty. Retrospectively assigning neat labels to survivors is not enough.

A practical ARDS roadmap for ACLF

So what should ACLF borrow?

1. Build a common capture framework, not a universal answer

The field needs an operational core dataset and transparent mapping between the major definitions. Regional differences should be studied rather than averaged away. A patient should be described by hepatic substrate, precipitant, organ failures and time, not reduced to “ACLF: yes/no”.

2. Standardise and measure supportive care

ACLF trials should define a minimum care bundle and report adherence to it. This should include precipitant management, antimicrobial and haemodynamic strategy, organ support, nutrition and timing of transplant assessment. Variation in ordinary care can otherwise obscure the effect of the treatment being tested.

3. Separate severity from biology

CLIF-C ACLF, MELD-Na and related scores are valuable prognostic tools. They were not designed to prove that patients share a treatment-responsive mechanism. Trial programmes should be clear whether a biomarker or classifier is prognostic, predictive or simply operational.

4. Make serial phenotyping routine

Clinical variables, biosamples and organ-support intensity should be collected repeatedly at prespecified time points. The aim is not to build the most complex model. It is to identify reproducible states and transitions early enough to change management.

5. Validate classifiers prospectively before allowing them to allocate treatment

Retrospective treatment interactions generate hypotheses. A bedside classifier must be feasible, calibrated across populations and able to express uncertainty. LIVE reminds us that a plausible treatment rule can become harmful when the patient is assigned to the wrong group.[12]

6. Design trials around mechanism and timing

An anti-inflammatory therapy, regenerative therapy, antimicrobial strategy and extracorporeal device should not all recruit “ACLF” in the same way. Each needs a clear rationale: which biological state is being targeted, when it is present, how target engagement will be measured and what competing events may intervene.

7. Treat transplantation as central, not as statistical inconvenience

This is where the ARDS analogy breaks most clearly. Selected patients with ACLF-3 can achieve one-year post-transplant survival above 80%, while survival without transplantation may be extremely poor.[19] Transplantation changes the clinical pathway and the meaning of study endpoints. Trials need prespecified approaches to listing, transplantation and transplant-free survival, alongside outcomes such as ACLF resolution and organ-support-free days.

Where the analogy stops

ARDS can be a roadmap, but it is not a template.

ARDS is centred on injury to one organ, even when embedded in multiorgan critical illness. ACLF begins with chronic hepatic disease and rapidly becomes a systemic syndrome. Hepatic reserve, portal hypertension, cirrhosis-associated immune dysfunction and the possibility of replacing the failing organ fundamentally alter its biology and clinical decisions.

ARDS has not completed the journey either. Its supportive-care evidence is strong, but prospectively validated biological treatment selection remains a work in progress. The hyperinflammatory phenotype is one of critical care’s most interesting research stories. It is not yet routine bedside care. The PANTHER precision-medicine platform was only recently established to test the approach prospectively.[20] ACLF could easily spend the next decade producing increasingly sophisticated clusters that predict death without telling us who benefits from treatment.

The lesson is not simply to “find two phenotypes”. The sequence matters:

1. define a syndrome well enough to recognise it;

2. standardise care well enough to study it;

3. measure its heterogeneity rather than average it away;

4. prove that a proposed phenotype is reproducible;

5. then show prospectively that it changes the effect of treatment.

So, is ARDS 10-15 years ahead?

My answer is yes, conceptually at least. ARDS is ahead not because it has solved heterogeneity, but because it has already encountered many of the problems ACLF is now facing. It has shown the value of a common definition and protocolised care, the repeated failure of one-size-fits-all pharmacology and the potential risks of phenotype-guided treatment.

That history is useful. ACLF does not need to repeat it at the same speed.

The next generation of ACLF studies should preserve the value of a common syndrome label without mistaking it for biological uniformity. They need robust case capture, serial measurement, a clear mechanistic rationale and prospective testing of treatment effects.

The next ACLF trial should not ask only: does this treatment work in ACLF?

It should ask: which ACLF, at what moment, through which mechanism, and towards which outcome?

Case capture gets patients into a trial. It should not be mistaken for the treatment rationale.

That, for me, is the roadmap ARDS offers. Learning from it could save ACLF a decade.

Key points

• A useful syndrome definition does not identify a biologically uniform population.

• ARDS made its clearest gains through better supportive care. ACLF should give the surrounding care bundle the same attention as the investigational treatment.

• Prognostic enrichment is not predictive enrichment. High-risk patients do not necessarily share a modifiable mechanism.

• Serial change should be designed into ACLF phenotyping and trial recruitment from the start.

• Classifiers need prospective validation before they allocate treatment, and transplantation must be built into both trial design and endpoint selection.