Chapter question: The patient is on pump, the cause has been found, the temperature is being held, and the family is asking the only question they care about. Every instrument that could answer it was validated in a population where doctors withdrew treatment on the basis of that same instrument — and half of those instruments behave differently on a circuit. What can honestly be said, and when?
Evidence search date: 14 September 2026.
Primary sources: ELSO Red Book 6th edition — Chapter 32 (neuroprognostication aligned with conventional CPR practice, the neurological care bundle), the adult systematic-review complication figures, and the paediatric and neonatal neuromonitoring chapters. ECPR and Resuscitative ECMO (Shinar and Badulak), Chapter 3 — the clinical examination findings at hospital arrival and the signs-of-life literature.
External evidence: the ECPR brain-injury epidemiology (an autopsy series, three ELSO Registry analyses, two meta-analyses), the neuroprognostication literature in full (the ERC/ESICM algorithm and its two validation cohorts, the two no-withdrawal cohorts, the ECMO-specific comparisons, the PRAGUE-OHCA biomarker sub-analysis, the SAVE-J II imaging analysis of 1,146 patients, the ECPR electroencephalography studies, and the good-outcome literature).
Everything external in this chapter comes from a structured abstract. Nothing was retrieved in full text. Every database row seeded from this chapter is marked Verified = No.
What this chapter covers — and what it does not
Chapter 22 ended where prediction begins. This chapter is about inference: what can be concluded about a brain from the instruments available, and with what confidence. It is not about the instruments themselves, and it is not about what is then done.
This chapter owns | Owned elsewhere — cite, do not re-argue |
What actually happens to ECPR patients neurologically, and how much of it is never detected | Cerebral monitoring as a set of instruments — near-infrared spectroscopy, transcranial Doppler, electroencephalography — Chapter 28 |
The self-fulfilling prophecy, and which studies escape it | The decision to stop, and the conversation with the family — Chapter 24 |
What the circuit does to each prognostic instrument | Futility and non-beneficial ECMO as a concept — Chapter 74; withdrawal of ECMO as a procedure — Chapter 75 |
Neuron-specific enolase on ECMO, and why its threshold is not the published one | Laboratory monitoring generally — Chapter 29; haemolysis — Chapter 37 |
Electroencephalography as an instrument of prognosis | Electroencephalography as a seizure monitor, and the treatment of seizures — Chapter 22 §22.12, Chapter 28, Chapter 51 |
Brain imaging for prediction, and the grey-white matter ratio in ECPR specifically | Brain imaging to find the cause of the arrest — Chapter 22 §22.8 |
When prediction may begin, and why sedation moves the clock | Sedation, analgesia and the ECMO sequestration problem — Chapter 57, Chapter 49 |
Predicting a good outcome — the neglected half of the problem | Pre-cannulation selection scores (TiPS65, the Pre-ECPR score) — Chapter 20; scores as a class — Chapter 30 |
How to run a prognostication process, and the errors that recur | Long-term functional and cognitive recovery — Chapter 59; palliative care — Chapter 76 |
The boundary with Chapter 24 is the one that matters. This chapter produces a probability. Chapter 24 decides what to do about it. Those are different acts, performed by overlapping people at the same bedside, and conflating them is the single commonest error in this field — because a prediction that is made in order to justify a decision has already stopped being a prediction.
23.1 Three destinations, and how many patients reach each
Begin with base rates, because almost every conversation in this chapter goes wrong when they are not known.
Evidence · Low certainty · Where ECPR patients end up
The largest synthesis covers 78 studies and 50,049 patients, of whom 6,261 (12.5%) received ECPR — median age 56, 63% male, 48% out-of-hospital, 37% initial shockable rhythm, median support duration 3.2 days.
Neurological events: at least one neurological complication in 27% (17–39%); hypoxic-ischaemic brain injury 23% (14–32%); ischaemic stroke 6% (2–11%); seizures 6% (1–16%); intracerebral haemorrhage 4% (1–10%); and brain death in 17% (12–23%).
Outcomes: survival 29% (26–33%); good neurological outcome 24% (21–28%).
Those last two numbers are the ones to hold. Survival 29%, good outcome 24% — which means roughly five in six survivors had a good neurological outcome.
That ratio is not an artefact of one synthesis. It recurs everywhere it is measured.
Cohort | Survival | Good outcome | Good outcome as a share of survivors |
Migdady 2020, 78 studies, 6,261 ECPR patients | 29% | 24% | ~83% |
Gravesteijn 2020, in-hospital arrest, 19 studies | 30% (28–33%) | — | 84% (80–88%), stated directly |
Pozzi 2026, 295 consecutive patients over 14 years, outcome at 6 months | 21.7% | 17.3% (28.2% in-hospital arrest vs 10.1% out-of-hospital) | ~80% — only 4.4% of the whole cohort survived with severe sequelae |
Rojas-Salvador 2023, 361 patients, single centre | 47% | 40% | 85% (145 of 170) |
Han 2019, 100 patients (Chapter 22 §22.1) | 14% | 12% | 85.7% (12 of 14) |
Clinical pearl · The outcome distribution in ECPR is close to bimodal, and this is not what most clinicians expect
Ask a clinician what they fear most about ECPR and the answer is usually the same: creating a survivor with a devastated brain. That outcome is the rarest of the three destinations. In the largest long-term series, 17.3% of patients reached a good neurological outcome at six months and 4.4% survived with severe sequelae — roughly a four-to-one ratio in favour of the outcome everyone wants.
The practical form: when a patient survives ECPR, the strong prior is that they will be neurologically reasonable. The uncertainty in this population is overwhelmingly about whether they will survive, not about what state they will survive in.
This is a prior, not a prediction, and §23.2 explains why it may be partly manufactured. But it is the correct starting point for a family conversation on day one, and it is very often not what the family is told.
Danger · The bimodality may be manufactured by the very practice this chapter describes
If a unit withdraws treatment from every patient whose neurological indicators look bad, and does so before those patients could have declared themselves, then the severely disabled survivor is rare because the unit prevented them from existing. The observed distribution would look exactly the same either way.
Two observations keep this from being a purely theoretical worry. First, in the 45-patient Prague sub-analysis with brain imaging, brain death accounted for 41% of deaths among poor-outcome patients, multiorgan failure 38% and recurrent cardiac arrest 13% — so a substantial share of deaths were not neurological at all, and a substantial share were irreversible by any definition. Second, and cutting the other way, Chapter 22 §22.9 recorded a registry analysis in which hypothermia was associated with survival but not with a favourable neurological outcome — the signature of additional survivors in poor neurological states.
Both facts are true at once. This book records the bimodality as the best available description of what happens, and records that its magnitude is not independently verifiable in a system that withdraws treatment.
The part nobody measures
There is a second problem with the base rates, and it is larger than it looks.
How brain injury was looked for | Population | How much was found |
ELSO Registry, voluntary reporting, no protocol | 10,775 ECPR patients | 16.5% |
ELSO Registry, same analysis, non-ECPR venoarterial | 35,855 patients | 7.7% |
Meta-analysis of published series | 6,261 ECPR patients | 27% any neurological complication; 23% hypoxic-ischaemic |
Standardised multimodal neuromonitoring at one centre | 123 venoarterial patients | 33% |
Electroencephalography actually applied | 92 ECMO patients monitored, 28% of them ECPR | 63% had epileptiform activity or ictal-interictal continuum patterns |
Post-mortem brain autopsy | 25 ECMO decedents, 22 venoarterial | 68% had acute brain injury — hypoxic-ischaemic 44%, intracranial haemorrhage 24%, ischaemic infarct 16%. Only 8 of 25 had none |
Physiology · The detection gradient is the finding
Read that table downwards and the prevalence of brain injury rises from 16.5% to 68% — and nothing about the brains changed. Only the intensity of looking changed.
The registry figure is a floor, and its own authors say so: the machine-learning analysis of the venovenous registry concluded that performance was suboptimal "likely due to lack of standardization of neuromonitoring/imaging protocols and data granularity," and called for standardised monitoring "to detect the true prevalence." The Red Book makes the same point about paediatric registry data: complications reported to ELSO "should be viewed as minimal estimates" because reporting is voluntary and evaluation is not systematic.
The operational consequence for this chapter is uncomfortable. A clinician predicting outcome from a registry-derived risk figure is using a denominator built from unsystematic detection. It also means that a unit which begins monitoring systematically will appear to get worse before it appears to get better.
Evidence · Low certainty · ECPR is the reason venoarterial ECMO looks neurologically dangerous
A meta-analysis of 73 studies and 16,063 patients compared brain injury between modes: venoarterial 19% versus venovenous 10% (P = 0.002), with ischaemic stroke 10% versus 1%, hypoxic-ischaemic injury 13% versus 1% and brain death 11% versus 1%. Intracerebral haemorrhage did not differ (6% versus 8%).
Then the authors excluded the studies containing ECPR — and the overall difference vanished: 13% versus 10%, P = 0.4.
The registry analysis makes the same point with different arithmetic: acute brain injury in 16.5% of ECPR patients against 7.7% of non-ECPR venoarterial patients.
The brain injury in this population is a property of the cardiac arrest, not of the venoarterial configuration. Chapter 17's vascular complications belong to the cannula; these belong to the forty minutes before it.
Pitfall · Two different injuries with two different clocks
The ECPR patient is at risk of hypoxic-ischaemic injury that was inflicted before cannulation, and of stroke and haemorrhage that occur on support. They do not arrive at the same time, and a single scan does not exclude both.
In an 878-patient venoarterial cohort, ischaemic stroke occurred after a median of 11 days (6–18) and intracranial bleeding after 5 days (2–9). Intracranial bleeding, not ischaemic stroke, was associated with higher mortality. Female sex, central cannulation and a platelet count below 100 at initiation were independently associated with bleeding — as, in a nested analysis, was a rapid carbon dioxide change at ECMO start, which is Chapter 22 §22.3 arriving from the haemorrhagic side.
A normal brain CT on day one says nothing about day eleven.
23.2 The problem that contaminates every number in this chapter
This is the organising idea, and it has to come before any individual test, because it changes how every subsequent number should be read.
The standard multimodal approach to neuroprognostication after cardiac arrest is built on predictors validated against a single figure of merit: specificity for a poor outcome, ideally 100%, because a false positive means predicting death or devastation in a patient who would have recovered. The literature delivers that figure repeatedly.
Study | n | Result |
Moseby-Knappe 2020, the ERC/ESICM four-step algorithm applied to the TTM trial cohort | 585 | Sensitivity 38.7% (33.1–44.7), specificity 100% (98.8–100). Exploratory variations reached at best 42.5% (36.7–48.5) sensitivity, still without false positives |
Bougouin 2024, prospective, 28 intensive care units | 337 | The algorithm predicted poor outcome in 175 patients; positive predictive value 100% [98–100]%. Individual predictor specificity ranged from 90% (electroencephalography) to 100% (clinical examination, evoked potentials) |
Sandroni 2020, systematic review | 94 studies, 30,200 patients | Bilaterally absent pupillary or corneal reflexes after day 4; high neuron-specific enolase from 24 h; absent N20 waves or unequivocal seizures; background suppression or burst-suppression from 24 h; diffuse oedema on CT from 2 h; reduced diffusion on MRI at 2–5 days — each had 0% false-positive rate in most studies. Risk of bias was high for every predictor |
Now read those studies again with one question in mind: what happened to the patients in whom the test was positive?
Danger · The specificity of these tests is partly a measurement of clinician behaviour
In almost every cohort above, a positive prognostic test was followed by withdrawal of life-sustaining treatment. The patient then died. The death was recorded as a poor outcome. The test predicted the outcome it caused.
This is not a fringe criticism; it is the authors' own. Moseby-Knappe's conclusion states that the results "should be validated prospectively, preferably in patients where withdrawal of life-sustaining therapy is uncommon to exclude any confounding from self-fulfilling prophecies." Bougouin's states that all patients meeting the criteria had a poor outcome "even if a self-fulfilling prophecy cannot be completely excluded." Sandroni rated the risk of bias high for every single predictor.
A 100% specificity generated inside a system that withdraws on the test is not a measurement of the brain. It is, in part, a measurement of the unit. And the direction of the bias is known: it inflates specificity, which is exactly the property clinicians rely on.
The two cohorts that escape the circle
The objection above is fatal to most of the literature. It is not fatal to all of it, because a small number of investigators did the hard thing and studied cohorts in which treatment was not withdrawn.
Evidence · Moderate certainty · Prognostication where nobody withdrew
ProNeCA, prospective, seven hospitals, 346 of 396 patients analysed. "None of the patients underwent withdrawal of life-sustaining treatments before the index event occurred." At six months, 223 (64%) had a poor outcome, of whom 68 were alive in a persistent vegetative state — the outcome that a withdrawing system would never have produced. Within 24 hours of arrest, at 100% specificity: bilaterally absent or absent-pathological cortical evoked potentials, sensitivity 57.4%; grey-to-white matter ratio below 1.21, 48.8%; isoelectric or burst-suppression electroencephalography, 34.5%. At least one unfavourable pattern was present in 74.4%; two simultaneously in 49.7%; all three in 17%.
Götze 2026, prospective, eight hospitals, patients comatose at 72 hours, and — critically — patients who underwent withdrawal within the first four weeks were EXCLUDED, leaving 101. Poor outcome (modified Rankin 4–6) at 12 months occurred in 67.3%. Individual markers had high specificity and limited sensitivity with variable false-positive rates. Two or more unfavourable markers within 14 days yielded no false positives, adjusted OR 34.7. Adding the best Coma Recovery Scale–Revised score improved discrimination, AUC 0.925 versus 0.888. Registered NCT02231060.
These two cohorts are the epistemic foundation of the whole field, and they are small: 346 and 101 patients, neither on ECMO.
Controversy 1 — If the specificity figures are circular, is the multimodal rule still usable?
The case that the literature is fatally circular. Every large validation cohort permitted withdrawal on the basis of the tests being validated. The mechanism by which this inflates specificity is not speculative — it is arithmetic: a patient predicted to do badly, in whom treatment is stopped, cannot become a false positive. The authors of the two most cited validations both say so explicitly. Risk of bias was rated high for every predictor in a 30,200-patient systematic review. On this reading, "100% specificity" is an uninterpretable number and clinicians who quote it to families are quoting their own practice back to themselves.
The case that the rule survives anyway. Two prospective cohorts broke the circle by not withdrawing, and both reproduce the central finding: individual predictors are highly specific and poorly sensitive, and the combination of two or more unfavourable markers produced no false positives. ProNeCA did this while allowing 68 patients to remain alive in a vegetative state — the population most capable of generating false positives, retained rather than removed. Götze did it with a 12-month endpoint, which is long enough for late recovery to appear if it were going to. If the multimodal rule were an artefact of withdrawal, it should have failed in exactly these two cohorts, and it did not.
Where this book lands. The rule survives the circularity. The certainty attached to it does not. Specifically:
- Use the multimodal rule — never a single test. This is the one conclusion supported by cohorts that are not circular.
- Do not quote 100% specificity, or any specificity, to a family. Say instead that two or more concordant unfavourable findings have not produced a recovery in the cohorts where recovery was allowed to happen — which is true, and is a different and more honest claim.
- Treat the confidence intervals as the honest quantity. Moseby-Knappe's specificity was 100% with a lower bound of 98.8%. In a hundred patients that lower bound permits one recovery. That is the number to carry, not the point estimate.
- Note what the circularity does NOT bias. It inflates specificity for poor outcome. It does not inflate the good-outcome predictors of §23.10, because nobody withdraws treatment on the basis of a favourable sign. The good-outcome literature is, epistemically, the cleaner half of this field — and it is the half that is routinely ignored.
What would settle it. More cohorts like ProNeCA and Götze, in ECPR patients specifically. None exists. [VERIFICATION REQUIRED] — both no-withdrawal cohorts are known here only through structured abstracts.
Clinical pearl · The self-fulfilling prophecy has an ECPR-specific aggravating factor
In conventional post-arrest care, withdrawal happens when the family and team agree the patient will not recover. In ECPR it can happen for an additional reason: the circuit is consuming a bed, a perfusionist, a machine and a blood bank, and somebody is waiting for it. Chapter 77 owns that argument. Its relevance here is that it applies pressure to prognosticate earlier and more confidently than the evidence supports, and that pressure is invisible in the medical record.
If a prognostic conversation is being held earlier than the protocol says, name the reason out loud. If the reason is resource, it is a Chapter 77 conversation wearing a Chapter 23 costume.
23.3 What the circuit does to the instruments
Every prognostic tool in the standard algorithm was validated in patients with a native circulation and no extracorporeal support. Four things change when a circuit is running.
What changes | Mechanism | Consequence for prediction |
Sedation is deeper and longer | Cannula safety, ventilator synchrony, shivering during temperature control — and drug sequestration in the circuit means plasma concentrations of fentanyl, propofol, dexmedetomidine and midazolam are all significantly reduced, so doses are higher (Chapter 22 §22.12, Chapter 49) | The clinical examination is unavailable for longer, and the 72-hour convention no longer marks the same biological moment. This is the single largest practical difference |
Haemolysis | Shear across the pump and membrane. Neuron-specific enolase is present in erythrocytes | Neuron-specific enolase may be inflated by a non-neurological source — the subject of §23.4 |
Magnetic resonance imaging is unavailable | The Red Book states it plainly: brain MRI "is a superior imaging modality but is incompatible with ECMO circuitry" | The most sensitive modality in the standard algorithm is simply removed, leaving computed tomography, which is markedly less sensitive (§23.8) |
Electrical and physiological noise | Pump, heat exchanger, non-pulsatile flow, continuous renal replacement | Degrades electroencephalography and evoked potentials; one narrative review calls them "unreliable" in this setting |
Evidence · Low certainty · Do the standard tools still work on ECMO?
A single-centre registry compared 397 patients without ECMO against 50 with, in whom venoarterial support had been initiated for refractory arrest with a no-flow under 5 minutes and a low-flow of 60 minutes or less.
The ECMO patients did worse — 74% with a poor functional outcome against 59% (P = 0.040) — and a non-significantly higher mortality, 60% against 47% (P = 0.080).
The prognostic items behaved similarly in both groups. Pupillary reflex, electroencephalographic background and reactivity, evoked potentials and myoclonus all related to unfavourable outcome in both, and the prevalence of having at least two such items concurrently was the same. The specificity of each item was between 92% and 100% in both groups, and of the combination of at least two items, 99.3% without ECMO and 100% with.
The one exception was neuron-specific enolase, which was higher in ECMO patients (P below 0.001).
And one result should temper the reassurance: the discriminative performance of the combination was 0.822 without ECMO against 0.681 with (P = 0.134). The specificity was preserved; the ability to separate the two groups was not. With fifty ECMO patients the comparison is underpowered and the difference is not significant — but the direction is the one that matters, and the authors' conclusion is explicitly provisional: "pending a prospective assessment on a larger cohort."
Clinical pearl · Specificity preserved, discrimination degraded — and what that means at the bedside
These two findings sound contradictory and are not. Specificity answers: if this test is positive, how often is the patient nonetheless going to recover? That appears to be preserved on ECMO. Discrimination answers: across the whole population, how well does this instrument separate those who will recover from those who will not? That appears to be worse.
The practical translation is precise. An unfavourable finding on ECMO can still be trusted to mean roughly what it means off ECMO. But the proportion of patients in whom the instruments give you no useful signal at all is larger, and those patients — the indeterminate group — are where the real work of this chapter happens. In the largest prospective cohort of conventional patients, the indeterminate group was half the population. On ECMO it is likely to be larger.
Plan for indeterminacy as the expected result, not the failure mode.
23.4 Neuron-specific enolase — the most useful test, and the one the circuit most disturbs
Neuron-specific enolase deserves a section of its own for two reasons: it is repeatedly the best-performing single instrument in this population, and it is the one whose published threshold is least likely to be correct on ECMO.
Evidence · Low certainty · Enolase is the strongest single marker in ECPR, and it is not degraded by the circuit in the way the others are
A sub-analysis of the PRAGUE-OHCA trial compared 164 patients resuscitated conventionally against 92 treated with ECPR, with a primary outcome of favourable neurological survival at 180 days (29.3% versus 21.7%, P = 0.191).
Biomarker trajectories were similar in both groups, with better values in patients who recovered. Procalcitonin was higher in the ECPR group at 24 to 72 hours (all P below 0.01). Neuron-specific enolase, C-reactive protein and the neutrophil-to-lymphocyte ratio did not differ between groups. Platelets, D-dimers and fibrinogen were lower in the ECPR group at 24 to 72 hours (all P below 0.001).
On discrimination, enolase was the best performer in both groups — areas under the curve at 24, 48 and 72 hours of 0.89, 0.90 and 0.91 conventionally, against 0.78, 0.90 and 0.90 in ECPR. Procalcitonin performed well specifically in ECPR (0.84, 0.87 and 0.86 against 0.72, 0.73 and 0.73 conventionally).
And the finding that governs practice: the optimal cut-off points for both enolase and procalcitonin were HIGHER in ECPR than in conventional resuscitation.
The authors' conclusion is the one this chapter adopts: "Prognostication algorithms should reflect the resuscitation method."
Controversy 2 — Does the circuit inflate neuron-specific enolase, and what threshold should be used?
The question is not academic. European guidance uses a threshold of roughly 60 µg/L at 48 to 72 hours to support a prediction of poor outcome. If the circuit raises enolase by a non-neurological mechanism, applying that threshold on ECMO produces false positives in exactly the direction that kills recoverable patients.
Position A — the circuit inflates it, through haemolysis. Enolase is present in erythrocytes; the pump and membrane shear them. A narrative review of neuroprognostication in ECMO states directly that enolase "is increased as a result of hemolysis." A registry comparison found enolase significantly higher in ECMO patients than in patients without (P below 0.001) while every other prognostic item was similar — precisely the pattern a circuit-specific artefact would produce.
Position B — the level is not inflated, but the threshold is different. The PRAGUE-OHCA sub-analysis found no difference in enolase levels between conventional and ECPR patients, in a directly comparable cohort from the same trial — yet found the optimal cut-off higher in ECPR. Those two findings together are incompatible with simple haemolytic inflation: if the circuit merely added enolase, levels would differ. A more coherent reading is that the ECPR population has a larger underlying neurological insult — longer low-flow, deeper ischaemia — so the level that discriminates recoverable from unrecoverable sits higher. The single dedicated ECMO study points the same way: in post-arrest ECMO patients, enolase at 48 hours discriminated best at a cut-off of 70 µg/L, with an area under the curve of 0.87, validated in a second cohort.
Why both positions can be partly right. Haemolysis is not constant. A circuit with a failing oxygenator or a high-resistance cannula haemolyses far more than a well-running one (Chapter 37). Enolase on ECMO is therefore a marker with a variable, measurable contaminant — and the contaminant can be checked, because haemolysis leaves other fingerprints.
Where this book lands.
- Do not apply the 60 µg/L threshold unmodified to a patient on ECMO. Every line of evidence — inflation, population severity, or both — points to the correct threshold being higher, and no line points to it being the same.
- Use the trajectory, not the value. Serial measurement outperformed single values in the one ECMO-specific study, where specificity was highest using all three time points. A falling enolase is informative in a way that no single number is, and a falling value cannot be produced by ongoing haemolysis.
- Interpret enolase against a haemolysis panel. If free haemoglobin is rising and the circuit is struggling, the enolase is contaminated and should be down-weighted rather than discarded.
- Interpret it against the electroencephalogram. The threshold is context-dependent even in conventional patients — see below.
What would settle it. A study measuring enolase and free haemoglobin simultaneously and serially in ECMO patients, with a neurological endpoint. None was found. That study is the single highest-value piece of missing work in this chapter.
[VERIFICATION REQUIRED] — all four studies in this controversy are known only through structured abstracts.
Clinical pearl · The threshold is conditional on the electroencephalogram, and this is the general form of the ECPR problem
A 155-patient study asked whether the enolase threshold should vary with the electroencephalographic pattern. It should, and substantially:
- With a malignant background, enolase above 45.2 µg/L predicted poor outcome at 100% specificity with 70.8% sensitivity, against 66% for the recommended 60 µg/L cut-off.
- With electrographic seizures, the figure was 53.5 µg/L, sensitivity 77.7%.
- With a benign background, 78.2 µg/L was needed to reach 100% specificity — the standard cut-off achieved only 94%.
Read that last line carefully. In a patient whose electroencephalogram looks benign, the conventional enolase threshold generates false positives. The same number means different things in different contexts.
This is exactly the structure of the ECMO problem in §23.4, arriving from a different direction. A biomarker threshold is not a property of the biomarker; it is a property of the biomarker in a population. ECMO is another population, and its thresholds have not been established.
23.5 The clinical examination, and the ECPR-specific obstacle
The clinical examination remains the most specific individual instrument in every cohort, and on ECMO it is the one most often unavailable when it is wanted.
Finding | When it is interpretable | What it is worth |
Bilaterally absent pupillary and corneal reflexes | After day 4 from return of circulation, per the systematic review — not before | 0% false-positive rate in most studies; specificity 100% in the prospective multicentre cohort. The most robust single item there is |
Quantitative pupillometry — Neurological Pupil index at or below 2 | First 48–72 hours | 99% specificity, 34% sensitivity in a 442-patient bicentric study. Correlates with malignant electroencephalography, absent N20 and high enolase. Notably, qualitative pupillary reflex had HIGHER sensitivity than quantitative measures (38.9% versus 23.6% and 19.0%, P below 0.001) with comparable specificity |
Motor response — extensor or absent | At or beyond 72 hours, and only off sedation | A gate for the algorithm rather than a predictor in itself. Prognosticating irrespective of motor score performed as well in one reanalysis |
Status myoclonus | Days 1–3 | A useful but less robust predictor. See the warning below |
Coma Recovery Scale–Revised, best score | Serial, over days | Improved discrimination from AUC 0.888 to 0.925 in the no-withdrawal cohort — the only instrument shown to add to the standard set in a cohort free of circularity |
Danger · On ECMO, "day 3" and "72 hours" are not the same moment they are in a conventional patient
The standard algorithm assumes a patient whose sedation has been stopped and cleared. The ECPR patient is frequently still deeply sedated at 72 hours because the circuit demands it, the temperature target demands it, and the drugs are being given at higher doses to overcome sequestration — so they take longer to clear once stopped.
The commonest error in this chapter is to perform the 72-hour examination on a patient who is still sedated, find an absent motor response, and enter it into the multimodal count. That is not a prognostic finding; it is a pharmacological one.
Two protections. First, document the sedation state alongside every prognostic examination, so that a later reader can see what the examination was worth. Second, the clock starts when sedation clears, not at a fixed hour after cannulation — and say so explicitly to the family, because "we will know more in three days" becomes untrue and erodes trust when day three arrives and nothing can be said.
Pitfall · Myoclonus is not the death sentence it is treated as
In a post-hoc analysis of 157 comatose post-arrest patients with rhythmic or periodic patterns, 98 (63%) had myoclonus. It was associated with a discontinuous or suppressed background (48% continuous with myoclonus versus 75% without, OR 0.31, 0.16–0.64) and with earlier onset of the abnormal pattern.
Myoclonus was associated with poor outcome — but not invariably: 96% poor with myoclonus versus 82% without (P = 0.004). The authors' conclusion is unambiguous: myoclonus "cannot predict a poor outcome without false positives."
Four per cent is not zero, and in a bimodal population the four per cent is the patient who walks out. Myoclonus belongs in the multimodal count as one item among several. It is not, on its own, a reason to stop.
Clinical pearl · Two examination findings that belong to the resuscitation, not to the ICU
Chapter 20 §20.5.4 established that signs of life during cardiopulmonary resuscitation are the strongest pre-cannulation variable there is. Those observations do not stop being informative once the patient reaches the intensive care unit, and they are frequently lost because nobody wrote them down.
From the ECPR textbook's review of arrival findings: in one 52-patient series, spontaneous breathing was present in 75% of those with a favourable outcome against 11% of those with a poor one, and among 26 patients with a pupil diameter greater than 6 mm, none survived. In an 85-patient analysis, only Glasgow Coma Score above 3 at arrival, pupil diameter at arrival, and the no-flow interval were independently associated with outcome — with a pupillary reflex present in 21% of favourable against 5.6% of poor outcomes (P = 0.049).
But the same source supplies the caution, and it is the right one: mean pupil diameter was 4.1 mm (SD 1.1) against 5.2 mm (SD 1.2), which the authors call "a difficult distinguishing finding due to the small difference and interrater variability."
Record the arrival examination in the notes as a prognostic observation, not merely as a resuscitation detail — and weight it as one item, not as an answer.
23.6 Electroencephalography — one instrument doing two unrelated jobs
Chapter 22 used the electroencephalogram as a seizure monitor. This chapter uses it as a prognostic instrument. These are different questions, they are answered by different features of the same recording, and they can point in opposite directions in the same patient.
How much is there, and how much is seen
Source | Population | Finding |
ELSO Registry, 2013–2023 | 13,783 ECPR patients | Seizures in 3% (346). Of these, 204 (59%) detected clinically, 193 (56%) by electroencephalography, 51 (15%) by both. Cases rose from 442 to 1,123 a year while reported seizure incidence fell from 5% to 3% (P = 0.02) |
Meta-analysis, 23 studies | 40,420 ECMO patients | Pooled seizure prevalence 3.0% overall; venovenous 2.0%, venoarterial 3.5%, ECPR 4.9% (1.3–17.2) |
Single centre, electroencephalography actually applied | 92 ECMO patients monitored (83% venoarterial, 28% ECPR), average 54 hours of recording | 58 of 92 (63%) had epileptiform activity or ictal-interictal continuum patterns — 3% non-convulsive status epilepticus, 36% generalised periodic discharges, 5% lateralised periodic discharges |
Practice survey, all 36 SAVE-J II institutions (Chapter 22 §22.9) | Japanese ECPR centres | Electroencephalography routinely applied in only 13.9% |
Danger · A registry seizure rate of 3% and a monitored rate of 63% are the same population
The registry figure is what gets reported when electroencephalography is applied to about one patient in seven. The 63% figure is what is found when it is applied systematically — and it includes patterns on the ictal-interictal continuum that a clinical examination cannot detect at all.
Two further observations sharpen this. In the registry, 41% of reported seizures were detected clinically but NOT electrographically, and 44% electrographically but not clinically — the two methods barely overlap, with only 15% detected by both. And the registry's reported seizure incidence FELL from 5% to 3% over a decade during which ECPR volume more than doubled, which the authors themselves attribute to possible underreporting and evolving practice rather than to a real decline.
A unit that does not monitor does not have a low seizure rate. It has an unmeasured one.
What the electroencephalogram is worth for prognosis in ECPR
Evidence · Low certainty · The two ECPR-specific studies
Intermittent electroencephalography within 96 hours, 69 adult ECPR patients: 17 (24.6%) had a favourable outcome. Malignant patterns occurred in 73.1% of poor-outcome patients against 5.9% of favourable ones (P below 0.001). Every patient with a highly malignant pattern — 43.5% of the cohort — had a poor outcome. On multivariable analysis, malignant patterns and duration of resuscitation were independently associated with poor outcome.
Amplitude-integrated electroencephalography within 24 hours, 22 ECPR patients with temperature control, 19 with data in the first day. Six patients who showed a continuous normal voltage pattern within 24 hours regained consciousness and had a good outcome, except one who died of a haemorrhagic complication. Patients with persistent flat trace or burst suppression at any time did not regain consciousness. Continuous normal voltage predicted good outcome at six months with 100% sensitivity, 93% specificity, 83% positive predictive value and 100% negative predictive value — and all four indices were 100% for the endpoint of wakefulness.
Twenty-two patients. The second study is the most striking good-outcome result in the ECPR literature and it is far too small to act on alone. It is recorded because it points the right way and because §23.10 needs it.
Pitfall · Post-anoxic status epilepticus is not automatically fatal, and treating it as such is a self-fulfilling prophecy in miniature
This is the place where bedside belief diverges most sharply from the data.
- In a pooled analysis of 274 patients with definite or possible status epilepticus after arrest, 24 (8.8%) had a good recovery. After excluding those with two or more poor-outcome criteria, 94 patients remained and 25% had a good outcome. Non-motor semiology and cessation of the status were associated with recovery; none of the patients with a good outcome had a suppressed or burst-suppression background before onset.
- In a trial substudy of 191 monitored patients, 52 (27%) developed status epilepticus. A "favourable profile" — late onset beyond 24 hours, arising from a continuous background, with no established poor-outcome indicators — was present in 20 (38%), of whom 12 survived to six months and 8 had a good outcome, tripling the probability of recovery. Every patient lacking the favourable profile had a poor outcome. At 24 months all 12 survivors were alive, seven with good functional outcome and most with only mild cognitive impairment.
- In a 166-patient cohort where refractory status was treated aggressively whenever multimodal indicators were not unfavourable, survival and good outcome were 72.4% and 71.1% for a benign background; 54.3% and 44.4% for refractory status epilepticus; 15.4% and 0% for generalised periodic discharges; and 2.4% and 0% for a malignant non-epileptiform pattern.
- A review of non-convulsive status after arrest found a favourable outcome in about one patient in four, and concluded that clinicians "should avoid equating NCSE after ROSC with poor prognosis."
The pattern across all four: what matters is not the status epilepticus but the background it arises from, and whether other unfavourable markers accompany it. Status arising late from a continuous background, in a patient without two other poor-outcome findings, carries a meaningful chance of recovery and is worth treating. Status arising early from a suppressed background does not, and no patient in these series recovered from that combination.
Chapter 51 owns the drugs. This chapter owns the inference, and the inference is: ask what the background was doing before the status started.
23.7 Somatosensory evoked potentials
The most specific electrophysiological instrument, and the one most degraded by the environment it is used in.
Finding | Performance | Source |
Bilaterally absent N20 | Specificity 100% for poor outcome in the prospective multicentre cohort; 57.4% sensitivity at 100% specificity in the no-withdrawal cohort within 24 hours | ProNeCA; AfterROSC |
Absent, or present with low amplitude (0.41 µV or below) | 100% specific (96–100%); adds only marginally to absence alone — sensitivity 47% versus 46% | 158-patient registry |
Low-voltage N20-baseline (under 0.88 µV) or N20-P25 (under 1 µV) | Specificity 94% and 93%, sensitivity 50% and 66%. Combining "absent or low voltage" nearly doubled sensitivity over absence alone, 58% versus 30% (P = 0.002) | 82-patient prospective study |
Ultra-early assessment within 6 hours, including middle-latency N70 | Bilaterally absent N20: 100% specificity (89–100), 67% sensitivity (48–82). Adding low-amplitude, prolonged N20 without N70 raised sensitivity to 93% (79–99) without loss of specificity. Evoked potentials outperformed every other early index for both good and poor outcome | 65-patient prospective study |
Danger · The evidence base for evoked potentials contains no ECPR patients worth the name
Every row of that table comes from patients with a native circulation. The only direct ECMO comparison found evoked potentials to perform similarly — in 50 patients. A narrative review of neuroprognostication on ECMO states flatly that evoked potentials are "unreliable" in this setting, without quantifying it.
The mechanism of the concern is electrical: a centrifugal pump, a heat exchanger and a renal replacement circuit sit within a metre of the recording electrodes, and the N20 is a sub-microvolt signal.
This book's position: an absent N20 recorded by an experienced neurophysiologist who is satisfied the recording is technically adequate retains its usual meaning. An absent N20 on a noisy recording means nothing at all, and the distinction is not one the requesting clinician can make. Insist on an explicit technical-adequacy statement in the report, and treat its absence as an uninterpretable study rather than a negative one.
23.8 Brain imaging — and the ECPR finding that should change practice
Computed tomography is the only structural imaging available on ECMO, and the grey-to-white matter ratio is the quantity extracted from it. The general literature is favourable. The ECPR literature is not, and the discrepancy matters more than anything else in this section.
Evidence · Low certainty · The grey-white matter ratio in the general post-arrest population
Pooled across 42 studies and 8,104 patients: manually computed ratio, area under the curve 0.77 (0.73–0.81), sensitivity 0.55, specificity 0.96; automated computation 0.84 (0.81–0.87), sensitivity 0.53, specificity 0.95.
Timing matters and moves only sensitivity. A ratio below 1.10 was 100% specific irrespective of timing in a 195-patient cohort, but sensitivity rose from 12% (7–20) on scans before 6 hours to 48% (38–58) after 24 hours. In 2,204 patients, sensitivity peaked around 25% after five hours while the false-positive rate stayed below 5% at every time point. Serial imaging in 123 patients gave areas under the curve of 0.72 before 24 hours, 0.78 at 24–96 hours and 0.81 at 96–168 hours, with the ratio falling over time in poor-outcome patients and remaining stable in good-outcome ones.
The summary for a conventional patient: a low ratio early is trustworthy and rare; a normal ratio early means very little; and the scan gets more informative as it gets later.
Danger · In the largest ECPR imaging study ever performed, the grey-white matter ratio barely works
A secondary analysis of the SAVE-J II registry examined head computed tomography within 24 hours in 1,146 out-of-hospital arrest patients treated with ECPR — an order of magnitude larger than any other ECPR imaging study.
Poor neurological outcome at 30 days, by average ratio band:
- 1.00–1.09: 94.6%
- 1.10–1.19: 87.8%
- 1.20–1.29: 78.5%
- 1.30–1.39: 70.3%
Adjusted odds ratios against the 1.30–1.39 reference: 10.01 (3.58–27.99), 4.83 (2.31–10.12) and 2.16 (1.02–4.55).
And then the two numbers that matter. The area under the curve was 0.628 (0.59–0.66). The threshold achieving 100% specificity was 1.005 — with a sensitivity of 0.1%.
Read the band table again. Even in the best band, 70.3% had a poor outcome. A reassuring grey-white matter ratio after ECPR is not reassuring; it is the base rate of the population showing through.
The authors' conclusion is the correct one and this book adopts it verbatim in substance: early neuroprognostication depending on the grey-white matter ratio may not be sufficient after ECPR and requires a multimodal approach.
Physiology · Why the ratio may transfer badly to ECPR, and a finding that supports it
A secondary analysis of the Prague study examined 45 patients scanned within 36 hours and found something specific: patients treated with ECPR had lower attenuation in the centrum semiovale than those who were not — 28.3 (SD 2.7) against 31.0 (SD 2.8), P = 0.003.
The ratio is grey matter divided by white matter. If the circuit, the haemodilution of priming, or the oedema of prolonged low flow systematically alters white-matter attenuation, the denominator moves and every threshold derived in a non-ECPR population shifts with it. That is the imaging counterpart of the enolase problem in §23.4, and it points the same way: the number is not wrong, the threshold is borrowed.
The same analysis produced a genuinely counter-intuitive result. Both decreased AND increased grey-to-white differentiation indicated poor outcome. Favourable patients clustered in a narrow band — caudate 1.18 to 1.30, putamen 1.20 to 1.33 — which separated them from poor-outcome patients with sensitivity 78% and 66%, specificity 85% and 100%, and areas under the curve 0.86 and 0.77. Intracranial oedema was present in 16 patients and occurred exclusively in the poor-outcome group.
A ratio that is too high is not a normal ratio. This is a single 45-patient analysis and is flagged as such, but it is a reason to treat an unexpectedly high value as a finding rather than as reassurance.
Clinical pearl · Contrast from the angiogram does not invalidate the scan on a dual-energy scanner
A practical obstacle in ECPR is that the patient has usually just had coronary angiography (Chapter 22 §22.8), and residual contrast is assumed to make the grey-white matter ratio uninterpretable.
In 25 comatose patients scanned with dual-energy computed tomography at a median of 2.08 hours after cannulation, the ratio retained prognostic value despite residual contrast, with areas under the curve of 0.85 for the cerebrum, 0.76 overall and 0.70 for the basal ganglia against in-hospital death.
Twenty-five patients, an abstract, and dual-energy scanners are not universal. But the practical instruction is worth having: do not cancel the scan because the patient has had contrast — ask the radiologist whether dual-energy reconstruction is available.
Clinical pearl · Do not read the computed tomogram as a single number
The one ECPR-specific study that combined imaging features did substantially better than any single measure. In 42 patients scanned within 48 hours, areas under the curve were 0.792 (0.639–0.901) for the basal-ganglia ratio, 0.745 (0.587–0.867) for optic nerve sheath diameter (5.57 mm in good-outcome patients against 6.07 mm in poor) and 0.817 (0.682–0.952) for loss of grey-white boundary or sulcal effacement — and 0.904 (0.773–0.973) for the three combined, significantly better than any component.
Ask for the qualitative reading as well as the ratio: loss of the grey-white boundary, sulcal effacement, cistern effacement, and oedema. These outperformed the ratio individually in that cohort, and are what a radiologist reports anyway.
23.9 When may prediction begin?
The conventional answer is 72 hours, off sedation, using the four-step algorithm. The ECPR-specific pressures push in both directions: sedation pushes the moment later, while bed pressure, family distress and the sheer visibility of the circuit push it earlier.
Controversy 3 — Can outcome be predicted earlier than 72 hours after ECPR?
The case for earlier prediction. Several instruments carry their full specificity within the first day. In the no-withdrawal cohort, evoked potentials, imaging and electroencephalography each reached 100% specificity within 24 hours, with sensitivities of 57.4%, 48.8% and 34.5%, and at least one unfavourable pattern present in 74.4% of poor-outcome patients. In an ultra-early study, evoked potentials recorded within 6 hours achieved 100% specificity with 67% sensitivity, rising to 93% when amplitude and middle-latency components were included. An amplitude-integrated electroencephalogram in ECPR patients separated outcomes within 24 hours. And a purpose-built ECPR score claims to work at 72 hours with a specificity of 100% (see below).
The case against. Every early instrument shares one property: high specificity, low sensitivity. They identify a minority of the patients who will do badly and say nothing about the rest — and the rest are the majority. Meanwhile the ECPR patient is the patient least likely to be assessable early, because sedation is deeper and longer. The asymmetry of error is extreme and one-directional: acting early on an unfavourable finding risks ending a recoverable life, while waiting risks a few more days of support in a patient who was going to die anyway. In a population where roughly four in five survivors do well, the cost of a false positive is enormous.
Where this book lands. Early testing, late deciding.
- Test early. An early unfavourable result is informative, it is specific, and the imaging and electrophysiology get harder to obtain later, not easier.
- Do not act early on a single unfavourable result, and never on an early clinical examination. The one thing every no-withdrawal cohort agrees on is that the combination is what carries no false positives.
- Treat early results as accumulating rather than deciding. The Coma Recovery Scale, repeated assessments, and enolase trajectory all improved prediction specifically because they were serial.
- Do not predict at 72 hours because it is 72 hours. The clock that matters is the one that starts when sedation clears (§23.5).
The one exception worth naming: an early scan showing established infarction, herniation or massive haemorrhage is not prognostication — it is diagnosis, and it belongs to Chapter 24 immediately.
Evidence · Very low certainty · The only prognostic score built for ECPR patients after cannulation
A single-centre study of 361 out-of-hospital arrest patients treated with ECPR between 2015 and 2023 derived and internally validated an eight-point score at 72 hours. Of the cohort, 170 (47%) survived, 85% of them with a favourable neurological outcome.
The eight retained predictors mix resuscitation variables with neurological ones: absence of witnessed arrest; absence of bystander resuscitation; downtime of 7 minutes or more; lactate of 9 mmol/L or more; low-flow of 55 minutes or more; enolase above three times the upper limit of normal; a highly malignant electroencephalogram; computed tomography showing anoxic injury; and absence of brainstem reflexes.
Discrimination was 0.87 (0.82–0.92) in derivation and 0.86 (0.85–0.92) in validation. A score of 6 or more predicted poor outcome with sensitivity about 21%, specificity 100%, false-positive rate 0%. Scores of 4 or more and 6 or more carried an 89% and 97% probability of poor outcome.
This is a conference abstract, internally validated by splitting one centre's cohort, and it has not been externally validated anywhere.
Danger · Two problems with that score, printed and not resolved
First, the calibration statistic is reported as supporting the model when it does not. The abstract reports "calibration (goodness of fit p=0.04)." A Hosmer-Lemeshow goodness-of-fit P value below 0.05 indicates that the model's predicted probabilities differ significantly from observed frequencies — that is, calibration is poor. It is presented alongside the discrimination figures as though it were a supporting result. The discrepancy is printed here and is not corrected; it may be a reporting convention this book is misreading, or the full paper may say otherwise. But a score whose stated purpose is to give "estimates of probability of poor neurological outcome" depends on calibration more than on discrimination, and the one calibration statistic offered does not support it.
Second, the validation interval is asymmetric in a way that suggests a transcription problem. Derivation is reported as 0.87 (0.82–0.92) and validation as 0.86 (0.85–0.92) — a point estimate sitting almost exactly on the lower bound of its own interval. Both figures are printed as they appear and neither is reinterpreted.
The practical consequence. The score's construction is sensible and its variable list is a reasonable checklist of what to gather by 72 hours. But a score offering a 97% probability of poor outcome, on a 100%-specificity threshold with 21% sensitivity, derived and validated in one centre, with a calibration statistic that argues against it, is not an instrument for ending a life. Use the variable list. Do not use the probability.
23.10 Predicting a good outcome — the neglected and cleaner half
Almost every instrument in this chapter was developed to predict a poor outcome. That emphasis has two consequences: it leaves the indeterminate group — roughly half the population — with nothing, and it means the literature's cleanest evidence is the least used.
Clinical pearl · Good-outcome prediction is not contaminated by the self-fulfilling prophecy
§23.2 established that withdrawal on the basis of a test inflates that test's specificity for poor outcome. Nothing analogous happens to favourable predictors, because no unit withdraws treatment because a sign is encouraging. A favourable predictor that is followed by recovery is a clean observation.
The good-outcome literature is therefore the epistemically stronger half of this field, and it is routinely ignored — because the question clinicians ask themselves at 72 hours is "may I stop?" rather than "should I continue?"
Favourable predictor | Performance | Population |
Two concordant favourable signs in patients whose outcome was indeterminate after the standard algorithm | Specificity 99% (96–100), sensitivity 23% (11–38) for good outcome | 162 of 337 prospective multicentre patients — half the cohort was indeterminate |
Falling enolase between 24 and 72 hours | Positive predictive value 82%, sensitivity 81% | 215 patients |
Benign electroencephalographic background | Positive predictive value 72%, sensitivity 94% | 215 patients |
N20 amplitude above 0.85 µV bilaterally | Positive predictive value 75%, sensitivity 100% | 215 patients |
Combination — enolase under 51.5 µg/L at 48 h, falling trend, benign background | Positive predictive value 96%, sensitivity 76% | 215 patients |
Six-item multimodal favourable score, 4 or more of 6 | Sensitivity 97.5% (92.9–99.5), accuracy 77.5%, area under the curve 0.88 (0.85–0.91), externally validated | 499 patients |
Continuous normal voltage on amplitude-integrated electroencephalography within 24 hours — the only ECPR-specific result | Sensitivity 100%, specificity 93%, negative predictive value 100% for good outcome at 6 months | 19 ECPR patients |
High-amplitude N20 with preserved N70 at 6 hours | Sensitivity 94% (79–99), specificity 100% (89–100) | 65 patients |
Clinical pearl · The three favourable findings to actively look for in an ECPR patient
- A continuous electroencephalographic background. It is the single most consistent favourable sign across every study above, it appears early, and it is the feature that determines whether status epilepticus means anything (§23.6).
- A falling enolase. §23.4 argued that the absolute value is unreliable on a circuit. The trajectory is not, and a falling value cannot be generated by ongoing haemolysis. Two measurements are worth more than one perfectly timed one.
- A present, good-amplitude N20 — where the recording is technically adequate (§23.7).
Write these into the prognostication note explicitly as favourable findings. A note listing only unfavourable items reads as a case for stopping even when it was not meant as one, and that is how a multimodal assessment quietly becomes a justification.
23.11 Running a prognostication process
Everything above resolves into a process rather than a threshold. The steps are ordered; the ordering is the content.
Step | What to do | Why it is in this order |
1. Exclude confounders | Document sedation and neuromuscular blockade, temperature, metabolic state, and whether seizures are being suppressed pharmacologically | Every prognostic item is invalid in their presence, and on ECMO sedation is the confounder that persists longest (§23.5) |
2. Look for a structural cause | Computed tomography, read qualitatively as well as quantitatively | Herniation or a large infarct is a diagnosis, not a prediction, and it belongs to Chapter 24 immediately |
3. Gather early | Electroencephalography, evoked potentials, serial enolase from 24 h, pupillometry, imaging | Specificity is available early even when the decision is not (§23.9) |
4. Start the clock at sedation clearance | Not at a fixed hour after cannulation. Tell the family this at the outset | Otherwise the day-three promise is broken on day three |
5. Count both directions | Tally unfavourable and favourable findings, in the same note | The good-outcome literature is the cleaner half and is routinely omitted (§23.10) |
6. Require two or more concordant items | Never predict on one. Never predict on a single examination finding | The only conclusion the no-withdrawal cohorts support |
7. Repeat | Serial Coma Recovery Scale, serial enolase, repeat electroencephalography | Repeated assessment reduced false positives in the no-withdrawal cohort and added most to discrimination |
8. Separate prediction from decision | The person writing the prognosis states a probability and stops. Chapter 24 owns what follows | A prediction made in order to justify a decision has stopped being a prediction (§23.2) |
Clinical pearl · What to actually say to a family on day one
Three things are defensible on day one, and all three are usually omitted.
- "Most people who survive this do well." Roughly four in five ECPR survivors reach a good neurological outcome. This is the base rate (§23.1) and it is the single most useful thing a family can hold.
- "The question right now is whether he survives, not what state he survives in." True in this population, and it directs the family's attention to the question that is actually open.
- "We cannot assess his brain while he is this sedated, and that will take longer than usual because of the machine. I will tell you when the assessment becomes possible." This replaces the day-three promise that cannot be kept.
None of these is a prediction. All three are honest, and together they buy the time that §23.9 argues for.
23.12 The errors that recur
Nine errors account for most of what goes wrong in this domain. Seven of them are errors of confidence and two are errors of omission, and the omissions are the ones that harm survivors.
Error | Why it is made | What to do instead |
Quoting a specificity of 100% to a family | The number is printed in the guideline and in the paper. It is arithmetically correct within the study that produced it | State that the finding makes a poor outcome highly likely and that the confidence interval and the withdrawal problem both prevent certainty. The two no-withdrawal cohorts reproduced the rule; neither reproduced the certainty (§23.2) |
Starting the prognostic clock at cannulation or at arrest | Seventy-two hours is the number everyone remembers, and it is the number promised to the family on day one | Start it when sedation has demonstrably cleared and the examination is interpretable. On ECMO that is later, sometimes much later, and the family must be told so at the outset rather than on day three (§23.5, §23.9) |
Applying the 60 µg/L enolase threshold unchanged on ECMO | It is the guideline threshold, and it is what the laboratory report flags | Treat the published cut-off as a lower bound on ECMO, insist on serial values and read the trajectory, and demand a haemolysis index with every sample. The one ECPR study that derived its own threshold found it higher, not lower (§23.4) |
Reading a single enolase value | One number is easier to act on than three | Measure at 24, 48 and 72 hours. A falling trajectory was the single best favourable predictor in the good-outcome literature, and a trajectory is immune to the one-off haemolytic spike in a way that a point value is not (§23.4, §23.10) |
Predicting on the clinical examination alone | It is free, immediate and at the bedside | Use it to raise or lower suspicion and never to conclude. The ECPR-specific series that reported a mean pupil diameter of 4.1 mm in survivors against 5.2 mm in non-survivors called it, in the authors' own words, a difficult distinguishing finding (§23.5) |
Treating an abnormal electroencephalogram as a prediction | The word malignant carries prognostic weight that the recording does not always earn | Separate the two jobs the instrument does. Highly malignant patterns carry prognostic weight. Status epilepticus does not, on its own: a quarter of patients with post-anoxic status epilepticus and no other poor-outcome feature recovered well, and none of those who recovered had a suppressed or burst-suppressed background before onset (§23.6) |
Concluding from a normal-looking computed tomogram in the first hours | The scan is done early for a different reason — to find the cause — and it is there to be read | Early imaging has high specificity and almost no sensitivity. A normal early scan excludes a catastrophe and excludes nothing else. Its sensitivity rises with time; its usefulness for prediction does not exist at hour two (§23.8) |
Omitting the favourable findings from the note | The literature, the guidelines and the algorithm are all built to identify poor outcome. There is no checklist for the other direction | Count both directions in the same note. Two concordant favourable items carried a specificity of 99% for a good outcome in the largest validation cohort, and a combination of favourable enolase trajectory, benign background and preserved evoked-potential amplitude reached a positive predictive value of 96% (§23.10) |
Treating indeterminacy as a failure of the assessment | It feels like an admission that the workup was inadequate, and it leaves the family without an answer | Indeterminate is the expected result on ECMO and was the result in half of the largest prospective algorithm cohort. The discriminative power of a multimodal combination fell from an area under the curve of 0.822 without a circuit to 0.681 with one. Saying so is the accurate answer, not the absence of one (§23.3, §23.9) |
Danger · The three sentences that should never be said about an ECPR patient
- "The test is 100% specific, so there is no chance of recovery." The interval is not zero, the study withdrew treatment on the basis of the test, and on a circuit the test performs differently from the population in which it was validated.
- "He has had seventy-two hours, so we can assess him now." Not if he is still sedated. On ECMO the drug is sequestered, the volume of distribution is enlarged and clearance may be impaired. The clock runs from sedation clearance, not from the calendar.
- "The scan is normal, so the brain is fine." Early computed tomography after ECPR has a sensitivity that barely exceeds a tenth in the first hours. It rules a catastrophe in. It rules nothing out.
23.13 Key points
- Roughly one in six ECPR patients reaches a good neurological outcome overall, and roughly four in five survivors do. The severely disabled survivor is the rarest of the three destinations, not the commonest. A family asking whether their relative will be left severely disabled is asking about the least likely outcome, and the honest answer on day one is that the open question is survival, not the state of survival (§23.1).
- Every specificity figure in this chapter was generated in a healthcare system that withdraws life-sustaining treatment on the basis of the test being evaluated. The circularity is not a technical caveat; it is the central epistemic fact of the field. Two cohorts escaped it — one prospective Italian multicentre study in which no treatment was withdrawn, and one German multicentre study that excluded withdrawal deaths. Both reproduced the rule and neither reproduced the certainty, and the Italian cohort left 68 patients alive in a persistent vegetative state (§23.2).
- The circuit degrades the instruments without abolishing them. Sedation clears later, haemolysis may inflate the biomarker, magnetic resonance imaging is unavailable while the patient is cannulated, and the discriminative power of the multimodal combination falls substantially. Specificity is largely preserved; sensitivity and discrimination are not. The predictable consequence is more indeterminate assessments, and indeterminacy is the expected result rather than the failure mode (§23.3).
- Neuron-specific enolase remains the most useful single test and the one the circuit most disturbs. The direction of the disturbance is disputed: one ECMO comparison found values significantly higher on a circuit, one ECPR sub-analysis found no difference between resuscitation methods, and a narrative review attributes elevation to haemolysis. What all three support is the same practical rule — serial measurement, trajectory over threshold, and a haemolysis index with every sample (§23.4).
- Detection, not incidence, explains most of the variation in the reported numbers. Registry-reported acute brain injury after ECPR is about 16%; an autopsy series found 68%. Registry-reported seizures are 3%; continuous monitoring of an ECMO population found epileptiform or ictal-interictal activity in 63%. Registry seizure incidence fell while case volume doubled. Electroencephalography is routinely applied in about one ECPR institution in seven. The literature measures what is looked for (§23.1, §23.6).
- Somatosensory evoked potentials are the instrument that performs best in both directions early. Bilaterally absent cortical responses within six hours retained full specificity; adding low-amplitude prolonged responses raised sensitivity from about two-thirds to over nine-tenths; and a high-amplitude response with a preserved late potential predicted a good outcome with a sensitivity of 94%. Amplitude carries information that presence-or-absence discards (§23.7).
- The largest imaging study in this population argues against the threshold it was designed to validate. In 1,146 ECPR patients the grey-white matter ratio behaved as a graded risk and not as a switch: the area under the curve was 0.628, and the cut-off that achieved perfect specificity did so at a sensitivity of one patient in a thousand. The authors' own conclusion is that the ratio is insufficient after ECPR. A separate ECPR cohort found favourable patients clustered in a narrow band rather than above a threshold, and found lower centrum semiovale attenuation in ECPR patients specifically (§23.8).
- Prediction may begin early for gathering and late for deciding. The evidence supports collecting every instrument early — specificity is available early — and deferring the conclusion until confounders are excluded. The only published ECPR-specific multimodal score reached an area under the curve of 0.86 on validation, and its own goodness-of-fit statistic indicates poor calibration, which its abstract presents as supporting evidence. It is not ready to drive a decision (§23.9).
- The favourable half of the literature is cleaner than the unfavourable half and is routinely omitted. Predicting a good outcome is not contaminated by the self-fulfilling prophecy in the same direction, because nobody withdraws treatment on the strength of a benign electroencephalogram. Two concordant favourable findings carried a specificity of 99% for recovery; a six-item score reached an area under the curve of 0.88 with external validation. Write the favourable findings into the same note as the unfavourable ones (§23.10).
- Never predict on one item. The single conclusion that survives every no-withdrawal cohort, every ECMO-specific comparison and every guideline is that two or more concordant findings are required, that the assessment must be repeated over days, and that the person producing the prognosis states a probability and stops there. What is then done belongs to Chapter 24 — and a prediction made in order to justify a decision has already stopped being a prediction (§23.2, §23.11).
Cross-references
Inherited from earlier chapters
- Chapter 19 — the ECPR trials and why all three randomised the decision to cannulate. Every outcome figure in §23.1 is downstream of that design.
- Chapter 20 — pre-cannulation selection, low-flow time and the selection scores. The variables that predict outcome before cannulation are largely the same variables that predict it afterwards, which is why §23.9 argues the post-cannulation score adds less than it appears to.
- Chapter 21 — cannulation, and the time cost of a difficult cannulation. Low-flow duration is the strongest modifiable predictor in §23.1.
- Chapter 22 — the first twenty-four hours: the carbon dioxide argument, oxygen, temperature and seizure detection. Chapter 22 owns the instruments as therapy; this chapter owns them as evidence. The hypothermia dissociation recorded there, in which hypothermia was associated with survival but not with favourable neurological outcome, cuts directly against the bimodality argued in §23.1 and is flagged rather than resolved.
Handed forward
- Chapter 24 — ECPR Failure and Withdrawal. Everything this chapter produces is an input to that one. The structural claim of §23.2 is that the two must be performed as separate acts.
- Chapter 28 — Cerebral Monitoring. Near-infrared spectroscopy, transcranial Doppler and the practicalities of applying continuous electroencephalography on a circuit. This chapter uses the outputs; that chapter owns the instruments.
- Chapter 29 — Laboratory Monitoring, and Chapter 37 — Haemolysis. The haemolysis index that §23.4 requires with every enolase sample, and the mechanisms that generate it.
- Chapter 30 — ECMO Data Interpretation. Scores as a class, calibration as distinct from discrimination, and why an area under the curve is not a licence to act. The defect flagged in the ECPR-specific score in §23.9 is an instance of a general problem.
- Chapter 49 — Sedation Pharmacology on ECMO, and Chapter 57 — Sedation and Analgesia. Sequestration, the enlarged volume of distribution and the clearance problem that together move the prognostic clock in §23.5 and §23.9.
- Chapter 51 — Antiepileptic Drugs on ECMO. The treatment of the status epilepticus that §23.6 declines to treat as a prognostic sign.
- Chapter 59 — Mobilization and Rehabilitation. What happens to the four survivors in five who reach a good outcome, and to the fifth who does not.
- Chapter 65 — Accidental Hypothermia, where the prognostic rules genuinely differ and the numbers in §23.1 do not apply.
- Chapter 68 — Paediatric Considerations. The paediatric ECPR neurological figures quoted in §23.1 are held there in full.
- Chapter 74 — Futility, Chapter 75 — Withdrawal of ECMO, Chapter 76 — Palliative Care, and Chapter 77 — Resource Allocation and ECPR Ethics. The ethical and procedural machinery that this chapter deliberately stops short of.
References
How to read this list
Every external reference in this chapter was read as a structured abstract. None was retrieved in full text. Numerical values quoted in the chapter are those printed in the abstract. Where a journal name, volume, page range, digital object identifier or PubMed identifier could not be confirmed from the abstract record, the entry is marked [VERIFICATION REQUIRED] rather than completed by inference — no bibliographic detail in this list has been reconstructed from memory. Every database row seeded from this chapter carries Verified = No.
Epidemiology of neurological injury after ECPR
- Migdady I, et al. Prevalence, incidence and outcomes of neurological complications in extracorporeal membrane oxygenation: a systematic review and meta-analysis. 2020. 78 studies, 50,049 patients, of whom 6,261 received ECPR. Source of the pooled ECPR complication rates and of the 17% brain-death figure in §23.1. [VERIFICATION REQUIRED] — journal, volume and identifier not confirmed.
- Gravesteijn BY, et al. Neurological outcome after extracorporeal cardiopulmonary resuscitation for in-hospital cardiac arrest: a systematic review and meta-analysis. 2020. 19 studies. Source of the finding that 84% of survivors reached a favourable outcome. [VERIFICATION REQUIRED].
- Pozzi M, et al. Long-term neurological outcome after extracorporeal cardiopulmonary resuscitation. 2026. 295 patients over 14 years; Cerebral Performance Category 1–2 at six months in 17.3%; severe sequelae in 4.4%. The single most useful outcome-distribution source in §23.1. [VERIFICATION REQUIRED].
- Rojas-Salvador C, et al. 361 ECPR patients; 47% survived and 85% of survivors reached a favourable outcome. Also the derivation and validation source for the multimodal ECPR score discussed in §23.9. 2023. [VERIFICATION REQUIRED] — read as a conference-style abstract only.
- Kalra A, et al. Machine-learning prediction of acute brain injury during extracorporeal membrane oxygenation: Extracorporeal Life Support Organization Registry analysis. 2024. 10,775 ECPR and 35,855 non-ECPR venoarterial runs; acute brain injury 16.5% against 7.7%; area under the curve 0.69 to 0.73; positive predictive value 29% and negative predictive value 90% in ECPR. [VERIFICATION REQUIRED].
- Shoskes A, et al. Brain injury is more common in venoarterial than venovenous extracorporeal membrane oxygenation: a systematic review and meta-analysis. 2020. 73 studies, 16,063 patients. Source of the critical secondary observation that the venoarterial-versus-venovenous difference disappeared once ECPR studies were excluded. [VERIFICATION REQUIRED].
- Cho SM, et al. Neurological autopsy findings in patients supported with extracorporeal membrane oxygenation. 2020. 25 decedents, 22 of them venoarterial; acute brain injury in 68%; only 8 of 25 brains free of injury. The detection-gradient anchor for §23.1. [VERIFICATION REQUIRED].
- Cholet C, et al. Neurological complications during venoarterial extracorporeal membrane oxygenation. 2018. 878 patients; ischaemic stroke at a median of 11 days and intracranial bleeding at a median of 5 days. Source of the timing separation in §23.1. [VERIFICATION REQUIRED].
The self-fulfilling prophecy, and the cohorts that escape it
- Moseby-Knappe M, et al. Performance of a guideline-recommended algorithm for prognostication of poor neurological outcome after cardiac arrest. 2020. 585 patients from the Targeted Temperature Management trial; sensitivity 38.7% and specificity 100% (98.8–100). [VERIFICATION REQUIRED].
- Bougouin W, et al. AfterROSC multicentre prospective validation, 337 patients across 28 intensive care units. Positive predictive value 100% (98–100); 162 patients, close to half the cohort, were indeterminate. Also the source of the two-favourable-sign specificity of 99% used in §23.10. [VERIFICATION REQUIRED].
- Sandroni C, et al. Prediction of poor neurological outcome in comatose survivors of cardiac arrest: a systematic review. 2020. 94 studies, 30,200 patients. Source both of the near-zero false-positive rates and of the finding that the risk of bias was high for every predictor examined. [VERIFICATION REQUIRED].
- Scarpino M, et al. ProNeCA prospective multicentre study, 346 of 396 patients analysed, in a system where life-sustaining treatment was not withdrawn. Sensitivities at full specificity: evoked potentials 57.4%, grey-white matter ratio 48.8%, isoelectric or burst-suppressed electroencephalogram 34.5%; 68 patients remained alive in a persistent vegetative state. 2019. The most important single reference in §23.2. [VERIFICATION REQUIRED].
- Götze I, et al. Multimodal neuroprognostication after cardiac arrest with withdrawal-of-treatment deaths excluded. Critical Care. 2026. Eight German hospitals, 101 patients, registration NCT02231060; two or more unfavourable markers within 14 days produced no false positives, adjusted odds ratio 34.7; adding the Coma Recovery Scale raised the area under the curve from 0.888 to 0.925. The second cohort to break the circularity, and the source of the repeated-assessment argument in §23.11. [VERIFICATION REQUIRED] — volume and identifier not confirmed.
What the circuit does to the instruments
- Maj G, et al. Neuroprognostication in patients supported with extracorporeal membrane oxygenation: a narrative review. 2025. Source of the three mechanisms in §23.3 — biomarker inflation by haemolysis, electrophysiological unreliability, and the unavailability of magnetic resonance imaging. [VERIFICATION REQUIRED].
- Ben-Hamouda N, et al. Multimodal prognostication in patients treated with extracorporeal membrane oxygenation after cardiac arrest. 2022. 397 patients without a circuit against 50 with one; poor outcome 74% against 59%; enolase significantly higher on the circuit; specificity preserved at 92–100% in both groups; the area under the curve of the multimodal combination fell from 0.822 to 0.681. The quantitative basis of §23.3. [VERIFICATION REQUIRED].
Neuron-specific enolase
- Brodská H, et al. Sub-analysis of the Prague out-of-hospital cardiac arrest study: 164 conventional against 92 ECPR patients. Enolase did not differ between resuscitation methods; optimal cut-offs for both enolase and procalcitonin were higher in the ECPR group; the authors conclude that prognostication algorithms should reflect the resuscitation method. 2024. [VERIFICATION REQUIRED].
- Schrage B, et al. Neuron-specific enolase as a predictor of neurological outcome in patients supported with extracorporeal membrane oxygenation after cardiopulmonary resuscitation. 2019. 65 derivation and 64 validation patients; enolase at 48 hours, area under the curve 0.87, cut-off 70 µg/L; specificity highest with serial measurement. The only ECPR-specific derived threshold in §23.4. [VERIFICATION REQUIRED].
- Pelle J, et al. Interaction between electroencephalographic pattern and the enolase threshold. 2025. 155 patients; with a benign background a value above 78.2 µg/L was required for full specificity, at which the standard threshold reached only 94%. [VERIFICATION REQUIRED].
- Besnard T, et al. Prediction of good neurological outcome after cardiac arrest using multimodal assessment. 2025. 215 patients; a falling enolase trajectory between 24 and 72 hours gave a positive predictive value of 82% with a sensitivity of 81%; the full favourable combination reached a positive predictive value of 96%. The single most useful source for §23.10. [VERIFICATION REQUIRED].
Clinical examination and pupillometry
- Benghanem S, et al. Quantitative pupillometry for neuroprognostication after cardiac arrest. 2026. 442 patients, two centres; neurological pupil index at or below 2 gave 99% specificity with 34% sensitivity; the qualitative light reflex had significantly higher sensitivity than the quantitative indices at comparable specificity. [VERIFICATION REQUIRED].
- Nutma S, et al. Myoclonus after cardiac arrest: post-hoc analysis of the TELSTAR trial. 2023. 157 patients; myoclonus in 63%; poor outcome 96% against 82%; the authors conclude that myoclonus cannot predict a poor outcome without false positives. [VERIFICATION REQUIRED].
Electroencephalography and seizures
- Kalra A, et al. Seizures during extracorporeal cardiopulmonary resuscitation: Extracorporeal Life Support Organization Registry analysis. 2025. 13,783 ECPR runs from 2013 to 2023; seizures in 3%; annual cases rose from 442 to 1,123 while reported seizure incidence fell from 5% to 3%; seizures associated with 90-day mortality, adjusted odds ratio 2.32. The core of the detection argument in §23.6. [VERIFICATION REQUIRED].
- Hwang J, et al. Seizures in patients supported with extracorporeal membrane oxygenation: a systematic review and meta-analysis. 2024. 23 studies, 40,420 patients; pooled incidence 3.0%, highest in ECPR at 4.9%. [VERIFICATION REQUIRED].
- Amorim E, et al. Continuous electroencephalographic monitoring during extracorporeal membrane oxygenation. 2022. 92 monitored patients, 83% venoarterial and 28% ECPR, average 54 hours of recording; epileptiform or ictal-interictal activity in 63%. The counterpart to the registry figure. [VERIFICATION REQUIRED].
- Kim YJ, et al. Electroencephalographic patterns within 96 hours after extracorporeal cardiopulmonary resuscitation. 2020. 69 patients; malignant patterns in 73.1% of those with poor outcome against 5.9% of those with favourable outcome; all patients with highly malignant patterns had poor outcomes. [VERIFICATION REQUIRED].
- Kobata H, et al. Amplitude-integrated electroencephalography after extracorporeal cardiopulmonary resuscitation. 2020. 22 patients, 19 with data within 24 hours; a continuous normal-voltage trace within 24 hours predicted good outcome at six months with a sensitivity of 100% and a negative predictive value of 100%. The strongest early favourable signal in §23.10, from the smallest sample. [VERIFICATION REQUIRED].
- Hofmeijer J, et al. Good recovery after post-anoxic status epilepticus. 2025. 274 patients; after excluding those with two or more established poor-outcome criteria, 25% of the remainder recovered well; none of those who recovered had a suppressed or burst-suppressed background before onset. [VERIFICATION REQUIRED].
- Admiraal MM, et al. Post-anoxic status epilepticus: substudy of the TTM2 trial. 2025. 191 patients with continuous electroencephalography; status epilepticus in 27% at a median of 42 hours; a favourable profile was present in 38% of those, of whom 60% survived and 40% reached a good outcome; all those lacking the favourable profile had a poor outcome. [VERIFICATION REQUIRED].
- Beretta S, et al. Neurological outcome by electroencephalographic pattern after cardiac arrest. 2018. 166 patients. [VERIFICATION REQUIRED].
- De Stefano P, et al. Non-convulsive status epilepticus after cardiac arrest. 2022. Favourable outcome in 24.5%; the authors advise against equating non-convulsive status epilepticus after return of circulation with a poor prognosis. [VERIFICATION REQUIRED].
- Hifumi T, et al. Survey of post-resuscitation practice in ECPR-capable institutions. 2022. Electroencephalography routinely applied in 13.9% of institutions. The single figure that explains most of the registry-versus-monitoring gap. [VERIFICATION REQUIRED].
Somatosensory evoked potentials
- Barbella G, et al. Prediction of poor outcome by somatosensory evoked potential amplitude. 2020. 158 patients; an amplitude at or below 0.41 µV was 100% specific (96–100) with a sensitivity of 47%, against 46% for absence alone. [VERIFICATION REQUIRED].
- Benghanem S, et al. Value of low-amplitude cortical responses for neuroprognostication. 2022. 82 patients; "absent or low voltage" raised sensitivity from 30% to 58% against absence alone. [VERIFICATION REQUIRED].
- Scarpino M, et al. Early somatosensory evoked potentials within six hours of cardiac arrest. 2025. 65 patients; bilaterally absent cortical responses 100% specific (89–100) with 67% sensitivity; adding low-amplitude prolonged responses without a late potential raised sensitivity to 93%; a high-amplitude response with a preserved late potential predicted good outcome with 94% sensitivity and 100% specificity. The bidirectional source in §23.7. [VERIFICATION REQUIRED].
Brain imaging
- Hongo T, et al. Grey-white matter ratio on early computed tomography after extracorporeal cardiopulmonary resuscitation: SAVE-J II analysis, 1,146 patients scanned within 24 hours. Graded risk across ratio bands; area under the curve 0.628; a ratio threshold of 1.005 achieved 100% specificity at a sensitivity of 0.1%; the authors conclude that the ratio may not be sufficient after ECPR and that a multimodal approach is required. 2024. The most important imaging reference in this chapter, and the one that argues against its own instrument. [VERIFICATION REQUIRED].
- Wang Y, et al. Grey-white matter ratio for neuroprognostication after cardiac arrest: systematic review and meta-analysis. 2025. 42 studies, 8,104 patients; manual measurement area under the curve 0.77, automated 0.84. [VERIFICATION REQUIRED].
- Streitberger KJ, et al. Timing of computed tomography and prognostic value of the grey-white matter ratio. 2019. 195 patients; a ratio below 1.10 was fully specific irrespective of timing, while sensitivity rose from 12% within six hours to 48% beyond 24 hours. [VERIFICATION REQUIRED].
- Case N, et al. Sensitivity of early computed tomography after cardiac arrest. 2024. 2,204 patients; maximum sensitivity 25% after five hours with a false-positive rate under 5% at all time points. [VERIFICATION REQUIRED].
- Liu Y, et al. Serial computed tomography for neuroprognostication. 2025. 123 patients; area under the curve 0.72 within 24 hours rising to 0.81 beyond 96 hours. [VERIFICATION REQUIRED].
- Ryu JA, et al. Brain computed tomography within 48 hours after extracorporeal cardiopulmonary resuscitation. 2017. 42 patients; optic nerve sheath diameter and basal-ganglia ratio each informative, and a composite reaching an area under the curve of 0.904. [VERIFICATION REQUIRED].
- Hrdlicka J, et al. Secondary analysis of the Prague out-of-hospital cardiac arrest study, 45 patients imaged within 36 hours. Oedema occurred exclusively in patients with poor outcome; favourable patients clustered within narrow ratio bands rather than above a threshold; ECPR patients had significantly lower centrum semiovale attenuation than conventionally resuscitated patients. 2023. The source of the band-not-threshold argument and of the flagged bidirectional finding in §23.8. [VERIFICATION REQUIRED].
- Li X, et al. Dual-energy computed tomography after extracorporeal cardiopulmonary resuscitation with coronary angiography. 2024. 25 patients imaged at a median of 2.08 hours; cerebral area under the curve 0.8485 despite residual contrast. Read as a conference abstract only. [VERIFICATION REQUIRED].
Predicting a good outcome
- Sandroni C, et al. Prediction of good neurological outcome after cardiac arrest: a systematic review. 2022. 37 studies; favourable predictors exceeding 80% specificity and 40% sensitivity. [VERIFICATION REQUIRED].
- Vanat A, et al. A six-item score for the prediction of good neurological outcome, with external validation. 2023. 499 patients; a threshold of four of six items gave 97.5% sensitivity and an area under the curve of 0.88. [VERIFICATION REQUIRED].
Textbooks
- ELSO Red Book, 6th edition — Chapter 32, neuroprognostication aligned with conventional post-arrest practice and the neurological care bundle; the adult systematic-review figures (44 studies, median acute neurological complication risk 13%, mortality 83% against 42% with and without neurological complications); the paediatric ECPR figures (22% acute neurological injury, 89% mortality, 84% of survivors discharged home rather than to rehabilitation); and the statement that magnetic resonance imaging is a superior modality but is incompatible with ECMO circuitry. [VERIFICATION REQUIRED] — page numbers not confirmed.
- Shinar Z, Badulak J. ECPR and Resuscitative ECMO, Chapter 3 — the clinical examination at hospital arrival: the 52-patient series in which spontaneous breathing separated 75% favourable from 11% poor outcome and none of 26 patients with a pupil diameter over 6 mm survived; and the 85-patient analysis in which mean pupil diameter was 4.1 mm against 5.2 mm, which the authors themselves call a difficult distinguishing finding. [VERIFICATION REQUIRED].
- ISCCM Manual of RRT and ECMO in ICU — haemolysis mechanisms relevant to §23.4 and sedation pharmacology relevant to §23.5. Held largely for Chapter 37 and Chapter 57. [VERIFICATION REQUIRED].
- Taha AR, Caridi-Scheible M, Leiendecker E, et al. ECMO: A Practical Guide to Management — consulted for neurological complication management; not separately quoted in this chapter. [VERIFICATION REQUIRED].
Chapter status
Drafted and audited 14 September 2026. Ten-pass quality control completed: clinical, physiology, evidence, citation, numerical, safety, contradiction, redundancy, bedside utility and literature-currency passes.
This chapter rests on a larger evidence base than Chapter 22 and a less trustworthy one. The neuroprognostication literature is large, prospective in places and guideline-codified — and almost all of it was generated in healthcare systems that withdraw life-sustaining treatment on the basis of the tests being evaluated. Two cohorts escaped that circularity by different means, one by not withdrawing treatment at all and one by excluding withdrawal deaths from analysis. Both reproduced the decision rules and neither reproduced the certainty, and the first left 68 patients alive in a persistent vegetative state — the outcome the certainty is meant to prevent. §23.2 states this as the chapter's central epistemic fact rather than as a limitation paragraph.
Three controversies were set out rather than smoothed. §23.4 asks whether the circuit raises neuron-specific enolase, and finds three sources disagreeing in three directions while converging on the same bedside rule. §23.6 separates the electroencephalogram's two unrelated jobs and argues that post-anoxic status epilepticus is not, by itself, a prognostic sign. §23.9 asks when prediction may begin and concludes that gathering and deciding have different clocks, and that on ECMO only the second one moves.
The numerical and contradiction audits flagged five problems and corrected none. (1) The only ECPR-specific multimodal score reports a goodness-of-fit statistic indicating poor calibration and presents it as supporting evidence, and (2) its validation point estimate sits almost on the lower bound of its own confidence interval. (3) The enolase literature contains a three-way contradiction — significantly higher on a circuit, no different by resuscitation method, and elevated by haemolysis — which the chapter prints intact. (4) The detection gradient is large enough to be the dominant explanation of the epidemiology: registry-reported brain injury 16.5% against 68% at autopsy, registry seizures 3% against 63% on continuous monitoring, and registry seizure incidence falling from 5% to 3% while annual case volume more than doubled. (5) The ECPR imaging cohort reports that both decreased and increased grey-white differentiation indicate poor outcome, with favourable patients confined to a narrow band, which is incompatible with a one-sided threshold and is not discussed by the source.
A structural finding that changes practice rather than wording: the largest imaging study in this population, 1,146 ECPR patients, produces an area under the curve of 0.628 and a threshold that reaches full specificity at a sensitivity of one patient in a thousand. The instrument most widely used for early prediction after ECPR is argued against by the largest study of it, in that study's own conclusion.
The citation audit was performed against the chapter tracker rather than against inherited assumption, following the discovery during Chapter 23's preparation that five chapter cross-references in published Chapters 20, 21 and 22 were wrong — hypothermia, haemolysis, revascularisation, rehabilitation and paediatrics — all now repaired, with the full 92-chapter map recorded in the project status document and the standing rule established that the tracker is the only authority for a chapter number.
The redundancy audit removed the instruments themselves, which belong to Chapter 28; laboratory monitoring generally, which belongs to Chapter 29; scores as a class, which belong to Chapter 30; sedation pharmacology, which belongs to Chapters 49 and 57; seizure treatment, which belongs to Chapter 51; long-term recovery, which belongs to Chapter 59; and the entire decision to stop, which belongs to Chapter 24. This chapter produces a probability and stops there.