Chapter question: The ECPR phone has rung. Which of these patients should be cannulated — and how much does the criteria list on the wall actually know?
Evidence search date: 11 September 2026.
Primary sources: ECPR and Resuscitative ECMO (Shinar and Badulak), Chapter 3, which reviews every eligibility component against its evidence and reproduces the Minnesota, Vancouver and Prague protocols; ELSO Red Book 6th edition Chapter 32, which reproduces the ELSO example inclusion criteria (Table 32-1); and ISCCM Manual Chapter 40.
The dominant evidence base for this chapter is SAVE-J II — a retrospective registry of 2,157 adults who received ECPR across 36 Japanese institutions between 2013 and 2018 — and its family of secondary analyses, which between them examine signs of life, pupillary findings, transient return of spontaneous circulation, initial asystole, initial pulseless electrical activity, rhythm conversion, prolonged low-flow time, sex and centre volume. All of it is observational, all of it is from patients who were selected for ECPR, and none of it tells you what would have happened to the patients who were refused.
Everything in this chapter is from structured abstracts. No full text was retrieved for any of the selection literature; the WebFetch quota was spent on Chapter 19's three anchor sources. Every row seeded from this chapter is marked Verified = No, and where two analyses disagree the disagreement is printed rather than resolved.
The ILCOR position, from the most recent systematic review of the field, is the honest starting point: "Recent randomized trials suggest potential benefit of ECPR, but the certainty of evidence remains low. It is unclear which patients might benefit from ECPR."
What this chapter covers — and what it does not
This chapter owns | Deferred to |
What a selection criterion is actually for; the definitional contamination that affects every number in the field; the published criteria sets side by side; how well the ELSO example criteria discriminate; each criterion examined separately — age, witnessed status, bystander CPR, no-flow, initial rhythm, rhythm at the door, signs of life, transient ROSC, pH, lactate, end-tidal CO₂, comorbidity and aetiology; why prognosis and treatment effect point in opposite directions; the published scores and the machine-learning models; how to write a unit's criteria; and the coupling between how tight the criteria are and how good the team becomes | Why time is the indication, the sixty-minute controversy and the trials themselves → Chapter 19
Cannulation during chest compressions → Chapter 21
The first hours after flow → Chapter 22
Neurological assessment and prognostication after cannulation → Chapter 23
Recognising futility, withdrawal, organ donation → Chapter 24
Prognostic scores as a general subject, including SAVE and RESCUE-IHCA → Chapter 30
Accidental hypothermia in depth → Chapter 65
Whether to offer ECPR at all, cost and equity → Chapter 77 |
20.1 What a selection criterion is for, and the three jobs it is doing at once
A unit's ECPR criteria list looks like a clinical decision rule. It is not one. It is doing three different jobs simultaneously, and most of the arguments about ECPR eligibility are arguments between people who are each thinking about a different one.
Physiology — the three jobs, and why they conflict
1. Prognostic selection: who is likely to survive? This is what the literature measures, and it is the only one with data. Every odds ratio in this chapter comes from asking which ECPR recipients did well.
2. Treatment-effect selection: who is likely to benefit from ECPR specifically? This is what the decision actually requires, and it is a different question. A patient who would have survived with conventional CPR gains nothing from cannulation but takes all its risk. §20.6 shows that these two questions can point in opposite directions, with quantitative evidence.
3. Operational selection: what can this team do, at this hour, without failing? A criterion that a night team cannot apply in ninety seconds is not a criterion. This job is why the list exists at all — Chapter 19 §19.1 established that the decision cannot be made well in the moment, so it must be made in advance.
The three jobs pull in different directions. Prognostic selection makes the list long. Operational selection makes it short. Treatment-effect selection would sometimes make it select the opposite patients.
What the randomised trials can and cannot contribute here
Chapter 19 set out the three randomised trials in full. For this chapter their contribution is sharply limited, and the ECPR textbook states the limitation precisely: "Clinical trials remain limited to a single inclusion criterion, and thus are only able to inform of benefits to that specific patient group rather than providing evidence for optimal eligibility criteria."
ARREST randomised only patients with refractory ventricular fibrillation. INCEPTION randomised only patients with ventricular arrhythmias. PRAGUE-OHCA randomised all rhythms and found that the non-shockable patients did badly. Between them they tell you that ECPR can work in shockable rhythms in a good system. They do not tell you where to put the age limit, what to do about an unwitnessed arrest, or whether a lactate of 15 should stop you. Every one of those decisions rests on observational data.
Danger — the shape of the evidence base, and what it structurally cannot show
Almost everything that follows comes from cohorts of patients who were selected for ECPR. That has two consequences that no amount of statistical adjustment removes.
First, the comparison group is missing. When a study reports that patients with signs of life did better, it is comparing ECPR recipients with signs of life against ECPR recipients without them. It is not comparing them against patients who had signs of life and did not get ECPR — those patients are usually not in the dataset at all.
Second, every criterion already in use is invisible. If a registry's contributing units all excluded patients over 75, the registry cannot tell you anything about patients over 75. Selection criteria are self-confirming: they remove the evidence that would refute them.
The practical consequence: treat every finding in §20.5 as describing who does well among those cannulated, which is a prognostic statement, and not who should be cannulated, which is a treatment-effect statement. §20.6 is where the difference becomes quantitative.
20.2 The definitional contamination that affects every number in this chapter
Before any criterion can be weighed, two definitional problems have to be stated, because they contaminate the numbers rather than merely qualifying them.
Pitfall — ECPR and post-ROSC ECMO are being counted together
The ECPR textbook flags it plainly: "Many studies have been ambiguous in regard to differentiating cases initiated during active chest compressions versus those commenced after ROSC. The latter represents a better prognosis and should not be classified as 'ECPR'."
A patient cannulated during compressions and a patient cannulated after a pulse returned are in different diseases with different prior probabilities, and some published ECPR series contain both. Every survival figure in this chapter is therefore an upper bound of uncertain tightness.
When reading any ECPR paper, find the sentence that says whether the circuit was started with or without a pulse. If it is not there, the survival figure is not comparable with one where it is.
The second problem is the one Chapter 19 §19.1 opened with: "refractory" has no agreed definition, with published thresholds ranging from 10 to 60 minutes. A unit that defines refractory as 10 minutes and a unit that defines it as 30 are selecting different patients before they apply a single other criterion, and their published survival figures are not comparable.
Clinical pearl — the two questions to ask of any ECPR outcome figure before you use it
- Were the patients in arrest when the pump started?
- What did this unit mean by "refractory"?
A 40% survival figure from a programme that cannulates at 15 minutes, some of them after ROSC, and a 20% figure from a programme that cannulates at 45 minutes strictly during compressions, may describe identical clinical skill. Benchmark against a programme whose definitions match your own, or do not benchmark.
20.3 The published criteria sets, side by side
Four published criteria sets, plus the three trials, are reproduced here because seeing them together shows what is agreed and what is not.
Criterion | ELSO example set (via Red Book Table 32-1) | Vancouver / St Paul's protocol | The three randomised trials |
Age | Under 70 | 65 or under | ARREST 18–75 · PRAGUE 18–65 · INCEPTION 18–70 |
Witnessed | Required | Required — "seen or heard" | PRAGUE and INCEPTION required it |
No-flow | Under 5 minutes, i.e. bystander CPR | Not stated as a number | INCEPTION required bystander CPR |
Initial rhythm | VF, pulseless VT or PEA | Either initial shockable rhythm or signs of life with CPR | ARREST and INCEPTION shockable only; PRAGUE all rhythms |
Time ceiling | Arrest to ECMO flow under 60 minutes | Dispatch to hospital under 50 minutes, using arrest time if EMS-witnessed and subtracting any periods of ROSC | ARREST transfer under 30 min · PRAGUE arrival under 50 min · INCEPTION cannulation within 60 min |
End-tidal CO₂ | Above 10 mmHg during CPR before cannulation | Above 10 mmHg pre-hospital; below 10 is an emergency-department exclusion | Not an entry criterion in any trial |
Intermittent ROSC | Listed as a favourable criterion | Periods of ROSC are subtracted from the elapsed time | Not specified |
Signs of life | "May be a positive predictive factor for survival" | An alternative route into eligibility — moving, or gasping, or pupils 5 mm or less | Not specified |
Comorbidity | Absence of known life-limiting comorbidity, consistent with goals of care | No CHF, COPD or significant lung disease, dialysis, liver failure, malignancy, alcoholism; no pre-existing major neurological deficit; re-checked against hospital records in the ED | Not specified |
Metabolic | Not specified | Lactate above 18 or PaO₂ below 50 is an ED exclusion — "do not delay ECMO initiation for results" | Not specified |
Cause | Not specified beyond comorbidity | Non-traumatic; no recent recreational drug use; cause either presumed cardiac or overdose of cardiac toxins (beta-blocker, calcium-channel blocker, psychotropics, digoxin) | PRAGUE required presumed cardiac origin |
Other | No known more-than-mild aortic incompetence | Not too large for the mechanical compression device; if initial rhythm was non-shockable, signs of life must still be present at the ED | Mechanical CPR mandatory in ARREST, used in 85% and 90% of the others |
Clinical pearl — three design features of the Vancouver protocol worth stealing
Whatever you think of its specific thresholds, three structural choices in that protocol are better engineering than most published lists.
1. Two gates, not one. A pre-hospital inclusion gate that must be met in full, and a separate emergency-department exclusion gate where any single item disqualifies. The information available at the two moments is different, so the criteria are different.
2. Signs of life are an alternative route in, not a bonus. A patient qualifies with either a shockable rhythm or signs of life. §20.5.4 shows the data now support this design strongly.
3. The list contains an anti-delay instruction. "Do not delay ECMO initiation for results." A criterion that stops the clock is worse than no criterion, and this protocol says so on its face.
And one explicit carve-out: "For hypothermia-related arrests these guidelines do not apply." Accidental hypothermic arrest is a different disease with different rules — see §20.5.8.
20.4 Controversy 1 — how well do the standard criteria actually discriminate?
Controversy 1 — Do the ELSO example selection criteria identify the patients who do well?
The question. The criteria set in Red Book Table 32-1 is the nearest thing to an international standard, and most units' lists are recognisably derived from it. It has never been prospectively validated. One study has tested it against a multivariable alternative.
What that study found. A single-centre cohort of 120 normothermic ECPR patients treated between 2010 and 2021, in which favourable outcomes occurred in 27.5%, was used to develop the Pre-ECPR score, incorporating age, a composite of no-flow and initial rhythm, total cardiac arrest time, signs of life, pupil dilation, regional cerebral oxygen saturation, arterial pH and end-tidal CO₂.
— The Pre-ECPR score achieved an area under the curve of 0.87 (95% CI 0.77–0.93) in development and 0.79 (0.67–0.88) in internal cross-validation.
— The ELSO example selection criteria achieved an area under the curve of 0.63 (95% CI 0.54–0.72) for one-year survival with favourable outcome, and the difference was statistically significant (P = 0.012).
— A Pre-ECPR probability above 6.4% gave 100% sensitivity with a positive predictive value of 40.5%.
What an AUC of 0.63 means at the bedside. It means that if you take one patient who will do well and one who will not, the ELSO criteria rank them correctly about 63 times out of 100 — against 50 for a coin. The lower bound of the confidence interval is 0.54. This is a criteria set that is doing something, but not much, and certainly not what a list of hard thresholds implies it is doing.
The case for the criteria anyway, and it is not a weak case. A criteria list is not only a prediction instrument. It is an operational instrument (§20.1), and its value lies partly in being applicable in ninety seconds by a team that did not design it. The Pre-ECPR score requires regional cerebral oximetry, an arterial pH and an end-tidal CO₂ — it cannot be applied by a dispatcher or a paramedic, which is where most of the ECPR decision actually happens. A discriminating instrument you cannot use at the decision point is worth less than a crude one you can.
Resolution — this book's reading. The finding does not mean the criteria should be abandoned. It means they should not be trusted to a decimal place, and a patient who fails one of them by a small margin has not been shown to be a poor candidate. Use the list as a fast operational filter that keeps the pathway reproducible, and hold its individual thresholds loosely. The strongest practical implication is that a criteria list should have an override route for a senior decision-maker, because the list itself is demonstrably a weak discriminator.
Certainty: low. One single-centre study of 120 patients, from abstract only, developing and testing in the same dataset with internal cross-validation and no external validation. It has not been replicated, and it should be. But nothing contradicts it, and no study has shown the standard criteria performing better.
Danger — the criteria are not the only thing being tested here
It is tempting to read the AUC of 0.63 as showing that the ELSO list is badly chosen. A second reading is at least as likely: that the outcome of ECPR is genuinely hard to predict from what is knowable at the bedside during an arrest, and that no list of pre-cannulation variables will do much better.
The multivariable models in §20.7 reach areas under the curve of roughly 0.70 to 0.89 — better than 0.63, but a long way from certainty, and the higher figures come from in-sample estimates that fall on validation.
The honest statement is that the ECPR decision is made under irreducible uncertainty, and that no published instrument removes it. That is a reason to be humble about refusing patients, not only about accepting them.
20.5 The criteria one at a time
20.5.1 Age — the criterion with the weakest evidence and the widest use
Age limits between 60 and 75 appear in almost every published list. The evidence beneath them is thin and partly contradictory.
Finding | Source and size | Certainty |
Young age was NOT associated with favourable neurological outcome at discharge | European cohort, 423 ECPR patients | Low — single large cohort, abstract only |
Several further studies report no association between age and ECPR outcome | Multiple cohorts as reviewed in the ECPR textbook | Low |
No favourable 1-month neurological outcome in any of 22 patients aged 75 or over | Single OHCA cohort | Very low — 22 patients |
18% favourable neurological outcome in patients over 75 — but none with a low-flow interval beyond 60 minutes | A second cohort | Very low, and it contradicts the row above |
Peak survival at age 60, with a bimodal peak of favourable neurological outcome at ages 40 and 65 | Single cohort | Very low — but the shape matters more than the values |
Age 45–60 was protective against in-hospital death | Chinese CSECLS registry, 990 patients collected, 351 in the development model | Low |
Age under 65 associated with favourable outcome, OR 6.17 | 2025 systematic review, 8 studies pooled, 4,353 patients | Low to moderate — the largest synthesis, but pooled from heterogeneous cohorts |
Physiology — why age behaves strangely here, and it is not a statistical artefact
The textbook offers the explanation and it is convincing: the cause of arrest differs systematically by age, and the causes differ in how treatable they are by a circuit.
Pathologies over-represented in younger arrest victims — aortic dissection, occult drug overdose, primary cerebral catastrophe — are among the least amenable to ECPR. The pathology that dominates in older adults — acute coronary syndrome — is the one ECPR is best at bridging, because there is a definitive treatment waiting in the catheter laboratory.
Age is therefore a proxy for aetiology pointing in two directions at once, which is exactly what a bimodal survival curve and a protective 45–60 band look like. It is also why a past history of cardiovascular disease is associated with better, not worse, ECPR outcomes (§20.5.7) — a finding that is otherwise baffling.
What to do with this. The textbook's own conclusion is that "age appears to be a poor predictor of ECPR-treatment outcomes; however, an age limit of 75 appears to be a reasonable ceiling for ECPR eligibility, especially for prolonged low-flow durations typically seen with out-of-hospital cardiac arrests." This book agrees, with one addition: an age ceiling is doing operational work more than prognostic work. It makes the decision fast and reproducible. It should not be defended as though it identified biology, and a patient one year over the line who is otherwise ideal is a legitimate case for the override route in §20.4.
20.5.2 Witnessed arrest, bystander CPR and no-flow — three names for one variable
These three criteria are correlated, they are all attempts to estimate the same quantity, and the quantity is the no-flow interval — the only part of the timeline that is pure normothermic ischaemia (Chapter 19 §19.2).
Criterion | What the data show | Verdict |
Witnessed arrest | Consistently associated with better outcomes. Several cohorts report no survivors at all among unwitnessed cases. But: 1 of 16 (6.2%) unwitnessed patients survived to discharge in one series of 111; 3 of 21 (14%) in another series of 156; and a programme restricted to initial shockable rhythms reported 33% (4 of 12) survival in its unwitnessed subgroup | Strong prognostic signal, but not an absolute exclusion — and the shockable-rhythm subgroup finding is the reason why |
Bystander CPR | Mixed. Some cohorts show an association with survival, several do not. Without bystander CPR, favourable neurological outcome was 3.8% in one cohort and 13% in another. A programme that did not exclude on bystander CPR status reported 2 of 12 (17%) such patients surviving to discharge. In SAVE-J II, bystander CPR was independently associated with favourable outcome (OR 1.63, 95% CI 1.03–1.88), and the 2025 systematic review found mixed results across studies | The textbook's conclusion stands: data remain insufficient to use bystander CPR as an exclusion criterion. It is a modest positive, not a gate |
No-flow interval | Several programmes build a no-flow threshold into eligibility, but evidence supporting a distinct cut-off is lacking, and estimating no-flow reliably at the scene is very difficult. In patients treated with conventional resuscitation, no favourable neurological outcomes were seen with no-flow beyond 10 minutes | 10 minutes is a reasonable ceiling. Treat the number as soft, because the measurement is soft |
Clinical pearl — the initial rhythm may tell you the no-flow time
This is the most useful single inference in the selection literature and it is easy to miss.
In patients treated with conventional resuscitation, an initial shockable rhythm carried 94% certainty that the preceding no-flow interval was under 10 minutes. Ventricular fibrillation degenerates; if the monitor still shows it, not much time has passed unperfused.
The consequence is that initial rhythm may supersede witnessed status. An unwitnessed arrest found in ventricular fibrillation has, with high probability, a short no-flow interval — which is precisely why the shockable-only programme achieved 33% survival in its unwitnessed patients while other series found none.
Bedside version: do not exclude an unwitnessed arrest that presents in VF or pulseless VT on that basis alone. The rhythm is better evidence about the no-flow time than the absence of a witness.
20.5.3 Initial rhythm, and the rhythm at the door
Physiology — the paradox at the centre of rhythm-based selection
The textbook states it exactly: "In theory, initial rhythm in refractory cardiac arrest cases treated with ECPR should not have predictive power, as none have ROSC. The degree of neurologic injury should not be dependent on rhythm but rather blood flow. However, studies consistently show differential outcomes based on initial rhythm."
The paradox resolves through two mechanisms, and both are visible in the data.
First, rhythm encodes elapsed unperfused time (§20.5.2) — it is a no-flow proxy.
Second, rhythm encodes aetiology. Ventricular fibrillation in an adult is usually ischaemic, and ischaemia has a definitive treatment at the end of the bridge. Asystole is usually the end-state of something else, and often of something a circuit cannot fix.
Neither mechanism is about the electrical activity itself, which is why the rhythm at the moment of cannulation turns out to matter as much as the rhythm at the start.
Comparison | Favourable neurological outcome | Source |
Shockable vs non-shockable initial rhythm | 32% vs 0% (mean low-flow 84 min) | 68 OHCA patients |
Shockable vs non-shockable initial rhythm | 24% vs 11% | European cohort, 423 patients |
Shockable vs PEA vs asystole (in-hospital) | 49% vs 34% vs 12%, with best low-flow cut-points of 46 minutes for shockable and 22 minutes for PEA | 190 patients, median low-flow 30 min |
Shockable rhythm at the scene | OR 2.11 (1.16–3.95) | SAVE-J II, 1,823 patients |
Shockable rhythm on hospital arrival | OR 2.59 (1.60–4.30) — a stronger association than the rhythm at the scene | SAVE-J II, 1,823 patients |
Shockable rhythm at ECMO initiation | OR 6.67; on hospital arrival OR 3.68 | 2025 systematic review, 4,353 patients |
The rhythm you should be asking about is the one on the monitor now, not the one the ambulance found. Two SAVE-J II analyses make this concrete in patients whose initial rhythm was shockable.
Rhythm on hospital arrival, in patients with an initial shockable rhythm | Favourable neurological outcome | Adjusted odds ratio |
Sustained shockable | 22.4% | Reference |
Pulseless electrical activity | 9.5% | 0.35 (0.21–0.58) |
Asystole | 2.7% | 0.08 (0.03–0.20) |
A companion analysis of 718 patients with an initial shockable rhythm found favourable neurological outcome in 12.9% of those who converted to PEA before ECPR versus 26.4% of those with a sustained shockable rhythm (adjusted OR 0.42, 0.27–0.66) — and, importantly, when the conversion happened mattered: 9.8% if it occurred during ambulance transport, 18.0% at hospital arrival, 21.4% before ECMO initiation. Conversion after hospital arrival was not significantly different from a sustained shockable rhythm (19.6% versus 26.4%, P = 0.19).
Danger — VF that has become asystole by the time you see it is a different patient
A patient whose initial rhythm was ventricular fibrillation and who is in asystole on arrival had a favourable neurological outcome in 2.7% of cases, with an adjusted odds ratio of 0.08. That is among the lowest yields reported anywhere in this literature for a patient who nominally satisfies a shockable-rhythm criterion.
A criteria list that asks only about the initial rhythm will admit this patient. Most published lists do exactly that.
Add the arrival rhythm to your list. It is free, it is available at the decision point, and on these data it carries more information than the criterion most lists already contain.
20.5.4 Signs of life — the strongest variable in the field
Gasping, any pupillary light reaction, or any body movement during resuscitation. Across every dataset that has examined them, signs of life outperform every other pre-cannulation variable, and they are absent from most published criteria lists as anything more than a footnote.
Study | Finding |
SAVE-J II, 1,395 patients | Signs of life on arrival in 250 (17.9%). Favourable neurological outcome 38.0% with versus 8.1% without; adjusted OR 5.65 (3.97–8.03) |
Pooled analysis of three French cohorts, 434 ECPR recipients | Any sign of life in 61% — pupillary light reaction 48%, gasping 32%, increased level of consciousness 13%. 30-day survival with CPC 1–2 was 15% overall. Adjusted odds ratios: any sign of life 7.35 (2.71–19.97), pupillary light reaction 5.86 (2.28–15.06), increased consciousness 4.79 (2.16–10.63), gasping 1.75 (0.95–3.21) |
SAVE-J II, 1,823 patients | Independently associated with favourable outcome: body movement OR 7.10 (1.79–32.90), gasping OR 4.33 (2.57–7.28), pupillary reflex on arrival OR 2.93 (1.73–4.95) |
2025 systematic review, 4,353 patients | Any life sign OR 9.63 — with sensitivity 0.89, specificity 0.46, positive predictive value 0.22 and negative predictive value 0.96 |
Single centre, 227 refractory OHCA patients | Favourable neurological outcome by rhythm and signs of life: shockable without signs of life 4.1%; shockable with signs of life 48.2%; non-shockable without signs of life 0.0%; non-shockable with signs of life 38.7% |
Danger — signs of life reorder the criteria, and that last row is why
Read the four cells of the 227-patient table in the order that matters: a patient with a non-shockable rhythm and signs of life did better (38.7%) than a patient with a shockable rhythm and no signs of life (4.1%).
That is a ten-fold inversion of the criterion most lists are built around. If it replicates, signs of life are not a tie-breaker to be applied after the rhythm criterion — they are the more informative variable, and the rhythm is the modifier.
The Vancouver protocol already encodes this, by accepting either an initial shockable rhythm or signs of life as a route into eligibility, and by requiring that in a patient whose initial rhythm was non-shockable, signs of life must still be present at the emergency department. On these data that is the best-designed criterion in the published sets.
Certainty: moderate for the direction, low for the magnitudes — see the numerical flag below.
Pitfall — two numerical problems in the signs-of-life literature, flagged and not corrected
1. The adjusted odds ratios from the 227-patient study — reported as 34.33 for shockable and 96.51 for non-shockable patients with signs of life — are not usable estimates. One of the four cells contains zero favourable outcomes, and an odds ratio cannot be computed across a zero cell without a correction that the abstract does not describe. A figure of 96.51 derived from 73 non-shockable patients has a confidence interval that would span orders of magnitude. The four proportions are used in this chapter. The two odds ratios are not.
2. The two largest analyses disagree about which sign of life carries the signal. SAVE-J II found gasping strongly associated with favourable outcome (OR 4.33, 2.57–7.28); the pooled French analysis found gasping the weakest of the four and not statistically significant (OR 1.75, 0.95–3.21), while ranking pupillary light reaction far higher. Both are recorded and neither is asserted. What survives the disagreement is the composite: any sign of life is a strong positive, and which one it is remains unresolved.
Clinical pearl — how to use a negative predictive value of 0.96
The pooled operating characteristics of "any sign of life" are sensitivity 0.89, specificity 0.46, positive predictive value 0.22, negative predictive value 0.96.
Read those four numbers as a pair of statements.
As a rule-in test it is poor. Specificity 0.46 and positive predictive value 0.22 mean that most patients with signs of life still do badly. Signs of life do not promise a good outcome.
As a rule-out test it is strong. A negative predictive value of 0.96 means that among patients with no sign of life at any point, roughly 96% will not achieve a favourable neurological outcome.
The pooled analysis draws the conclusion carefully, and this book adopts the same caution: "The lack of any sign of life might obviate the provision of ECPR for patients without shockable cardiac rhythm." Note the final clause. The absence of signs of life is a strong argument against ECPR in a patient who also has a non-shockable rhythm — where the observed rate of favourable outcome in that 227-patient study was 0.0%. In a patient with a shockable rhythm it is a serious negative, not a veto, and the observed yield was still 4.1%.
20.5.5 Transient return of spontaneous circulation
Defined in the largest analysis as any palpable pulse lasting at least one minute before ECMO initiation.
In SAVE-J II, 328 of 1,501 patients (22%) had transient ROSC. Favourable neurological outcome was 26% versus 12%, and survival to discharge 46% versus 24%, with adjusted odds ratios of 3.34 (2.35–4.73) for neurological outcome and 3.99 (2.95–5.40) for survival.
The ELSO example criteria already list intermittent ROSC as a favourable feature, and the Vancouver protocol goes further by subtracting periods of ROSC from the elapsed time used to test the 50-minute ceiling. That is the right operational treatment: a minute of native circulation is a minute the brain was perfused, and the stopwatch has been counting it as low-flow.
Pitfall — persistent ventricular fibrillation is not the same claim as intermittent ROSC
Some protocols use "persistent" or "refractory" VF as a qualifying feature, on the intuition that a heart still fibrillating is a heart still viable.
The data do not support it. A study of 423 ECPR cases found no association between the number of defibrillations and survival — a median of 2 (IQR 0–5) in survivors and 2 (IQR 0–4) in non-survivors — and the textbook's conclusion is explicit: "Refractory VF does not appear to confer superior prognosis compared to simply an initial VF rhythm."
Intermittent ROSC earns its place in a criteria list. Refractory VF, as distinct from initial VF, does not.
20.5.6 pH, lactate and end-tidal CO₂ — the metabolic gates
These are the criteria most likely to be applied wrongly, because they arrive as numbers and numbers feel like thresholds.
Finding | Certainty and comment |
Systematic reviews find higher pH and lower lactate associated with better ECPR outcomes | Low — consistent direction, no validated cut-off |
Initial pH at or above 7.0: OR 2.01 for favourable outcome | Low to moderate — 2025 systematic review of 4,353 patients |
94% mortality when lactate was 13 or above, in 52 ECPR-treated OHCAs | Very low — 52 patients, and 94% mortality is not 100% |
Survivors have been reported with a lactate of 19 and a pH of 6.68 | Case-level, and decisive against absolute thresholds |
One programme restricts eligibility to lactate 18 or below; "while evidence supporting this threshold is lacking, the results of the program are excellent" | Very low as evidence for the number; the honesty of the statement is the point |
Adding lactate and pH to a clinical prediction model did not improve it. Clinical-only versus clinical-plus-laboratory models: AUROC 0.79 vs 0.83 for survival and 0.80 vs 0.79 for favourable neurological outcome — differences not statistically significant | Low — single centre, but directly relevant |
End-tidal CO₂ and PaO₂ thresholds are used by at least one programme with good results, but the outcomes of the patients excluded by those thresholds are unknown | Very low — the textbook's own verdict is that further data are required |
Danger — the metabolic numbers are prognostic, weakly, and they are the wrong tool for a gate
Four reasons to hold these thresholds loosely.
1. Survivors exist beyond every published threshold — a pH of 6.68 and a lactate of 19 among them.
2. Adding them to a model did not improve prediction of neurological outcome at all. The clinical variables you already have appear to carry the same information.
3. The end-tidal CO₂ value is confounded by the drugs given during the resuscitation that produced it — adrenaline lowers it, bicarbonate transiently raises it (Chapter 19 §19.2).
4. They arrive late. The Vancouver protocol's own instruction is the correct one: the metabolic exclusion is checked on a sample taken during cannulation, and "do not delay ECMO initiation for results." A blood gas that arrives after the cannula is in can inform the conversation about continuing; it cannot inform the decision to start.
And §20.6 goes further than caution: in one analysis, a lower pH and a higher lactate predicted a LARGER treatment effect from ECPR, not a smaller one.
The textbook's own conclusion is the safe one: "pH and lactate values are correlated with outcomes; robust cut-off values to classify non-survival have not been identified."
20.5.7 Comorbidity — the criterion that removes the fewest patients
Every published list excludes patients with significant comorbidity, and the intuition is obvious. The data are surprising.
Comorbidity | Association with ECPR outcome |
Past history of cardiovascular disease | Associated with IMPROVED outcomes |
Diabetes, COPD, asthma, chronic renal disease, cirrhosis, cancer, hypertension, smoking, immunocompromise | No association detected with ECPR outcome |
Dyslipidaemia | Associated with non-survival in one study — a lone finding, direction unexplained |
History of cerebrovascular disease | A mortality risk factor in the Chinese registry model |
Elevated serum creatinine | Associated with unfavourable outcome, OR 2.22 |
A stepwise exclusion analysis removing terminal illness, cancer, end-stage liver disease and pre-existing severe cognitive impairment | Changed the rate of favourable outcomes from 21.5% to 22.2% |
Evidence — the stepwise exclusion finding deserves to be read slowly
A large investigation applied a step-wise approach to constructing optimal exclusion criteria and found that excluding terminal illness, cancer, end-stage liver disease and pre-existing severe cognitive impairment moved the proportion of favourable outcomes from 21.5% to 22.2% — an absolute gain of 0.7 percentage points.
Those are four of the most confidently applied exclusions in ECPR practice, and together they improved the yield by less than one point.
Two readings are possible and both are probably true. The first is that these conditions are simply rare in the population that reaches an ECPR decision, so excluding them removes few patients. The second is that comorbidity is much less important than the arrest characteristics — which is what the rest of this section shows, with age, rhythm, signs of life and time all carrying far larger effects.
Certainty: low — single analysis, abstract only.
What it does not mean. It does not mean comorbidity exclusions should be deleted. A patient with metastatic cancer and a documented ceiling of care should not be cannulated, and that is a decision about goals rather than about yield (Chapter 77). What it means is that comorbidity exclusions are not where the discriminating power of a criteria list lives, and a unit that spends its decision time on the comorbidity question is spending it in the wrong place.
And the cardiovascular-disease finding is not a paradox once §20.5.1 is in view. A patient with known coronary disease who arrests is more likely to be arresting from something with a definitive treatment at the end of the bridge. The comorbidity that predicts a good ECPR outcome is the comorbidity that explains the arrest.
20.5.8 Aetiology — the criterion that is hardest to apply and matters most
The cause of the arrest determines whether there is anything at the far end of the bridge. It is also, during a resuscitation, usually unknown.
Cause | What is known | Verdict |
Acute coronary syndrome | The commonest cause in refractory arrest treated with ECPR — 63% of OHCA and 55% of IHCA in one 252-patient registry. In an ELSO-database analysis, survival after ECPR for myocardial infarction was 27%. In the SAVE-J II PEA analysis, favourable neurological outcome 9.7% | The canonical indication. There is a definitive treatment waiting |
Pulmonary embolism | 4 of 19 ECPR-treated cases survived with favourable neurological outcome (median low-flow 77 minutes) versus 0 of 20 matched historical controls. A case series reported 8 of 9 surviving. A "high-risk" PE series reported 2 of 18. In SAVE-J II's PEA analysis, 19.3% favourable neurological outcome — the highest of any cause in that subgroup | An appropriate indication, including with prolonged low-flow. The difficulty is identifying it among other non-shockable rhythms |
Myocarditis and primary cardiac causes | ELSO-database survival after ECPR: myocarditis 61%, cardiomyopathy 43%, versus sepsis 18%. A separate series: 4 of 12 myocarditis patients survived | The highest-yield aetiology in the registry data. Rarely identifiable during the arrest |
Accidental hypothermia | In the SAVE-J II asystole analysis, hypothermia was the commonest cause among the few favourable outcomes — 7 of 12. Dedicated prediction tools exist for this population, including an individual-patient-data meta-analysis and a specific outcome score | A different disease with different rules. The Vancouver protocol's carve-out — "For hypothermia-related arrests these guidelines do not apply" — is the correct design. Do not apply an ECPR time limit to a hypothermic arrest. Depth in Chapter 65 |
Cardiotoxic drug overdose | Included in several eligibility sets — beta-blocker, calcium-channel blocker, psychotropic, digoxin. One series reported 3 of 10 surviving after ECPR for toxicological causes. Overall the incidence of VA ECMO for toxicological indications is low | A reasonable indication when the toxin is cardiotoxic, because the poisoning is self-limiting and the circuit buys the time. Note that the Vancouver protocol simultaneously excludes recent recreational drug use — the two are not the same thing |
Overdose causing respiratory depression then arrest | Probably unfavourable, but the textbook states that evidence supporting this is unavailable | Reasoning, not evidence. Certainty: very low |
Acute aortic disease | In the SAVE-J II PEA analysis, no patient with acute aortic disease survived | A genuine exclusion where it can be suspected — and a reason to think before cannulating a young patient with an unexplained non-shockable arrest |
Primary cerebral disease | In the same analysis, no survivors | A genuine exclusion where it can be suspected |
Sepsis | ELSO-database survival 18% | Low yield; the arrest is a late event in a systemic illness |
Trauma | Excluded by the Vancouver protocol and most others | Outside this chapter's scope |
Clinical pearl — the aetiology question you can actually answer during an arrest
You will rarely know the cause. You can almost always answer a narrower question that carries most of the same information: is there something at the far end of this bridge?
If the answer is a catheter laboratory, a thrombectomy, a rewarming circuit or a drug that will clear, the bridge has a destination. If the answer is that the circuit will restore a circulation to a brain or an aorta that is already destroyed, it does not — and Chapter 19's bridge-to-nowhere warning applies with full force.
Two aetiologies with zero survivors in the SAVE-J II data are worth holding in mind as the arrest is running: acute aortic disease and primary cerebral catastrophe. Both are suspectable — a young patient with chest or back pain before collapse, a patient with a sudden severe headache, a patient found with unexplained non-shockable arrest and no cardiac history. Neither is excluded by any criteria list in §20.3.
20.6 Controversy 2 — prognosis and treatment effect point in opposite directions
Controversy 2 — Should ECPR go to the patients most likely to survive, or to the patients ECPR helps most?
The question. Every criteria set in §20.3, and every variable in §20.5, selects patients with the best expected outcome. That is prognostic selection. But a patient's expected outcome is not the same as the difference ECPR makes to it — and if those two quantities rank patients differently, then every criteria list in current use is selecting the wrong patients.
Chapter 11 named this principle for VA ECMO in cardiogenic shock — "prognosis is not indication" — as a conceptual argument. There is now a quantitative test of it in ECPR.
What the study did. A 2026 analysis of adult, non-traumatic, emergency-medical-services-attended out-of-hospital cardiac arrests from four tertiary centres in Taiwan between 2016 and 2024. After propensity-score matching on shockable rhythm and witnessed arrest, 1,953 matched patients, of whom 977 received ECPR. A gradient-boosted causal forest model estimated each patient's individualised treatment effect — the predicted difference ECPR would make to that patient's survival — rather than their predicted survival.
What it found.
— Overall, survival was no different: 11.1% with ECPR versus 12.8% without (standardised mean difference 0.054). Averaged across everyone, ECPR did nothing.
— In the top 10% by predicted benefit, survival to discharge was 50.0% with ECPR versus 20.0% without — an absolute observed treatment effect of 30.0% (95% CI 1.56–57.58, P = 0.042).
— The model identified subgroups with higher observed benefit than those selected by rule-based criteria, including the entry criteria of the ARREST, PRAGUE-OHCA and INCEPTION trials and the hospitals' own protocols.
— The factors associated with a LARGER treatment effect were: lower serum pH, lower PaCO₂, higher lactate, younger age, and bystander CPR.
Read that last line against §20.5.6. Every prognostic model in this chapter treats higher pH and lower lactate as favourable. This analysis found that lower pH and higher lactate identified the patients in whom ECPR made the biggest difference.
Why that is not a contradiction. A profoundly acidotic patient does badly whether or not they are cannulated — but they do very badly without the circuit and merely badly with it, and the gap between those two is what a treatment-effect model measures. A patient with a near-normal pH may do reasonably well either way, so the gap is small. Prognosis ranks patients by where they end up. Treatment effect ranks them by how much the intervention moves them. These are different orderings and there is no reason they should agree.
The case for staying with prognostic selection anyway.
— Treatment effect is not observable in an individual. It is a modelled counterfactual, and the model here is a causal forest fitted to observational data with propensity matching on two variables. The "20.0% survival without ECPR" in the top decile is an estimate, not a measurement.
— The confidence interval on the headline effect runs from 1.56% to 57.58%. It excludes zero, and it excludes very little else.
— Single health system, retrospective, one paper, not externally validated, abstract only.
— And there is a resource argument that cuts the other way: a programme that cannulates the sickest patients will have lower absolute survival, which — for reasons §20.9 explores — may cost it the institutional support it needs to exist at all.
Resolution — this book's reading, stated as reasoning.
The finding should not yet change any unit's thresholds, and it should change how the thresholds are defended. A metabolic exclusion presented as "this patient is too acidotic to benefit" is making a treatment-effect claim that the only direct evidence on the question contradicts. If the exclusion is to be kept, it should be defended honestly as a resource-allocation decision and a prognostic one — this patient is unlikely to do well — and not as a claim that ECPR would not help them.
The broader point survives the weakness of the single study, because it is structural rather than empirical: an observational literature that only ever measures outcome among the cannulated cannot, in principle, tell you whom to cannulate. That is a limitation of the design, not of any particular dataset, and no amount of further registry analysis will fix it.
Certainty: very low for the specific model; moderate for the principle, which follows from the structure of the evidence and is supported independently by Chapter 11's reasoning and by Chapter 19's finding that the randomised trials measure systems rather than patients.
Clinical pearl — the sentence to keep
"Prognosis is not indication" (Chapter 11) now has a quantitative counterpart in ECPR: the patients with the worst predicted outcome were the patients with the largest predicted benefit.
At the bedside this does not mean cannulating everyone. It means that when you are about to refuse a patient, you should know which of two things you are saying:
— "This patient will probably die anyway" — a prognostic claim, usually well supported by §20.5.
— "ECPR would not help this patient" — a treatment-effect claim, which the data in this section do not support and in some cases invert.
They feel like the same sentence in the moment. They are not.
20.7 The scores, and what they are worth
Several instruments now exist. None is a substitute for a criteria list, and none has been prospectively validated as a selection tool.
Score | Population and components | Performance | Certainty |
START-ECPR
Signs of life, Transient ROSC, Not Asystole Rhythm | SAVE-J II, 648 OHCA patients with initial PEA or asystole. Three predictors: shockable rhythm or PEA at hospital arrival; transient ROSC before arrival; signs of life at arrival | Survival to discharge 4.4% (7/159) at score 0; 10.7% (38/356) at score 1; 30.8% (39/130) at scores 2–3. Bias-corrected C-index 0.696 for survival, 0.761 for favourable neurological outcome | Low — internal bootstrap validation only. The most directly useful score in the field, because it addresses the hardest decision: the non-shockable patient |
Pre-ECPR score | 120 normothermic ECPR patients, single centre. Age, no-flow/initial-rhythm composite, total arrest time, signs of life, pupil dilation, regional cerebral oxygen saturation, arterial pH, end-tidal CO₂ | AUC 0.87 (0.77–0.93) in development, 0.79 (0.67–0.88) cross-validated, versus 0.63 (0.54–0.72) for the ELSO example criteria (P = 0.012). Probability above 6.4%: 100% sensitivity, PPV 40.5% | Low — 120 patients, no external validation. Requires cerebral oximetry and a blood gas, so it is not a pre-hospital instrument |
TiPS65 | As reported in the 2025 systematic review | At a cut-off of 4 points, accuracy 88.4% and 88.6% in two studies — but sensitivity 0.172 and 0.193 with specificity 0.971 and 0.985, PPV 0.423 and 0.646, NPV 0.906 and 0.896 | Low. Read the sensitivity, not the accuracy — see the flag below |
RESCUE-IHCA | In-hospital arrest. Patient type, presenting rhythm, time of day, age, renal insufficiency, and duration of arrest | 72% prediction ability; 68% on external validation | Low. The only instrument here developed for in-hospital arrest. Chapter 30 owns it |
Machine-learning model, shockable OHCA | 376 refractory OHCA patients with a shockable rhythm, 301 cannulated, at a quaternary centre; 40% neurologically favourable survival. Most predictive of 11 variables: rhythm at cannulation, intermittent or sustained ROSC, arrest-to-perfusion time, lactate | In-sample AUC 0.89 (misclassification 0.19); out-of-sample AUC 0.80 (misclassification 0.23) | Low — single centre, internal holdout. Note that its top four variables are exactly the ones §20.5 identifies |
Clinical-only versus clinical-plus-laboratory models | Single centre, 2012–2019 | Survival AUROC 0.79 versus 0.83; favourable neurological outcome AUROC 0.80 versus 0.79 — no statistically significant difference | Low. The clinical variables appear to carry the information; the laboratory values add little |
Chinese CSECLS registry model | 990 ECPR patients across 61 hospitals; 351 development, 68 external validation. 64.8% died before discharge. Risk: cerebrovascular disease history, PEA/asystole, higher lactate. Protective: age 45–60, higher pH, intra-aortic balloon pump during ECPR | C-statistic 0.70 internal, 0.65 external, with adequate calibration | Low. The external C-statistic of 0.65 is the realistic expectation for this class of model |
Pitfall — an "accuracy" of 88% with a sensitivity of 17%
The TiPS65 figures are a textbook illustration of why accuracy is the wrong headline metric in a population with a low event rate.
At a cut-off of 4 points the score was reported as 88.4% and 88.6% accurate. Its sensitivity was 0.172 and 0.193. In a cohort where most patients do badly, a rule that says "this patient will do badly" is right most of the time by saying it about almost everyone, and it identifies fewer than one in five of the patients it is supposed to find.
Its specificity — 0.971 and 0.985 — is genuinely high, which means a positive result is informative. Use these instruments through their specificity and their negative predictive value, never through their accuracy.
Danger — what every score in this table shares
All of them were developed in patients who received ECPR. They predict who does well after cannulation. None of them was developed by comparing cannulated patients with similar patients who were not cannulated, which is the comparison a selection instrument requires — and it is precisely the comparison that the treatment-effect model in §20.6 attempted, with a different answer.
These scores are therefore prognostic instruments being used as selection instruments, and the mismatch is the same one §20.6 describes.
Use them to inform the conversation with the family and the team about expected outcome. Be much more careful about using them to refuse a cannulation.
20.8 Writing your unit's criteria
Chapter 19 established that the ECPR decision cannot be made well in the moment and must therefore be made in advance. This section is what "in advance" looks like.
Clinical pearl — seven design rules for a criteria list, drawn from everything above
- Two gates, not one (§20.3). The dispatch-level gate uses what a paramedic knows. The door-level gate uses what the resuscitation room can see. Do not put a variable in gate 1 that cannot be observed at gate 1.
- Put the arrival rhythm in gate 2 (§20.5.3). Most lists ask only about the initial rhythm, and a patient who was in VF and is now in asystole had a 2.7% favourable outcome rate with an adjusted odds ratio of 0.08.
- Make signs of life a route in, not a footnote (§20.5.4). A non-shockable patient with signs of life outperformed a shockable patient without them.
- Subtract periods of ROSC from the elapsed time (§20.5.5), and record transient ROSC explicitly — it roughly triples the odds of a favourable outcome.
- Never let a criterion stop the clock (§20.5.6). Send the gas; do not wait for it. A criterion that costs five minutes has to be worth more than five minutes of low-flow time, and none of them is.
- Build in a named senior override (§20.4). The standard criteria discriminate at an AUC of about 0.63; a list that weak should not be the final word without a human being attached to it.
- Record every refusal and the reason. Every finding in §20.1's danger callout follows from the fact that refused patients are invisible. Your own refusal log is the only dataset that can ever tell you whether your criteria are right.
Danger — two things a criteria list must not become
It must not become a prediction instrument in disguise. The moment a threshold is defended as "this patient will not survive" rather than "this is where we have decided to draw the line", the conversation stops being about a policy that can be audited and becomes an unfalsifiable claim about an individual. §20.6 shows how badly that claim can be wrong.
And it must not become a substitute for the goals-of-care question. The criteria list answers can this work? It does not answer should we? — which is about what this person would have wanted and what the unit can justify. That is Chapter 77, and it is not optional just because there is no time to ask it.
20.9 The coupling between how tight the criteria are and how good the team becomes
This is the part of selection that no individual criterion contains, and the ECPR textbook states it directly:
"When establishing a program, one may want to restrict candidacy to those most likely to survive, in order to maintain support and enthusiasm. However, this will limit experience and skill development among providers within the prehospital, emergency department, and critical care areas. Previous data has shown that volume of ECMO experience is correlated with outcomes, which also appears to apply for ECPR. Thus, institutional criteria may balance the requisite volume for institutional competency with the desire for an optimal proportion of positive outcomes."
Physiology of a programme — the selection loop
Chapter 19 §19.9 established that the unit of intervention in ECPR is the programme, and that outcome tracks volume. This chapter adds the mechanism by which a unit's criteria feed back into its own competence.
Tighten the criteria → fewer cannulations → higher reported survival, which protects institutional support → but fewer cases per operator per year, which is the variable most strongly associated with outcome → and a longer door-to-flow time, which Chapter 19 §19.7 identified as the interval where the largest randomised trial lost its advantage.
Loosen the criteria → more cannulations and more practice → but lower reported survival, which is the number the hospital board will see.
The two goals that pull against each other are "our survival rate" and "our competence", and the criteria list is the dial between them. A unit that optimises purely for its published survival figure can select itself into incompetence.
The volume evidence, and a numerical problem in it. A SAVE-J II analysis of 1,740 patients grouped its 36 centres into tertiles by annual ECPR volume — high (21 or more sessions per year), medium (11–20) and low (fewer than 11) — and reported survival at discharge of 33.4%, 24.1% and 26.8% respectively (P = 0.001), concluding that high-volume centres did best.
Pitfall — the volume finding is not the clean dose–response it is often quoted as
Two problems in the reported figures are recorded here and neither is corrected.
1. The relationship is not monotonic. Survival was 33.4% high, 24.1% medium, 26.8% low. The middle tertile did worse than the lowest. A true volume–outcome gradient should not have that shape, and the abstract offers no explanation.
2. One adjusted odds ratio is internally implausible. The analysis reports adjusted odds ratios of 0.657 (P = 0.003) for medium-volume and 0.983 (P = 0.006) for low-volume centres relative to high-volume. An odds ratio of 0.983 is an effect of under 2%, and it is reported as more significant than an odds ratio of 0.657 in the same model. That cannot both be as printed.
What is used here: only the direction of the high-volume finding, that the highest-volume tertile had the best survival, which is consistent with the independent meta-regression cited in Chapter 19 §19.7. The tertile figures and the two odds ratios are not used. This is the fourth published numerical problem flagged in Part IV.
It also complicates Chapter 19 §19.9's statement that "the volume signal turns at about 10 cases a year." On these data the separation is at roughly 21 per year, and the behaviour below that is not orderly. Chapter 19's claim should be read as "more is better, with the best-supported threshold somewhere between 10 and 21 per year", and an addendum has been appended to that chapter.
20.10 The errors that recur
Error | Why it is wrong |
Treating the criteria list as a validated prediction rule | The ELSO example criteria discriminated at AUC 0.63 (0.54–0.72) in the one study that tested them |
Asking only about the initial rhythm | A patient who was in VF and is in asystole on arrival had a 2.7% favourable outcome rate, adjusted OR 0.08. The arrival rhythm is free and more informative |
Treating signs of life as a tie-breaker | Non-shockable with signs of life (38.7%) beat shockable without them (4.1%) in one series. They are the strongest variable in the field |
Excluding an unwitnessed arrest that presents in VF | An initial shockable rhythm carries 94% certainty of a no-flow interval under 10 minutes. The rhythm is better evidence than the witness |
Counting refractory VF as a favourable feature | No association between number of defibrillations and survival. Intermittent ROSC earns its place; refractory VF does not |
Waiting for a lactate or a gas before cannulating | Adding laboratory values to a clinical model did not improve prediction of neurological outcome. Send it, do not wait |
Refusing on a metabolic threshold and calling it futility | Survivors are reported at pH 6.68 and lactate 19, and the one treatment-effect analysis found lower pH and higher lactate predicted LARGER benefit |
Spending decision time on the comorbidity list | Excluding terminal illness, cancer, end-stage liver disease and severe cognitive impairment moved favourable outcomes from 21.5% to 22.2% |
Applying the ECPR time ceiling to a hypothermic arrest | A different disease. Hypothermia was the commonest cause among favourable outcomes in initial-asystole ECPR |
Cannulating an unexplained non-shockable arrest without considering aorta or brain | No survivors with acute aortic disease or primary cerebral disease in the SAVE-J II PEA analysis, and neither appears in any published criteria list |
Reading a score's "accuracy" | 88% accuracy with 17% sensitivity. Use specificity and negative predictive value |
Using a prognostic score to refuse a cannulation | Every score was built in patients who received ECPR. They predict outcome after cannulation, not benefit from it |
Tightening criteria to protect the survival figure | The criteria list is the dial between "our survival rate" and "our competence." A unit can select itself into incompetence |
Not recording refusals | Refused patients are invisible in every published dataset. Your refusal log is the only way you will ever learn whether your criteria are right |
20.11 Key points
- A criteria list is doing three jobs at once — prognostic selection, treatment-effect selection and operational selection — and they pull in different directions. Most disagreements about ECPR eligibility are disagreements about which job is being done.
- The randomised trials cannot tell you where to draw the lines. Each tested a single narrow inclusion set. Every threshold in current use rests on observational data drawn from patients who were already selected.
- The evidence base is structurally incapable of answering the selection question, because the refused patients are not in it, and because criteria already in use remove the evidence that would refute them.
- Two definitional problems contaminate every number: some series mix patients cannulated during compressions with patients cannulated after ROSC, and "refractory" ranges from 10 to 60 minutes across the literature. Ask both questions of any figure before you benchmark against it.
- The ELSO example selection criteria discriminated at an AUC of 0.63 (0.54–0.72) against a multivariable score's 0.79. Hold the individual thresholds loosely and build in a senior override.
- Age is a poor predictor and a reasonable operational ceiling. Its strange behaviour — a bimodal survival peak, a protective 45–60 band — is explained by aetiology, not by biology: young arrests include dissection and overdose, older arrests include treatable coronary disease.
- An initial shockable rhythm carries 94% certainty that the no-flow interval was under 10 minutes. The rhythm may therefore supersede witnessed status, and an unwitnessed VF arrest should not be excluded on that basis alone.
- Ask about the rhythm now, not the rhythm the ambulance found. Among patients whose initial rhythm was shockable, favourable outcome was 22.4% if still shockable on arrival, 9.5% in PEA, and 2.7% in asystole (adjusted OR 0.08).
- Signs of life are the strongest single pre-cannulation variable and are missing from most criteria lists. Favourable outcome 38.0% versus 8.1% in 1,395 registry patients; adjusted odds ratios from 5.65 to 9.63 across datasets. In one series a non-shockable patient with signs of life (38.7%) did far better than a shockable patient without them (4.1%).
- Signs of life are a strong rule-out and a weak rule-in: negative predictive value 0.96, positive predictive value 0.22. Their absence argues hard against ECPR in a non-shockable patient; in a shockable patient it is a serious negative, not a veto.
- Transient ROSC — any palpable pulse lasting a minute — roughly triples the odds of a favourable outcome (adjusted OR 3.34) and should be subtracted from the elapsed low-flow time. Refractory VF, by contrast, confers nothing beyond initial VF.
- Metabolic thresholds are weak, late and possibly inverted. Survivors exist at pH 6.68 and lactate 19; adding laboratory values to a clinical model did not improve prediction of neurological outcome; and the only treatment-effect analysis found lower pH and higher lactate identified LARGER benefit.
- Comorbidity exclusions barely move the yield — 21.5% to 22.2% for the four commonest — and a past history of cardiovascular disease is associated with better outcomes, because the comorbidity that predicts a good result is the one that explains the arrest.
- Aetiology decides whether the bridge has a destination. Pulmonary embolism and myocarditis do well; acute aortic disease and primary cerebral catastrophe had no survivors and appear in no published criteria list. Hypothermic arrest is a different disease and the time limits do not apply.
- Prognosis and treatment effect can rank patients in opposite directions, and one causal-forest analysis found exactly that. When refusing a patient, know whether you are saying "they will probably die anyway" or "ECPR would not help them" — the evidence supports the first far better than the second.
- Every published score predicts outcome after cannulation, not benefit from it. Use them for the family conversation; be much more careful using them to refuse.
- The criteria list is the dial between the unit's survival figure and the unit's competence. Tighten it and the numbers improve while the team practises less — and volume is among the strongest correlates of outcome in the whole field.
- Record every refusal and its reason. It is the only dataset that can tell you whether the list on your wall is right, and nobody else's data will ever contain it.
Cross-references
Backwards
- Chapter 4 — pre-ECMO optimisation and the three questions. The structure of a selection decision under uncertainty is the same; the time available is not.
- Chapter 11 §11.3 — "prognosis is not indication." §20.6 is the quantitative test of that principle, and it survives it.
- Chapter 19 §19.1 — why the ECPR decision must be made in advance, and the undefined word "refractory" that §20.2 returns to.
- Chapter 19 §19.2 — the no-flow / low-flow vocabulary, the compression physiology, and the end-tidal CO₂ drug confounder that §20.5.6 depends on.
- Chapter 19 §19.3 and §19.4 — the low-flow gradient, the rhythm-specific cut-points, and the sixty-minute controversy. §20.5 assumes all of it.
- Chapter 19 §19.7 and §19.9 — the volume–outcome evidence and the programme as the unit of intervention, which §20.9 extends into the criteria themselves. An addendum to §19.9 has been appended following the volume analysis discussed in §20.9.
- Chapter 30 — prognostic scoring as a subject, including RESCUE-IHCA and the still-uncovered SAVE score.
Forwards
- Chapter 21 — cannulation during chest compressions. Everything in §20.8's gate 2 happens while the wire is going in.
- Chapter 22 — the first hours after flow, including the reperfusion decisions Chapter 19 §19.5 identified as ownerless.
- Chapter 23 — neurological assessment and prognostication after cannulation. The pupillary findings in §20.5.4 are a selection variable here and a prognostic variable there, and the SAVE-J II pupillary analysis belongs to both: among 723 patients with a negative pupillary light reflex and mydriasis on arrival, 10.2% still achieved a favourable neurological outcome, and a positive pupillary light reflex at ICU admission carried an odds ratio of 11.3 (5.17–24.7) while pupil diameter at that point did not predict anything.
- Chapter 24 — failure, withdrawal and organ donation. §20.8's refusal log and Chapter 24's withdrawal decisions are the two ends of the same policy.
- Chapter 60 — transport, where the dispatch-level gate is applied in practice.
- Chapter 65 — accidental hypothermia, the aetiology that suspends every time limit in this chapter.
- Chapter 77 — resource allocation and ethics. §20.9's dial between survival figures and competence is a resource question, and §20.8's danger callout hands the goals-of-care question there.
References
All rows seeded from this chapter are marked Verified = No — every source below is from a structured abstract.
SAVE-J II and its secondary analyses — a retrospective registry of 2,157 adults receiving ECPR across 36 Japanese institutions, 2013–2018
- Bunya N, et al. Prognostic significance of signs of life in out-of-hospital cardiac arrest patients undergoing extracorporeal cardiopulmonary resuscitation. Critical Care Medicine. 2023. DOI 10.1097/CCM.0000000000006116.
- Tominaga N, et al. Factors associated with favourable neurological outcomes following cardiopulmonary resuscitation for out-of-hospital cardiac arrest. Resuscitation Plus. 2024. DOI 10.1016/j.resplu.2024.100574.
- Otani T, et al. Transient return of spontaneous circulation related to favourable outcomes in out-of-hospital cardiac arrest patients resuscitated with extracorporeal cardiopulmonary resuscitation. Resuscitation Plus. 2022. DOI 10.1016/j.resplu.2022.100300.
- Shirasaki K, et al. Clinical characteristics and outcomes after extracorporeal cardiopulmonary resuscitation in out-of-hospital cardiac arrest patients with an initial asystole rhythm. Resuscitation. 2023. DOI 10.1016/j.resuscitation.2023.109694.
- Ijuin S, et al. Analysis of factors associated with favorable neurological outcomes in patients with initial pulseless electrical activity who underwent ECPR. Journal of Critical Care. 2024. DOI 10.1016/j.jcrc.2024.154917.
- Kawauchi A, et al. Clinical prediction score for patients with initial nonshockable rhythm receiving extracorporeal cardiopulmonary resuscitation for out-of-hospital cardiac arrest: development and internal validation (the START-ECPR score). Journal of the American Heart Association. 2025. DOI 10.1161/JAHA.125.042734.
- Ijuin S, et al. Association between conversion from an initial shockable rhythm to pulseless electrical activity before extracorporeal cardiopulmonary resuscitation and outcome. American Heart Journal. 2025. DOI 10.1016/j.ahj.2025.08.012.
- Inoue K, et al. Waveform conversion as a prognostic factor of poor prognosis in patients undergoing extracorporeal cardiopulmonary resuscitation. American Journal of Emergency Medicine. 2025. DOI 10.1016/j.ajem.2025.03.041.
- Shirasaki K, et al. Prognostic factors for out-of-hospital cardiac arrest patients with prolonged low-flow time undergoing extracorporeal cardiopulmonary resuscitation. European Heart Journal: Acute Cardiovascular Care. 2025. DOI 10.1093/ehjacc/zuaf072. 708 non-hypothermic patients with low-flow time of 60 minutes or more; 10.0% achieved a favourable neurological outcome. The direct evidence behind Chapter 19 §19.4.
- Hamaguchi T, et al. Association between pupillary examinations and prognosis in patients with out-of-hospital cardiac arrest who underwent extracorporeal cardiopulmonary resuscitation. Annals of Intensive Care. 2024. DOI 10.1186/s13613-024-01265-7. Held for Chapter 23.
- Kawauchi A, et al. Sex differences in extracorporeal cardiopulmonary resuscitation for out-of-hospital cardiac arrest. Critical Care. 2024. DOI 10.1186/s13054-024-05086-9. 1,819 patients; female sex independently associated with favourable neurological outcome, adjusted OR 1.60 (1.05–2.43). Not used as a selection criterion anywhere and not proposed as one here.
- Misumi K, et al. Impact of center volume on in-hospital mortality in adult patients with out-of-hospital cardiac arrest resuscitated using extracorporeal cardiopulmonary resuscitation. Scientific Reports. 2024. DOI 10.1038/s41598-024-58808-y. Two numerical problems flagged in §20.9 and not corrected.
Signs of life
- Bunya N, et al. Signs of life as a favorable predictor for non-shockable cardiac arrest undergoing extracorporeal cardiopulmonary resuscitation. American Journal of Emergency Medicine. 2024. DOI 10.1016/j.ajem.2024.10.046. The four-cell table in §20.5.4. Its two adjusted odds ratios are not used — see the flag.
- Debaty G, et al. Prognostic value of signs of life throughout cardiopulmonary resuscitation for refractory out-of-hospital cardiac arrest. Resuscitation. 2021. DOI 10.1016/j.resuscitation.2021.02.022. Pooled data from three French cohorts, 434 ECPR recipients. Disagrees with SAVE-J II about gasping.
Scores, models and syntheses
- Redfors B, et al. The Pre-ECPR score: developing and validating a multivariable prediction model for favorable neurological outcomes in patients undergoing extracorporeal cardiopulmonary resuscitation. Journal of Cardiothoracic and Vascular Anesthesia. 2024. DOI 10.1053/j.jvca.2024.09.009. The only published head-to-head test of the ELSO example criteria. The source of the AUC 0.63 finding in §20.4.
- Chen C-H, et al. Optimizing extracorporeal cardiopulmonary resuscitation candidate selection in out-of-hospital cardiac arrest: a machine-learning individualized treatment effect approach versus rule-based criteria. Journal of the American Heart Association. 2026. DOI 10.1161/JAHA.125.047815. The source of §20.6.
- Crespo-Diaz R, et al. Machine learning identifies higher survival profile in extracorporeal cardiopulmonary resuscitation. Critical Care Medicine. 2024. DOI 10.1097/CCM.0000000000006261.
- Chiu C-C, et al. Comparing clinical only and combined clinical laboratory models for ECPR outcomes in refractory cardiac arrest. Scientific Reports. 2025. DOI 10.1038/s41598-025-87200-7. Adding lactate and pH did not improve prediction of neurological outcome.
- Li Z, et al. Mortality risk factors in patients receiving ECPR after cardiac arrest: development and validation of a clinical prognostic prediction model. American Journal of Emergency Medicine. 2023. DOI 10.1016/j.ajem.2023.11.048. Chinese CSECLS registry, 61 hospitals.
- Woszczyk D, et al. Neurological outcome predictors after extracorporeal cardiopulmonary resuscitation: a systematic review. Systematic Reviews. 2025. DOI 10.1186/s13643-025-02818-y. 8 studies, 4,353 patients. Source of the pooled operating characteristics for signs of life and of the TiPS65 figures.
- Holmberg MJ, et al. Extracorporeal cardiopulmonary resuscitation for cardiac arrest: an updated systematic review. Resuscitation. 2022. DOI 10.1016/j.resuscitation.2022.12.003. The source of this chapter's opening quotation.
- Maekawa K, et al. Extracorporeal cardiopulmonary resuscitation for patients with out-of-hospital cardiac arrest of cardiac origin: a propensity-matched study and predictor analysis. Critical Care Medicine. 2013. DOI 10.1097/CCM.0b013e31827ca4c8. Pupil diameter on hospital arrival, adjusted hazard ratio 1.39 per 1 mm increase (1.09–1.78).
Protocols and guidance — all second-hand
- Richardson ASC, Tonna JE, Nanjayya V, et al. Extracorporeal cardiopulmonary resuscitation in adults: interim guideline consensus statement from the Extracorporeal Life Support Organization. ASAIO Journal. 2021;67:221–228. Not retrieved. The ELSO example inclusion criteria are read through Red Book Table 32-1. Still the highest-value outstanding target in Part IV.
- Belohlavek J, et al. The Prague hyperinvasive approach to out-of-hospital cardiac arrest. Journal of Translational Medicine. 2012;10:16. Reproduced as Appendix C of the ECPR textbook; not retrieved.
- Pasquier M, Hugli O, Paal P, et al. Hypothermia outcome prediction after extracorporeal life support for hypothermic cardiac arrest patients: the HOPE score. Resuscitation. 2018;126:58–64. Not retrieved. Held for Chapter 65.
- Saczkowski RS, Brown DJA, Abu-Laban RB, et al. Prediction and risk stratification of survival in accidental hypothermia requiring extracorporeal life support: an individual patient data meta-analysis. Resuscitation. 2018;127:51–57. Not retrieved. Held for Chapter 65.
Studies cited through the ECPR textbook, not retrieved and identifiers not established
- The 423-patient European cohort that supplies the age, rhythm and defibrillation-count findings. [VERIFICATION REQUIRED] — the textbook's body text attributes the defibrillation finding to "Linz and colleagues" while the corresponding reference in its own list is Lunz D, et al., Intensive Care Medicine, February 2020. The spelling discrepancy is recorded and neither form is asserted. This is the seventh bibliographic-metadata problem logged in this book.
- The cohorts supplying the unwitnessed-arrest figures (1 of 16; 3 of 21; 4 of 12), the bystander-CPR figures (3.8%; 13%; 2 of 12), the stepwise comorbidity-exclusion analysis (21.5% to 22.2%), the lactate and pH observations (94% mortality at lactate 13 or above; survival at lactate 19 and pH 6.68), the 94%-certainty no-flow inference, and the aetiology series. All read through ECPR and Resuscitative ECMO Chapter 3. [VERIFICATION REQUIRED].
Textbooks
- Shinar Z, Badulak J. ECPR and Resuscitative ECMO. Chapter 3 — the component-by-component review of ECPR eligibility, and Figures 1, 2, 3 and 5 reproducing the Prague, Minnesota and Vancouver protocols. The primary source for this chapter. [VERIFICATION REQUIRED] — page numbers not confirmed.
- ELSO Red Book, 6th edition, Chapter 32 — Table 32-1, the example inclusion criteria; Table 32-2; the RESCUE-IHCA score; the timing statements. [VERIFICATION REQUIRED].
- ISCCM Manual of RRT and ECMO in ICU, Chapter 40 — indications as bridge to recovery, transplantation or decision; the minimum institutional criteria; the statement that all contraindications to routine ECMO apply to ECPR; accidental hypothermia and cardiotoxic poisoning as additional out-of-hospital indications. [VERIFICATION REQUIRED].
Chapter status
Drafted and audited 11 September 2026. Ten-pass quality control completed: clinical, physiology, evidence, citation, numerical, safety, contradiction, redundancy, bedside utility and literature-currency passes.
This chapter deliberately does not restate Chapter 19. The time criteria, the low-flow gradient and the sixty-minute controversy live there and are referenced, not repeated. What is new here is every other criterion, weighed one at a time against its evidence.
Two controversies were set out rather than smoothed. §20.4 reports that the ELSO example selection criteria — the nearest thing to an international standard, and the ancestor of most units' lists — discriminated at an area under the curve of 0.63 in the only study to have tested them, and argues that this is a reason to hold the thresholds loosely and build in a senior override rather than to abandon the list. §20.6 sets prognostic selection against treatment-effect selection using a causal-forest analysis in which lower pH and higher lactate identified the patients ECPR helped most, and concludes that the principle is sound even though the specific model is weak, because the limitation is structural: an observational literature that only ever measures outcome among the cannulated cannot, in principle, say whom to cannulate.
The numerical audit flagged three problems and corrected none. Two adjusted odds ratios of 34.33 and 96.51 were computed in a 227-patient study across a cell containing zero events and are not used; the four proportions are. A centre-volume analysis reports a non-monotonic survival gradient (33.4%, 24.1%, 26.8%) and an adjusted odds ratio of 0.983 with P = 0.006 — an effect of under 2% reported as more significant than an effect of 34% in the same model; only the direction of the high-volume finding is used. A seventh bibliographic-metadata problem was logged, a first-author spelling discrepancy between a textbook's body text and its own reference list, and no identifier was reconstructed.
The contradiction audit recorded one genuine disagreement that is printed rather than resolved: SAVE-J II and the pooled French analysis disagree about whether gasping carries signal (OR 4.33, 2.57–7.28 versus OR 1.75, 0.95–3.21). The composite — any sign of life — survives the disagreement; the components do not.
An addendum has been appended to Chapter 19 §19.9, because the SAVE-J II centre-volume analysis puts the separation at roughly 21 cases per year rather than the 10 quoted there, and does not behave monotonically below it.