The question this chapter answers
Part V has spent five chapters teaching how to read one instrument at a time — the haemodynamic monitor, the echocardiogram, the blood gas, the cerebral monitors, the laboratory panel. This chapter is about what happens when you have all of them at once, and what happens to the numbers afterwards.
The clinical form of the question is ordinary and asked every day: the console says one thing, the echo says another, and the laboratory says a third. Which do I believe? The second form is asked much less often and matters just as much: the paper I am about to change my practice on was built out of numbers like the ones on my flow sheet. What survived the journey, and what did not?
Evidence search date: 22 September 2026. Primary sources for this chapter are the ELSO Red Book 6th edition (Chapter 58, Research; Chapter 60, The ELSO Registry; Chapter 55, Quality Improvement), ECMO: A Practical Guide to Management (programme documentation and charting), ECPR and Resuscitative ECMO (daily bedside ECMO rounds), and ECMO in the Adult Patient (registry structure and the clinical data analyst role). External evidence was retrieved through PubMed and a full-text academic index. One external source was read as complete full text and is marked Verified in the Evidence Database; the remainder are abstracts and are marked accordingly.
What this chapter covers — and what it does not
This chapter owns | This chapter does not own |
How the measurements from Chapters 25 to 29 are combined into one read, and what to do when two of them disagree | How to perform or interpret any single measurement — Chapters 25, 26, 27, 28 and 29 own their own instruments and this chapter re-teaches none of them |
The measurement pyramid: what the console measures, what the flow sheet keeps, what the registry receives, and what a published analysis finally uses | The circuit hardware that generates the signals — Chapter 3 owns the circuit and Chapters 31 to 38 own what goes wrong with it |
The three levels of ECMO documentation, and what each is for | The operational artefacts themselves. Chapter 87 owns the daily checklist, Chapter 88 the morning round template, Chapter 92 the handover template. This chapter supplies the reasoning those three are built from |
Registry structure, registry safeguards and registry limitations — a scope gap identified during the Chapter 28 audit and closed here | Any individual registry finding. Registry-derived numbers belong to the chapter that owns the clinical question |
How to read a number in an ECMO paper — denominators, missingness, immortal time, subgroups, and the difference between a definition and a measurement | The evidence tables themselves — Appendix 10, Appendix 11 and Appendix 12 hold them. This chapter gives them their reading rules |
Prediction models built on ECMO data, and the specific ways they mislead | Prognostic scores used for patient selection — Chapter 11 owns SAVE and SCAI, Chapter 4 owns RESP, and Chapter 20 owns the ECPR prediction literature |
30.1 The measurement pyramid
An adult on extracorporeal support is one of the most heavily instrumented patients in medicine. The console measures pump speed, blood flow, and in most modern systems inlet pressure, outlet pressure and transmembrane pressure, continuously. A bedside monitor measures the electrocardiogram, arterial pressure, peripheral saturation and often central venous oxygen saturation, continuously. The ventilator measures pressure, flow and volume breath by breath. The heat exchanger sets and reports temperature. Around that sit the intermittent measurements — the gas, the panel, the echocardiogram, the cerebral oximeter.
Almost none of it is kept.
The structural claim of this chapter · Every level of the pyramid loses about an order of magnitude, and every loss is biased in the same direction
Consider a single 10-day run, and follow one number down.
- The console measures continuously. At even one sample per second, ten days of pump flow is roughly 864,000 values. On almost every ECMO system in routine clinical use, these are displayed and then discarded; the console holds a short trend and overwrites it.
- The flow sheet records every one to two hours. ECMO: A Practical Guide to Management states the standard directly: flow sheet data are entered by the bedside specialist "at regular intervals (generally every 1 to 2 h)", and it calls this level of documentation a standard of care. Ten days at hourly entry is about 240 values. That is a reduction of roughly 3,600-fold, and it is the largest single loss in the chain.
- The registry receives a summary of the run. The ELSO Registry records a defined set of elements per run — demographics, configuration, indication, duration, complications as categories, outcome. For most variables this is one value, or a small number of values, per run. Another reduction of roughly 100-fold.
- A published analysis uses whatever was not missing. The 2022 ELSO Registry International Report states its own method plainly: "Missing values were excluded before calculating descriptive statistics." Source: Tonna JE, Boonstra PS, MacLaren G, Paden M, Brodie D, Anders M, Hoskote A, Ramanathan K, Hyslop R, Fanning JJ, Rycus P, Stead C, Barrett NA, Mueller T, Gomez RD, Malhotra Kapoor P, Fraser JF, Bartlett RH, Alexander PMA, Barbaro RP. "Extracorporeal Life Support Organization Registry International Report 2022: 100,000 Survivors." ASAIO J 2024;70(2):131-143. DOI 10.1097/MAT.0000000000002128. Abstract only.
None of these losses is an error. Each is a reasonable engineering or administrative decision taken for a good local reason. Together they are the reason the field knows far less than its data volume suggests.
And the losses are not random. At every level, what is lost is preferentially what was brief, what happened between entries, what the patient did not survive to have recorded, and what the centre did not choose to report. What survives is the stable, the charted, the completed and the reported. That is a bias with a direction, and the direction is toward reassurance.
Clinical pearl · The interval between entries is the resolution of the record, and it is the resolution of your memory too
A circuit event that begins and resolves between two hourly entries did not happen, as far as any later reader of the chart is concerned — including you, tomorrow.
This has an immediate bedside consequence that costs nothing: when something transient happens, write it down with a time, even though it has resolved. A flow that swung, an inlet pressure that went sharply negative during coughing or turning, a brief desaturation, an alarm that silenced itself. ECMO: A Practical Guide to Management includes exactly this in the flow sheet standard: "Histories of alarms or other circuit problems should be both documented and reported to the ECMO team providers."
The circuit problem that declares itself as an emergency at 03:00 almost always announced itself earlier as something transient that nobody recorded. Chapter 31 and Chapter 34 both depend on this, and neither can be practised from a chart that contains only the values that were stable enough to be typed in.
Pitfall · Treating a charted value as a measurement of the hour it sits in
A value on an hourly flow sheet is a measurement of a moment, labelled with an hour. Three things follow, and all three are routinely forgotten.
- A value recorded after an intervention describes the intervention, not the problem. The flow written at 14:00 in the row that also contains "repositioned cannula" is the flow after repositioning. The flow that prompted the repositioning is frequently never written at all.
- A value that was normal gets charted promptly; a value that was alarming gets charted once somebody has finished dealing with it. The delay is not constant, and it correlates with severity.
- A blank is not a normal value. It usually means nobody was free. On a flow sheet, the rows with gaps in them are the rows describing the worst hours of the run.
This is the bedside form of the missingness problem that §30.7 meets again at registry scale. Missingness on an ECMO record is informative missingness, at every level of the pyramid.
30.2 Level 2: the three levels of documentation, and what each is for
ECMO: A Practical Guide to Management sets out the documentation structure of an ECMO programme in three levels. It is the clearest statement of the subject in any of this book's primary sources, and the whole of §30.10 is built on it.
Level | Who writes it, and when | What it contains | What it is for |
1. Programme documentation | The programme, maintained continuously | Written description of all processes, guidelines and protocols: how emergencies are handled, inventory, training guidelines, expectations for all roles, anticoagulation, weaning, the consult process, acceptance criteria. Plus gathered data on patient outcomes and complications, team training and evaluations, and consult requests and outcomes | "Everything that is important to the smooth functioning of the program should be written and readily available to every member of the team as a reference." It is also a required part of becoming an ELSO Center of Excellence, and it feeds the quality assurance cycle — data are analysed and compared against the team's established guidelines and predetermined goals, which then informs revision of the guidelines |
2. Flow sheet data in the patient's chart | The bedside specialist, generally every 1 to 2 hours | All current pump parameters (revolutions per minute, flows), sensor readings, blood gas results, and haemodynamics if not automatically captured. At the same time, circuit checks: cannula position at the skin, insertion-site integrity, presence of clot in the line, oxygenator or pump, and presence and correct position of any clamps. Plus histories of alarms or other circuit problems, both documented and reported to the ECMO team providers | "This level of documentation is a standard of care." It is also the only layer with enough time resolution to reconstruct an event |
3. Provider documentation | The ECMO team providers, daily | The progress of the ECMO run apart from the non-associated critical care issues: weaning attempts and other adjustments, complications, current anticoagulation strategy, current setting snapshots | Meets documentation and billing requirements, and "provide information for later analysis of things like complication rates and timing of weaning that may be useful for program improvement." The same source recommends that the providers and the Programme Director agree a common standard for it |
Clinical pearl · The single most useful line in the three-level structure is the one that separates the run from the patient
Level 3 exists to document the progress of the ECMO run apart from the non-associated critical care issues. That separation is worth adopting whether or not your billing system requires it, because it changes what gets thought about.
A combined note written by a busy team records the sodium, the antibiotics and the family meeting, and the circuit appears only when it misbehaves. A separate circuit entry forces a daily answer to the questions nobody asks on a stable day: what is the transmembrane pressure doing compared with three days ago, what is the sweep doing compared with the carbon dioxide, has the anticoagulation strategy actually changed or has it only drifted, and when was the last weaning assessment.
ECPR and Resuscitative ECMO describes the institutional version of the same idea, adopted from the Alfred Hospital in Melbourne: separate daily bedside ECMO rounds, in addition to the ordinary intensive care round. Its stated purposes are the exchange of ideas between faculty on complex issues and the review of fundamental management concepts for learners — and it adds routine morbidity and mortality reviews with set criteria for whom to evaluate.
Chapter 88 owns the round template. What belongs here is the argument for holding the round at all, and the observation that a circuit reviewed only when it alarms is a circuit with no trend.
Pitfall · Charting the setting instead of the measurement
Two different things get written in the same column and they are not interchangeable.
- A setting is what someone asked the machine to do: pump revolutions per minute, sweep gas flow, set fraction of inspired oxygen to the oxygenator, target temperature.
- A measurement is what the machine reports back: blood flow, inlet and outlet pressure, transmembrane pressure, delivered temperature.
The clinically informative quantity is almost always the relationship between the two, and that relationship is invisible if only one is charted. Revolutions per minute without flow says nothing about circuit resistance or drainage; flow without revolutions per minute says nothing about what the pump is having to do to achieve it. A falling flow at unchanged revolutions per minute is a circuit finding; the same flow at reduced revolutions per minute is a decision. Chapters 31 to 34 rest entirely on this distinction, and a flow sheet that carries only one of the pair cannot support them.
The practical rule: every set value on the ECMO chart should have its measured partner on the same line. Speed with flow. Sweep with post-membrane carbon dioxide. Set temperature with delivered temperature.
30.3 Reading Part V as one panel: the integrated read
Chapters 25 to 29 each produced a rule about their own instrument. Read together, they produce a single rule that is stronger than any of them, and it is the rule this section is about.
Chapter | Its instrument | The caution it carries (summarised here; each chapter is authoritative for its own) | What it contributes to the combined read |
Chapter 25 | Haemodynamic monitoring | On a circuit the usual haemodynamic signals are generated partly by the pump, so their conventional interpretation does not carry over unchanged | Pressure and flow have come apart. Blood pressure is a poor proxy for systemic flow when a pump supplies much of it, and pulse pressure becomes a statement about the native heart rather than about volume |
Chapter 26 | Echocardiography | Echocardiography is the only instrument that sees the heart itself, and its weaning criteria descend from a single small cohort | The only direct look at the organ. Also the instrument most dependent on who is holding the probe, so its findings need a recorded image, not a recorded conclusion |
Chapter 27 | Blood gas and oxygenation | Where the sample was drawn determines what it means; the numbers describe a mixing problem, not a lung | Site is part of the value. A gas without its sampling site is uninterpretable and should not be charted as though it were a result |
Chapter 28 | Cerebral monitoring | Every instrument is a rule-in test; none can reassure. Every published complication rate is a detection rate | Asymmetry between instruments is more informative than any of them. And a normal cerebral panel licenses no reduction in suspicion |
Chapter 29 | Laboratory monitoring | Every assay was validated by changing one variable at a time; the useful comparison is the patient's own earlier value, ideally pre-cannulation | Specific assays keep their meaning; global ones do not. And a disagreement between two assays localises the abnormality |
The rule Part V reached from five directions
On extracorporeal support, the reference interval on the report describes a person who was never here. The comparison that carries information is this patient's own value, earlier.
Chapter 25 reached it through haemodynamics whose normal ranges assume a native circulation. Chapter 26 through weaning criteria derived from a population of fifty-one. Chapter 27 through gases whose meaning depends on a sampling site the reference interval knows nothing about. Chapter 28 through complication rates that are detection rates. Chapter 29 through assays validated one variable at a time, and through free haemoglobin and lactate dehydrogenase that were already abnormal before the pump started.
Five instruments, five literatures, one conclusion. That convergence is the strongest thing Part V produces, and it is worth more than any individual threshold in it.
Its operational form is a column, not a sentence. A panel displayed as today-only is a list of values against reference intervals. The same panel displayed with yesterday, day 3, and pre-cannulation alongside is a set of trajectories, and the trajectories are what the five chapters were about. Chapter 88 should specify the columns; this chapter specifies why there must be more than one.
Worked example · The same three numbers, read four ways
A venoarterial patient on day 3. Three values are available: circuit flow 3.2 L/min, mixed venous oxygen saturation 58 percent, arterial lactate 4.1 mmol/L.
Read as a list against reference intervals, two of the three are abnormal and the conclusion is "inadequate perfusion — increase the flow." That is the reflex, and on a busy round it is what happens.
Read as a set — which is Chapter 2's Fick reasoning, not this chapter's — a low mixed venous saturation with a raised lactate is consistent with delivery failing to meet consumption, and the lever is delivery: flow, haemoglobin or oxygenation. Still incomplete, but no longer reflexive.
Read against the patient's own earlier values, the picture can invert. If flow was 4.4 L/min yesterday at the same pump speed, the finding is not the saturation — it is that the circuit is delivering less for the same work, and the question is drainage, cannula position, volume status or a developing resistance, none of which is treated by turning the dial up. If lactate was 8.0 yesterday, a value of 4.1 is the best it has been.
Read against the other instruments, the differential narrows further. An echocardiogram showing a distending left ventricle points one way (Chapter 15); a mixed venous sample drawn from a site the configuration makes uninterpretable points at the measurement rather than the patient (Chapter 27); a rising free haemoglobin with a falling haptoglobin points at the circuit (Chapter 29).
The teaching point: the same three numbers supported four different actions, and the difference between them was entirely context that a single-column display removes. Chapter 30 has no opinion about which action was right in this invented case. It has a strong opinion about the display.
This example is constructed to illustrate the reasoning. The values are illustrative and are not drawn from a published report.
30.4 Discordance as a diagnostic instrument
Chapter 29 established this for two assays: when the anti-factor Xa and the activated partial thromboplastin time disagree, the disagreement localises the abnormality, because the two tests are sensitive to different things. This section generalises it across the whole of Part V, because the generalisation is where most of the diagnostic value in an ECMO monitoring panel actually sits.
Physiology · Why instruments on a circuit disagree more than instruments in an intact patient, and why that is useful
In a native circulation, most monitors are downstream of the same thing. Cardiac output, blood pressure, mixed venous saturation, lactate and urine output are all reading, at different removes, from one pump and one set of lungs. They move together, and when they do not, the commonest explanation is that one of them is wrong.
A patient on extracorporeal support has two circulations in parallel, joined at a mixing point whose position moves (Chapter 3). The consequences run through all of Part V.
- A monitor's reading now depends on where it is relative to the mixing point. Two saturation probes on the same patient are measuring genuinely different blood (Chapter 16, Chapter 27). They are not disagreeing; they are answering different questions.
- Systemic flow and native cardiac output have come apart. A rising arterial pulse pressure with unchanged circuit flow means the native heart is contributing more, which is a finding and not a discrepancy (Chapter 18).
- Delivery and consumption are no longer coupled through one organ. The oxygenator can be fixing the gas while the lung gets worse, or the reverse (Chapter 27, Chapter 32).
So on ECMO, two instruments disagreeing is the normal state, and the pattern of disagreement carries the diagnosis. The corollary is uncomfortable and worth stating: a monitor that always agrees with another monitor is adding no diagnostic information, and two concordant readings are not two pieces of evidence. That does not make it useless — a redundant channel still guards against the failure of the other one, which is why two saturation probes are standard — but it is a reason to stop treating agreement as confirmation.
Discordance | What it usually means | Where it is owned |
Two pulse oximetry sites disagree (right hand versus lower limb) on venoarterial support | Differential hypoxaemia — the mixing point has moved distally and the upper body is being perfused by a recovering, poorly oxygenating native heart | Chapter 16; the measurement caveats are Chapter 27 |
Circuit flow falls while pump speed is unchanged | A circuit or drainage problem, not a decision. Inlet pressure, cannula position, volume status, developing resistance | Chapters 31 to 34; the charting rule is §30.2 |
The activated clotting time moves without a heparin change, and the anti-factor Xa does not | Not heparin. Platelets, fibrinogen, haematocrit, temperature — the things the activated clotting time is actually correlated with | Chapter 29 for the measurement, Chapter 42 for the decision |
Free haemoglobin rises while lactate dehydrogenase and haptoglobin do not move | Usually a sampling artefact rather than intravascular haemolysis | Chapter 29 |
Near-infrared spectroscopy is asymmetric while the rest of the neurological panel is unremarkable | A rule-in signal for a focal cerebral problem. High specificity, low sensitivity — so it rules in and cannot rule out | Chapter 28 |
Echocardiography shows recovery while the weaning trial fails | The echocardiographic criteria were derived in a small cohort and one cohort found them running backwards. The trial is the harder evidence | Chapter 26 for the criteria, Chapter 18 for the weaning decision |
Post-membrane gas is good while the patient's arterial gas is not | The oxygenator is working. The problem is recirculation, native lung, cardiac output or mixing — not the membrane | Chapter 27, and Chapter 9 for the differential |
Clinical pearl · Three questions that resolve most discordances at the bedside, in order
Before deciding which instrument to believe, ask them in this order. The order matters, because each one is cheaper than the next and each one removes a large fraction of cases.
- "Are these two instruments measuring the same blood?" On a parallel circulation the answer is frequently no, and if it is no, there is no discordance to resolve — only two correct answers to two different questions. This single question disposes of most apparent contradictions on venoarterial support.
- "Which of these two has the better operating characteristics for the thing I am actually worried about?" Part V supplies these chapter by chapter. A specific assay beats a global one (Chapter 29). A high-specificity test rules in and cannot rule out (Chapter 28). A direct image of the organ beats an inference about it (Chapter 26).
- "What did each of them do yesterday?" A discordance that is new is a finding. A discordance that has been present since cannulation is usually anatomy, configuration or calibration.
If all three fail to resolve it, the answer is to repeat the cheaper measurement rather than to escalate to the more expensive one — which is also Chapter 29's instruction about a first abnormal free haemoglobin, and Chapter 27's about an unexpected gas.
30.5 Level 3: the registry — what it is, how it is built, and what it can hold
This section closes a scope gap. The Chapter 28 audit established that the book depended on registry-derived numbers in several places while no chapter owned registry structure or its limitations. Chapter 30 owns them, and every later chapter that quotes a registry figure should cite this section for the caveat rather than restating it.
Evidence · How the ELSO Registry is actually constructed
The ELSO Red Book 6th edition, Chapter 60, describes the registry's own machinery. It is more rigorous than most clinicians assume, and knowing the machinery is what makes the limitations legible.
Scale. The registry began in 1989, with data going back to 1976. ECMO in the Adult Patient records that data entry is voluntary and requires membership, that members pay an annual fee which entitles them to enter data and receive regular reports including comparison of outcomes with similar centres, and that the registry was initially essentially paediatric. Red Book Chapter 58 records over 175,000 patients at the start of 2022; the 2022 International Report records 154,568 ECMO runs entered from 2009 to 2022, from 780 centres over that period and 557 submitting in 2022, with the median annual number of adult runs per centre rising from 4 to 15 while paediatric and neonatal runs per centre fell from 12 to 7.
Reliability, first mechanism: definitions. Since 2017 the ELSO Registry Definitions Document is a publicly available resource that identifies mandatory data elements, defines individual data points, sets out data entry rules, and illustrates examples designed to clarify nuance between clinical scenarios, updated as fields change.
Reliability, second mechanism: certification. Since 2018, all individuals involved with management, oversight or collection of registry data are required to complete a data entry certification exam, in patient-vignette multiple-choice format, designed to ensure data definitions are clear and consistently applied.
Validity, first tier: internal. Data elements are validated at the time of entry via programmed hard and soft limits, flagging values outside prescribed limits for review and prompting the user to address them before submission.
Validity, second tier: external, and still a pilot. ELSO "has launched a pilot program with a small number of member centers" for external validation through a vendor. The goal is to validate 10 percent of each centre's patients from the previous year (minimum 5, maximum 25) using 64 data elements selected for audit, with centres allowed to adjudicate discrepancies. All of it is explicitly non-punitive.
Governance of change. Additions and revisions go through a Registry Development Advisory Group, or through optional addenda — currently Cardiac, ECPR, SARS-CoV-2 and Trauma. Published guidelines for registry changes require a demonstrated need, agreement to participate from 10 percent of active centres, and consideration of funding. Updates are communicated biannually.
Benchmarking. Since a 2016 collaboration with a commercial analytics provider, the ELSO Quality Reporting Platform displays risk-adjusted centre-specific outcomes benchmarked against peer institutions, with no fewer than 10 centres presented as comparators and only the member centre identified. Reports are available by age cohort, by unit, by run duration, per patient or per ECLS run — because some patients have multiple runs in one admission — with complications clustered or individual, time to complication, and complications per 1000 ECLS hours.
Certainty: high that this is how the registry is built. All of it is the registry's own description of itself in a primary source.
Danger · Four properties of registry risk adjustment that change how a benchmark should be read
These come from the Red Book's own account of the ELSO Quality Reporting Platform, and each one is invisible on the output graphic.
- The pragmatic risk adjustment models were built with candidate variables "that were missing under 10 percent of the time." That is a sensible engineering rule, and it has a consequence: a variable that is important but poorly recorded cannot enter the model. The adjustment is therefore built from what is well recorded, which is not the same as what matters.
- The models are reliability-adjusted: "if a given center has fewer patients, then the model shrinks the estimate back towards the group average outcome." This is correct statistics and it is essential to avoid ranking centres on noise. But it means a small centre's benchmarked result is partly a statement about the group, not about that centre — so a small programme cannot read its own dot as a measurement of itself, in either direction.
- Generalised estimating equations account for clustering within hospitals. Patients within a centre resemble each other; the model says so. The practical reading: between-centre differences are estimated with less precision than the number of patients suggests.
- "No fewer than 10 centers are presented as comparators." The peer group is a construct, and its composition determines where your dot sits.
None of this is a criticism of the platform, which does all four of these things because they are the right things to do. It is a warning against the use the platform is most often put to at a local level: taking the benchmark as a verdict on the programme rather than as one noisy estimate among several. The Red Book's own quality chapter frames it correctly, as an input to a Plan-Do-Study-Act cycle rather than a score.
Pitfall · Reading "complications per 1000 ECLS hours" as a rate of occurrence
It is a rate of reported occurrence per unit of support time, and three things follow.
- The numerator is a detection rate. Chapter 28 established this for neurological events specifically; the Red Book says so in its own words, that registry central nervous system complications "should be viewed as minimal estimates because the registry data is inherently limited by the voluntary nature of reporting, the methods of evaluation are not systematic, and the context in which data are collected is acute, intensive care based, and short term." The same three sentences apply to every complication category, not only the neurological ones.
- The denominator rewards long runs. A complication that happens once per run — cannulation injury, a first circuit change — has a per-hour rate that falls as runs get longer, with no change in the underlying risk.
- A centre that looks for a complication finds more of it. Chapter 28's demonstration is the sharpest available: standardised neuromonitoring carried an odds ratio of 2.24 for detection. A programme that improves its surveillance will appear to get worse on the benchmark.
The defensible use is within-centre over time with a stable detection policy. The indefensible use is between-centre comparison of a complication whose detection depends on how hard people look.
30.6 Controversy 1 · One dataset, one team, two opposite answers
The clinical question. Does mechanical left ventricular unloading during extracorporeal cardiopulmonary resuscitation improve survival? And — the question this chapter is actually about — what do you do when the same investigators, using the same data, publish two analyses that point in opposite directions in the same paper?
The paper. Ling RR, Chen Y, Low CJW, Agerstrand C, Jung JS, Lim SL, Lorusso R, Mueller T, Okada Y, Tonna JE, Shekar K, Brodie D, MacLaren G, Ramanathan K. "Left ventricular unloading during extracorporeal cardiopulmonary resuscitation: a target trial emulation of the ELSO registry." Crit Care 2025;29(1):186. DOI 10.1186/s13054-025-05345-3. Read as complete full text — the only Verified source in this chapter.
Position A: the target trial emulation found nothing. The investigators specified a hypothetical randomised trial and then emulated it: adults receiving extracorporeal cardiopulmonary resuscitation for in-hospital or out-of-hospital arrest between 1 January 2020 and 31 December 2023, first run only, confounders selected by a causal directed acyclic graph built after a prior literature review, missing data handled by multiple imputation with chained equations, and adjustment by propensity-score matching. Of 3,215 patients included, 621 pairs were matched.
- No difference in survival time to 90 days: hazard ratio 0.92 (95 percent confidence interval 0.79 to 1.08).
- Concordant across every sensitivity analysis: inverse probability of treatment weighting 0.92 (0.82 to 1.03); complete-case analysis 0.99 (0.86 to 1.15); re-matched including sex 0.92 (0.80 to 1.06); double-adjusted 0.94 (0.81 to 1.09); excluding potential contraindications 0.93 (0.78 to 1.11); adding percutaneous coronary intervention to the matching algorithm 0.94 (0.81 to 1.09).
- Concordant when the eligibility criteria of each published ECPR randomised trial were emulated in turn: ARREST 0.89 (0.75 to 1.07); PRAGUE-OHCA 0.92 (0.77 to 1.10); INCEPTION 1.05 (0.79 to 1.08, as printed — see the note below).
- No difference in survival with favourable functional outcome: odds ratio 1.15 (0.67 to 1.99).
- And harm signals in five directions: renal complications odds ratio 1.55 (1.16 to 2.07); cardiovascular 1.60 (1.14 to 2.26); central nervous system bleeding 1.75 (1.03 to 2.96); arrhythmias 1.56 (1.04 to 2.36); haemolysis 1.85 (1.10 to 3.09). Duration of ECMO was longer by 23.3 hours (8.3 to 38.4); hospital length of stay was unchanged.
Position B: the meta-analysis in the same paper found benefit. As a declared protocol deviation, the same team performed a post-hoc random-effects meta-analysis, pooling five other propensity-score-matched studies contributing six cohorts — and no randomised controlled trials — together with their own adjusted 30-day mortality estimate. Pooled odds ratio for 30-day mortality 0.55 (0.35 to 0.87), which they graded low certainty. By device: intra-aortic balloon pump 0.66 (0.37 to 1.17) and microaxial left ventricular assist device 0.51 (0.25 to 1.04), both graded very low certainty.
Where the evidence actually sits, and why this is a data-interpretation problem rather than a cardiology one. The two results are not a contradiction in the data. They are a difference in what was being estimated. The emulation asked what would have happened had these patients been randomised, with confounders selected in advance from a causal model. The meta-analysis pooled effect estimates from studies whose confounder sets were chosen differently — the authors say so explicitly, noting that previous studies estimated propensity scores from a different set of variables, and that three of the pooled studies came from Japan, one from Germany and one multinational European, in centres likely to have well-defined protocols for initiating unloading in a highly specific subset of patients.
The decisive mechanism is named by the authors themselves: immortal time bias. A patient must survive long enough to receive an unloading device. Any analysis that does not model the timing of the exposure will therefore attribute some of that survival to the device. The emulation's design is more resistant to this than the pooled observational estimates are, which is the strongest reason to prefer the null result over the pooled one.
What this book concludes. Where two analyses of the same data disagree, ask which one modelled time. The target trial emulation is the better estimate here, and the honest summary is that mechanical left ventricular unloading during ECPR has not been shown to improve survival and is associated with more complications. The authors' own conclusion is the same, and they call for randomised data. Chapter 15 owns left ventricular unloading and Chapter 22 owns post-ECPR management; both are owed this finding as an addendum, and Chapter 15's existing position should be re-read against it.
What would resolve it. A randomised trial of unloading during ECPR. The authors note why none exists — enrolment difficulty, and the ethical and logistical problem of withholding what is believed to be a lifesaving therapy — and that trials of unloading during venoarterial ECMO for cardiogenic shock are under way and may define the criteria first.
Certainty: moderate that unloading during ECPR does not improve 90-day survival; moderate for the complication signals; very low for the pooled 30-day mortality benefit, which is the authors' own grading.
Danger · Three things in this paper that a careful reader must notice, and one of them is a printing error
This is not a criticism of an unusually well-conducted study. It is the demonstration that even the best paper in a chapter needs to be read arithmetically, which is the whole argument of §30.9.
- A percentage that does not match its own count. The results state that "1,876 patients (74 percent) had an IHCA" in a cohort of 3,215. 1,876 divided by 3,215 is 58.4 percent, not 74 percent. Either the count, the percentage, or an unstated denominator is different from what is printed. This book reproduces both figures as printed and reconciles neither. [VERIFICATION REQUIRED: the correct in-hospital arrest proportion.]
- A confidence interval that is identical to a different analysis. The INCEPTION emulation is reported as hazard ratio 1.05 (95 percent confidence interval 0.79 to 1.08) — an interval identical to that of the primary analysis, whose point estimate was 0.92. An interval that reaches only 0.03 above a point estimate of 1.05 while extending 0.26 below it is implausibly asymmetric for a hazard ratio. The most likely explanation is a transcription error in the published text. Reproduced as printed, not corrected. [VERIFICATION REQUIRED.]
- A subgroup that will be quoted out of context. In patients with a non-shockable presenting rhythm, unloading was associated with a hazard ratio of 0.79 (0.64 to 0.98) — nominally significant — against 1.06 (0.84 to 1.33) in shockable rhythms, with a p value for interaction of 0.067. The authors state in their methods that subgroup results should be viewed as exploratory and that they did not adjust for multiplicity. The correct reading is that the subgroups do not differ; the correct use of the 0.79 is as a hypothesis, not a finding.
The general rule, which §30.9 develops: check that percentages match their counts, that confidence intervals are plausibly shaped around their point estimates, and that a "significant" subgroup has a significant interaction test behind it. Three arithmetic checks, none requiring statistical training, and they are not routinely done.
30.7 Denominators, missingness, and the addendum problem
Danger · The most consequential number in this chapter is a denominator
The ECPR analysis above opens with a sentence that should change how every registry-derived ECPR figure in this book is read:
"Between 1 January, 2020, and 31 December, 2023, 7,911 patients receiving ECPR were reported to the ELSO registry, of which 3,736 patients had a valid ECPR addendum."
That is 47.2 percent. More than half of the extracorporeal cardiopulmonary resuscitation runs reported to the world's largest ECMO registry over four years could not be used to answer an ECPR question, because the optional addendum carrying the ECPR-specific fields was not completed.
And the reason is structural, not negligent. The Red Book's own account of registry governance states that an optional addendum requires agreement to participate from 10 percent of active centres to be created. An addendum is therefore designed to be optional, and optional means partial. The Cardiac, ECPR, SARS-CoV-2 and Trauma addenda all carry this property.
The consequence for reading. Whenever a registry analysis depends on addendum fields — arrest rhythm, no-flow and low-flow times, bystander cardiopulmonary resuscitation, signs of life — the effective study population is the subset of centres that complete the addendum, not the registry. Those centres are unlikely to be a random sample of the registry: completing an optional addendum is itself a marker of programme resourcing and research engagement.
This does not invalidate the analyses. It means that the population they describe is narrower and better-resourced than the registry's headline count implies, and every ECPR figure carried in Part IV of this book should be read with that qualification. It also means an individual programme can close part of the gap for itself simply by completing the addendum.
Kind of missingness | Where it appears | What it does to the answer | What can be done |
Values excluded before analysis | The 2022 ELSO International Report states its method: "Missing values were excluded before calculating descriptive statistics" | Every descriptive percentage has its own unstated denominator, and the denominators differ between variables in the same table. Two percentages in one row are frequently not about the same patients | Read a registry report's descriptive statistics as complete-case proportions, and do not compute derived figures by multiplying two of them together |
Fields the registry does not record at all | The ECPR analysis names several: the degree of cardiac dysfunction, the indication for unloading, the exact timing and duration of unloading, the timing and indication for percutaneous coronary intervention, and the exclusion criteria used by prior randomised trials (life-limiting comorbidity, treatment limitations, baseline functional status) | An unrecorded confounder cannot be adjusted for by any method. This is the hard ceiling on registry causal inference and no amount of statistical sophistication raises it | Nothing analytical. This is a registry design question, and it is the argument for granular longitudinal datasets that the Red Book's research chapter makes |
Fields recorded by too few centres to use | The same analysis: "very few centres sufficiently recorded neurological outcomes at discharge" | The outcome the patient and family care most about is the one least available for analysis — the same structure Chapter 23 found in the prognostication literature | Record it. A functional outcome at discharge costs nothing to capture, and a published international core outcome set exists to standardise which instruments are used (see the note below) |
Optional addendum not completed | 52.8 percent of ECPR runs, 2020 to 2023 (see above) | Narrows and selects the population without changing the headline count | Complete the addendum your programme's case mix needs |
Informative missingness on the flow sheet | §30.1 — blanks cluster in the worst hours | Biases every within-run trend toward stability | A timed free-text entry when an event resolves, even retrospectively |
Clinical pearl · What a programme can record at essentially no cost that would change what it can answer later
None of these requires software, funding or a research nurse. Each one is a field that recurs as a stated limitation in the literature this book has read.
- A functional outcome at hospital discharge, on every patient, recorded with a named instrument. Its absence is an explicit limitation in the ECPR analysis above, and an international modified Delphi has already done the work of choosing instruments: Hodgson CL, Fulcher B, Mariajoseph FP, Burrell AJC, Pellegrino V, Brodie D, Fan E. "A Core Outcome Set for Research in Patients on Extracorporeal Membrane Oxygenation." Crit Care Med 2021;49(12):e1252-e1254. DOI 10.1097/CCM.0000000000005110. Two rounds, participants from the International ECMO Network and ELSO including consumers, clinicians, researchers and industry partners across 15 countries and five continents, response rates 71 percent and 87 percent, seven measurement tools identified. Abstract only — the seven tools and their measurement timings are not named in the abstract and are not reproduced here. [VERIFICATION REQUIRED: the seven instruments and their timing.]
- The timing of every device and major intervention relative to cannulation, in hours. Unrecorded timing is what makes immortal time bias unanalysable (§30.6).
- The pre-cannulation laboratory panel, which Chapter 29 identified as the highest-value single item in that chapter and which no chapter currently owns as a step.
- The indication, in the clinician's own words, for any decision that a later analyst will treat as an exposure — a circuit change, an unloading device, a switch of anticoagulant. Confounding by indication is unfixable when the indication was never written down.
- The daily phlebotomy volume (Chapter 29), and the number and type of circuit alarms (§30.1) — two numbers that no one collects and that would each support a question the field currently cannot answer.
The test for whether a field is worth adding: name the question it would let you answer in two years. If you cannot, do not add it. A data burden with no question behind it degrades the fields that do have one.
30.8 Controversy 2 · Does high-granularity data improve prediction?
The clinical question. Machine learning models trained on the dense time-series data an ECMO patient generates are now being published. Should a unit trust them more than the simple scores it already uses?
Position A: yes — granularity beats a score, and the numbers say so. A retrospective single-centre study of adult venovenous ECMO patients built models from electronic medical record time-series data across four intervals and compared them against a model built on the Respiratory ECMO Survival Prediction score alone.
- 119 patients, 59.6 percent female, 62 (52.1 percent) dying at discharge. Random Forest, CatBoost, AdaBoost and XGBoost, 80/20 train-test splits, feature importance by Shapley additive explanations.
- Model E, built on the full ECMO run, reached an area under the receiver operating characteristic curve of 0.904 against 0.815 for the RESP-score model.
- The authors' conclusion is that high-granularity models "accurately predicted V-V ECMO weaning" with "superior performance against RESP score-based ML models."
Source: Sivakumar S, Fan L, Cao M, Brown P, Kim BS, Whitman GJ, Cho S, et al. "A High Granularity Machine Learning Model to Predict Successful Weaning During V-V ECMO." Artif Organs 2026. DOI 10.1111/aor.70198. Abstract only.
Position B: no — read the same models by when their information would actually be available, and the conclusion reverses. The four intervals were pre-ECMO, the first 24 hours, the last 24 hours before decannulation, and the full run. Their reported performances were:
Model | Data window | AUROC | When would you have this? |
A | RESP score alone | 0.815 | Before cannulation |
B | Pre-ECMO | 0.715 | Before cannulation |
C | First 24 hours | 0.515 | Day 1 |
D | Last 24 hours before decannulation | 0.881 | Only once decannulation has happened |
E | Full ECMO run | 0.904 | Only once the run has ended |
The two models that performed best are the two that require the outcome to have already occurred before their data exist. A model that needs the last 24 hours before decannulation in order to predict decannulation is not predicting; it is describing. And the two windows in which a clinician would actually want a prediction performed at 0.715 and at 0.515 — the latter being indistinguishable from a coin toss — while a four-variable score available before cannulation reached 0.815 and beat both of them.
Where the evidence actually sits. This is a textbook demonstration of two problems that recur across the clinical prediction literature and are not specific to ECMO.
- Outcome-adjacent information. The closer the data window sits to the outcome, the more the model is reading the consequences of the outcome rather than its antecedents. The gradient across models C, D and E — 0.515, 0.881, 0.904 — is almost a pure measure of proximity to the endpoint.
- Sample size. 119 patients, 62 events, an 80/20 split, and four algorithm families compared. That is roughly 12 events in the test set. Estimates from a test set that size are extremely unstable, and the apparent superiority of 0.904 over 0.815 is well inside what that instability can produce.
A further ambiguity that must be flagged rather than resolved. The study is presented throughout as a model of weaning outcomes, yet the only outcome statistic in the abstract is that 62 of 119 (52.1 percent) experienced mortality at discharge. Whether the modelled endpoint is successful weaning, survival, or a composite is not recoverable from the abstract, and this book does not assume which. [VERIFICATION REQUIRED: the endpoint definition.] That matters because the two endpoints have different relationships to each data window.
And one detail worth noticing because this is a chapter about reading numbers: the published abstract reports the cohort's "median age = 48%". An age carrying a percent sign is a transcription error in a paper whose entire claim rests on careful handling of numbers. It proves nothing about the analysis. It is a reminder that peer review does not check arithmetic.
What this book concludes. A prediction model should be judged by its performance in the window in which the decision is taken, not by its best reported figure. Read that way, this study does not show that high-granularity data improve prediction of weaning; it shows that the first 24 hours of a venovenous run, densely measured, carried almost no information about eventual weaning in this cohort — which is itself a genuinely interesting and clinically humbling result, and the opposite of the headline.
A second study points the same way for a different endpoint and is worth having for calibration. A machine learning model predicting bleeding on ECMO, built from a 71-hospital database, used red cell transfusion as a surrogate marker for bleeding. Of 470 ECMO intensive care patients, 357 entered model development with 47 variables; the best models reached an area under the curve of about 0.703 to 0.705, and Shapley analysis identified circulating blood volume, haemoglobin and weight as the most important predictors. Source: Kamio T, Ikegami M, Mizuno M, Ishii S, Tajima H, Machida Y, Fukaguchi K. "Machine learning-based prediction of bleeding risk in extracorporeal membrane oxygenation patients using transfusion as a surrogate marker." Transfusion 2025;65(6):1051-1060. DOI 10.1111/trf.18261. Abstract only.
The structural problem there is the outcome definition, not the algorithm. Transfusion is a treatment decision, not a physiological event. A model trained to predict transfusion learns the transfusion practice of the contributing hospitals, and its most important features — blood volume, haemoglobin, weight — are precisely the variables a transfusion protocol uses. Chapter 46 owns transfusion strategy and should carry that caution.
What would resolve it. Prospective, externally validated models reporting performance separately for each decision window, with the window stated in the abstract; adequate event counts; and outcomes defined physiologically rather than by the treatment given for them.
Certainty: low for any specific model's performance; moderate-to-high for the structural reading, because the gradient across data windows is visible in the published numbers themselves.
Pitfall · Accepting a model's stated purpose instead of checking its inputs
Four questions dispose of most published clinical prediction models in under a minute, and none needs statistical training.
- "At what moment would I use this, and does its data window end before that moment?" If any input is measured after the decision point, the model cannot inform the decision.
- "How many events were in the test set?" Not patients — events. Under about 100 events, a difference of 0.05 in area under the curve is noise.
- "What exactly is the outcome?" If the outcome is a treatment (transfusion, circuit change, escalation), the model has learned local practice and will not transfer.
- "Was it validated anywhere other than where it was built?" An internally split cohort is not external validation.
A model that survives all four is worth reading. Almost none does.
30.9 How to read a number in an ECMO paper
This section exists because most of what a clinician will ever know about ECMO arrives as numbers in papers built from data of the kind §§30.1 to 30.7 described. Appendix 10, Appendix 11 and Appendix 12 hold this book's evidence tables; these are their reading rules.
Check | What you are looking for | Worked instance from this chapter |
1. Does the percentage match its count? | Divide. It takes five seconds and it fails more often than anyone expects | "1,876 patients (74 percent)" in a cohort of 3,215, where the division gives 58.4 percent (§30.6) |
2. What is the denominator, and is it the same one as the row above? | Registry descriptive statistics are typically complete-case, with a different denominator per variable | "Missing values were excluded before calculating descriptive statistics" (§30.7) |
3. Is the confidence interval plausibly shaped around its point estimate? | Ratios are roughly symmetric on a log scale. A wildly lopsided interval usually means a transcription error | Hazard ratio 1.05 (0.79 to 1.08) (§30.6) |
4. Does a "significant" subgroup have a significant interaction test? | The interaction test is the question. A nominally significant subgroup with a non-significant interaction is a hypothesis | Non-shockable rhythm 0.79 (0.64 to 0.98) with p for interaction 0.067, declared exploratory by the authors (§30.6) |
5. Could the exposure have required survival to receive it? | Immortal time bias. Devices, procedures, transfers and second runs are all vulnerable | Left ventricular unloading during ECPR — named as a limitation by the authors themselves (§30.6) |
6. Is the outcome a physiological event or a treatment decision? | A treatment outcome makes the model or the analysis a description of local practice | Transfusion as a surrogate for bleeding (§30.8) |
7. Is this a definition or a measurement? | Thresholds adopted for registry comparability are frequently mistaken for physiological boundaries | The 50 mg/dL haemolysis threshold — a definition of an event, with the strongest adult outcome signal sitting below it (Chapter 29) |
8. Is this rate an incidence or a detection rate? | Anything whose ascertainment depends on how hard people looked | Neurological complications: 4 to 13 percent in routine care against 68 percent at autopsy (Chapter 28) |
9. Which units? | The same quantity is reported in mg/dL and mg/L, in seconds and in ratios, in IU/mL and units/mL | 50 mg/dL equals 500 mg/L; the ISCCM fibrinogen row printed in mg/L where mg/dL is meant (Chapter 29) |
10. Is the population the population in front of you? | Neonatal, paediatric and adult ECMO differ; short perioperative runs differ from long ones; addendum-completing centres differ from registry centres | Centrifugal pumps cause more haemolysis in neonates than in adults (Chapter 29); short perioperative support produced no clinically significant multimer loss (Chapter 29) |
Clinical pearl · The single most useful habit is to read the methods section's data source paragraph before the results
It takes about a minute and it determines the meaning of every number that follows. Four things to find.
- Where did the data come from? A registry, a single centre's electronic record, a prospective protocol, or a trial.
- What was the time resolution? Once per run, daily, hourly, continuous. This is the ceiling on every temporal claim the paper can make, and it is the difference between the ECPR analysis above — which could not model the timing of unloading because the registry does not record it — and the prospective PROTECMO study, which recorded anticoagulation daily across 41 centres, 652 patients and 8,471 ECMO days, and could therefore show that anticoagulation is a dynamic process with frequent stopping and restarting, and that the median activated partial thromboplastin time fell by 5.3 seconds (95 percent confidence interval minus 7.4 to minus 3.2) after a patient's first bleeding event. Source: Martucci G, Giani M, Schmidt M, et al. "Anticoagulation and Bleeding during Veno-Venous Extracorporeal Membrane Oxygenation: Insights from the PROTECMO Study." Am J Respir Crit Care Med 2024;209(4):417-426. DOI 10.1164/rccm.202305-0896OC. Abstract only. Chapter 42 owns the anticoagulation content; what belongs here is that a registry field recording which anticoagulant was used cannot represent a practice that changes daily, and only daily data showed that it does.
- How was missing data handled? Excluded, imputed, or not mentioned. All three are common and they give different answers, as §30.6 showed within a single paper.
- Who decided which variables to adjust for, and how? A causal diagram built in advance is a different thing from a list of variables that were available.
30.10 Building a data system a unit can answer questions with
The Red Book's quality improvement chapter frames all of this correctly: data exist to drive a Plan-Do-Study-Act cycle — state the objective, predict what will happen and why, plan who and what and when and what data need to be collected, test on a small scale, document problems and unexpected observations, analyse, compare the data against the prediction, refine, and repeat. It calls this "the scientific method applied to action-oriented learning" and names error proofing as the change most worth making: redesigning the system so that errors become less likely, including "by writing it down as a policy, creating a checklist, or by making it inherent in the product or process."
Clinical pearl · A minimum viable ECMO data system, in four layers
This is this book's synthesis of the primary-source structures above. It is offered to be adapted and audited, not as a guideline. Nothing in it requires a research budget.
Layer 1 — the flow sheet, with three additions to the standard.
The standard content is in §30.2. The additions that this chapter's evidence supports:
- Every set value paired with its measured partner on the same line — speed with flow, sweep with post-membrane carbon dioxide, set with delivered temperature (§30.2).
- A timed free-text line for transient events, written when the event resolves, even retrospectively (§30.1).
- The sampling site recorded with every gas (Chapter 27), because a gas without its site is not a result.
Layer 2 — the daily circuit entry, separate from the critical care note.
Its minimum content, from ECMO: A Practical Guide to Management plus Part V: weaning attempts and adjustments, complications, the current anticoagulation strategy stated as a strategy rather than a dose, a settings snapshot, and — the addition this chapter argues for — the trend of three numbers against three days ago rather than against yesterday. Which three is a unit decision; transmembrane pressure, plasma free haemoglobin and the pulse pressure are defensible choices for most programmes.
Layer 3 — the registry entry, completed including the addendum your case mix requires.
The ELSO Registry Definitions Document and the data entry certification exam exist precisely so that this layer is comparable between centres (§30.5). The single highest-yield act available to an individual programme is to complete the relevant optional addendum, because that is where the 52.8 percent loss in §30.7 occurs.
Layer 4 — the local dataset that answers your own questions.
This is the layer the registry cannot provide, and the Red Book's research chapter says why: "Existing data sources do not, however, include all the necessary data elements (eg, granular, longitudinal data) or use the standardized definitions needed for research." A local dataset should hold what §30.7's pearl lists — functional outcome at discharge, the timing of every device and major intervention relative to cannulation, the pre-cannulation panel, the written indication for anything a later analyst would treat as an exposure — and nothing whose question you cannot state.
Pitfall · Measuring the programme by the two numbers that are easiest to obtain
ECMO: A Practical Guide to Management is direct about this. Survival to decannulation and survival to discharge are "straightforward" and easily benchmarked against ELSO, "However, these do not alone capture the measure of success for your program." Its own suggested list is broader and includes number of consults received internally and externally; runs per year; complication rates; blood product usage; length of stay; patient and family satisfaction; patient follow-up and long-term outcomes; disposable costs and identified savings; team member retention and satisfaction; team participation in outside activities; and publications.
The same source adds a point that is uncomfortable and true: what the hospital holds the programme accountable for is not always what the programme measures, and it lists financial impact as the first thing to understand. An ECMO programme that cannot describe itself in the terms its institution uses is a programme whose data will not protect it.
Two additions from this chapter. First, an outcome measured only to discharge is an outcome measured at the point the field's own literature says is least informative — Chapter 23 established this for neurological outcome and the ECPR analysis in §30.6 found that very few centres recorded even a discharge functional outcome. Second, any complication rate a programme reports about itself is a detection rate about itself (§30.5), so improving surveillance will make the numbers worse before it makes the patients better. Say so in advance, in writing, before the surveillance changes.
30.11 A worked case: the data problem behind a clinical disagreement
Worked example · Three clinicians, one patient, one absent column
The patient. Day 6 of venoarterial ECMO for fulminant myocarditis. Circuit flow 3.6 L/min at 3,400 revolutions per minute. Noradrenaline 0.08 micrograms/kg/min. The morning round has three positions and cannot resolve them.
The three positions.
- The intensivist: "The lactate is 2.1 and the mixed venous saturation is 71 percent. This patient is well supported and we should be talking about weaning."
- The perfusionist: "The transmembrane pressure has gone from 18 to 31 mmHg. I want to plan a circuit change."
- The cardiologist: "The echo yesterday showed the aortic valve opening every beat and the pulse pressure is 22 mmHg. The heart is recovering and we should reduce flow."
What is actually being disagreed about. All three statements are true. They are three different questions, answered by three instruments that measure different things, and the round is treating them as competing opinions about one question. §30.4's first question disposes of most of it: are these instruments measuring the same thing? No. One is measuring the adequacy of systemic delivery, one the state of the oxygenator, and one the state of the native heart. None of the three contradicts either of the others, and all three findings should be acted on, in a sequence.
What the round is missing, and it is not an instrument. It is a column. Every number quoted was quoted as a value. Put yesterday, day 3 and day 1 beside them and the argument resolves itself.
Value | Today | Day 5 | Day 3 | Day 1 | What the column shows |
Circuit flow (L/min) | 3.6 | 3.6 | 4.1 | 4.3 | Falling at unchanged speed since day 3 |
Pump speed (rpm) | 3,400 | 3,400 | 3,400 | 3,400 | Unchanged — so the flow is a circuit finding, not a decision |
Transmembrane pressure (mmHg) | 31 | 26 | 21 | 18 | Rising steadily, not abruptly |
Pulse pressure (mmHg) | 22 | 18 | 9 | 6 | Rising — native ejection returning |
Mixed venous saturation (percent) | 71 | 70 | 68 | 61 | Improving |
Lactate (mmol/L) | 2.1 | 2.4 | 3.8 | 6.9 | Improving |
Plasma free haemoglobin (mg/dL) | 34 | 22 | 15 | 12 | Rising inside the "normal" band |
Working it through.
- The intensivist is right that delivery is adequate, and the column shows it has been improving for five days. Weaning is a legitimate agenda item.
- The perfusionist is right that the oxygenator is changing, and the column tells you something the single value could not: the rise is gradual and monotonic rather than abrupt. That distinction is Chapter 32's, and it separates progressive fibrin deposition from an acute event. It changes the urgency, not the diagnosis.
- The cardiologist is right that the heart is recovering, and the pulse pressure column is the cleanest evidence of it — 6 to 9 to 18 to 22 mmHg. Chapter 18 owns what to do about it.
- The finding nobody raised is the one the column makes visible. Plasma free haemoglobin has nearly tripled, from 12 to 34 mg/dL, and remains "normal" by the ELSO definition throughout. Chapter 29 established that this is the band in which the only adult cohort that looked found an adjusted hazard ratio of 7.6 for non-haemorrhagic stroke, graded with concentration. It is also concordant in direction with the rising transmembrane pressure, which supplies a mechanism the free haemoglobin alone would not have.
- So the three arguments collapse into one coherent account. The circuit is progressively failing — falling flow at fixed speed, rising transmembrane pressure, rising free haemoglobin, all moving together since day 3. The patient is recovering — rising pulse pressure, improving saturation and lactate. These are not in tension. They are the two halves of the same decision, which is whether this patient reaches weaning before this oxygenator reaches replacement.
What should happen. Assess the oxygenator formally (Chapter 32) and plan, not perform, a circuit change. Proceed with a weaning assessment (Chapter 18, Chapter 26) — because if the patient can be decannulated, the circuit question disappears. Repeat the free haemoglobin from a different site before acting on it, and look for the concordant markers (Chapter 29). Raise neurological surveillance (Chapter 28). And record all four decisions with their indications, in the daily circuit entry, because in six months somebody will want to know why this circuit was changed and the chart will otherwise say only that it was.
The teaching point. No new measurement was needed to resolve a three-way disagreement between senior clinicians. What was needed was four columns of numbers the unit already had and was not displaying, and one question — "are we measuring the same thing?" — asked before the argument started.
This case is constructed to illustrate the reasoning in §§30.1 to 30.4. The values are illustrative and are not drawn from a published report.
30.12 What this chapter could not close
Open questions, recorded rather than concealed
- No study was retrieved that quantifies what is lost between continuous console data and the hourly flow sheet on ECMO. The arithmetic in §30.1 is a calculation from sampling frequencies, not a measured finding, and it is labelled as such. The empirical question — how many clinically meaningful circuit events occur entirely between two flow sheet entries — appears not to have been asked. It is answerable with a single centre's console logs and would be among the more useful studies in this chapter's subject.
- No evidence at all was retrieved on ECMO alarm burden, despite the alarm history being part of the documentation standard. A targeted search returned no ECMO-specific data on alarm frequency, alarm validity or alarm fatigue. Recorded as a finding.
- The external validation programme for ELSO Registry data was described as a pilot in a small number of centres. No published results of it were retrieved, so the accuracy of registry data against source records is not known to this book — which is a different and larger statement than any of the specific limitations in §30.7.
- The core outcome set for ECMO research exists, but its seven measurement tools and their timings were not retrievable from the abstract and are not reproduced anywhere in this book. [VERIFICATION REQUIRED.]
- Two internal inconsistencies in the best paper in this chapter are recorded unresolved (§30.6): a percentage that does not match its count, and a confidence interval identical to that of a different analysis. Neither is corrected here.
- INDEX (the International Research Database for Extracorporeal Support) and EXCEL (the Australia and New Zealand national ECMO registry) are named by the Red Book as the vehicles for granular longitudinal data, and nothing was retrieved about either. They are the most likely source of an answer to item 1 and are a retrieval target.
- No ECMO-specific prediction model retrieved for this chapter had external validation. Both models discussed in §30.8 were internally split only.
Corrections and additions owed to other chapters, arising from this chapter's audit
- Chapter 15 (LV Distension and LV Unloading) and Chapter 22 (Post-ECPR Management) are owed the ECPR unloading target trial emulation in §30.6 as a dated addendum: no survival benefit (hazard ratio 0.92, 0.79 to 1.08) and five complication signals, from 621 matched pairs. Chapter 15 already carries a section on unloading in ECPR (§15.9) and a section distinguishing observational from randomised evidence; both should be re-read against this before Part XII. BOOK_STATUS already records that Thiele's unloading-device arm was deliberately not quoted in Chapter 15; this is the ECPR-specific evidence that was missing then.
- Chapters 19, 20, 21 and 22 are owed the addendum denominator in §30.7: only 47.2 percent of ECPR runs reported to the registry between 2020 and 2023 had a valid ECPR addendum. Every registry-derived ECPR figure in Part IV should carry that qualification.
- Chapter 46 (Transfusion Strategy) is owed §30.8's caution that a model predicting transfusion has learned transfusion practice.
- Chapters 87, 88 and 92 are owed the four-layer data system in §30.10 and the paired set-and-measured-value rule in §30.2 as design requirements for the checklist, round template and handover template.
- Appendix 10, Appendix 11 and Appendix 12 are owed the ten reading checks in §30.9 as their preface.
- Chapter 1 is owed a cross-reference to §30.5, which now closes the registry-structure gap that the Chapter 28 audit recorded against it.
Carried forward from earlier chapters and still open: a home for extracorporeal carbon dioxide removal (Chapter 27); a home for neurological complications as complications (Chapter 28); the pre-cannulation laboratory panel, still unowned (Chapter 29); the four numeric anticoagulation disagreements awaiting Chapter 42 (Chapter 29); the alpha-stat versus pH-stat gap, for which no ECMO-specific evidence exists; and the Hunt near-infrared spectroscopy cohort.
30.13 Evidence and certainty summary
Claim | Best evidence | Certainty |
Flow sheet documentation on ECMO is performed by the bedside specialist every 1 to 2 hours and is a standard of care, with a defined content including circuit checks and alarm history | ECMO: A Practical Guide to Management, programme chapter, stated as a standard | High as a description of practice; none as evidence that the interval is correct |
ECMO documentation has three distinct levels with different purposes: programme, flow sheet, and provider | Same source, set out explicitly | High — an organisational framework, not a clinical claim |
Each level of the measurement pyramid loses roughly an order of magnitude, biased toward the stable and the reported | This book's calculation from the sampling frequencies in the sources above, plus the registry's own complete-case method | Low as a quantitative claim — the order-of-magnitude figures are arithmetic from sampling intervals, not a measured finding. High as a directional claim, because every mechanism in the chain removes the transient and the unreported |
The ELSO Registry maintains published data definitions, a mandatory data entry certification exam, internal hard and soft limits at entry, and a pilot external validation programme auditing 64 elements in 10 percent of each centre's prior-year patients | ELSO Red Book 6th edition, Chapter 60 — the registry's own account | High that these mechanisms exist; unknown what they achieve, because no results of the external validation pilot were retrieved |
Registry complication figures are minimal estimates, limited by voluntary reporting, non-systematic evaluation methods, and a short acute-care data collection context | Red Book, stated in its own words about central nervous system complications; generalised here to all categories | High for the statement as made; moderate for the generalisation, which this book makes and the source does not |
Registry risk-adjusted benchmarking shrinks small centres' estimates toward the group average and is built only from variables missing under 10 percent of the time | Red Book, Chapter 60 | High — the platform's own documented method |
Fewer than half of ECPR runs reported to the ELSO Registry between 2020 and 2023 had a valid ECPR addendum (3,736 of 7,911; 47.2 percent) | Ling and colleagues, Crit Care 2025, results section. Full text read | High — a directly reported count, in the paper's own flow description |
Mechanical left ventricular unloading during ECPR is not associated with improved 90-day survival | Ling and colleagues: 3,215 patients, 621 matched pairs, hazard ratio 0.92 (0.79 to 1.08), concordant across seven sensitivity analyses and three emulated trial-eligibility criteria | Moderate — the strongest available evidence, but observational; the authors call for randomised data |
Unloading during ECPR is associated with more renal, cardiovascular, haemorrhagic and haemolytic complications | Same study: renal odds ratio 1.55, cardiovascular 1.60, CNS bleeding 1.75, arrhythmias 1.56, haemolysis 1.85 | Moderate — consistent with prior reports and with a plausible mechanism (additional device, additional shear) |
A pooled meta-analysis of the same literature, including the above cohort, suggests lower 30-day mortality with unloading | Same paper, post-hoc: pooled odds ratio 0.55 (0.35 to 0.87) from six cohorts, no randomised trials | Very low — the authors' own grading. Contradicted by their primary analysis, and immortal time bias runs in the direction of the pooled result |
Immortal time bias inflates the apparent benefit of any ECMO exposure that a patient must survive to receive | Named as a limitation by the investigators of the study above, with reference to prior ECMO literature | High as a methodological principle; unquantified in this specific case |
High-granularity machine learning models for VV ECMO weaning perform best in data windows that require the outcome to have occurred | Sivakumar and colleagues, Artif Organs 2026: 119 patients; pre-ECMO 0.715, first 24 hours 0.515, last 24 hours before decannulation 0.881, full run 0.904, against RESP-based 0.815. Abstract only | Low for any individual figure — roughly 12 events in the test set. Moderate to high for the structural reading, because the gradient is visible in the published numbers |
A machine learning model predicting transfusion has learned transfusion practice, not bleeding | Kamio and colleagues, Transfusion 2025: 357 patients from 71 hospitals, 47 variables, area under the curve about 0.70, top features circulating blood volume, haemoglobin and weight. Abstract only | Moderate — the surrogate is stated by the authors and the feature ranking is consistent with it |
Daily prospective data capture reveals ECMO anticoagulation to be a dynamic process that a per-run registry field cannot represent | PROTECMO: 41 centres, 652 patients, 8,471 ECMO days; aPTT monitoring on 86 percent of days; median aPTT fell 5.3 seconds after a first bleeding event. Abstract only | Moderate to high — a prospective multicentre design purpose-built for the question |
An international core outcome set for ECMO research exists, identifying seven measurement tools | Hodgson and colleagues, Crit Care Med 2021: two-round modified Delphi, 15 countries, response rates 71 and 87 percent. Abstract only; the seven tools are not named in the abstract | Moderate that it exists and was rigorously derived; none for its content, which this book does not have |
Programme success cannot be measured by survival to decannulation and discharge alone | ECMO: A Practical Guide to Management, with an eleven-item suggested measure set | Moderate — expert practice, and an organisational rather than clinical claim |
Quality improvement on ECMO should run as a Plan-Do-Study-Act cycle with prespecified data collection | ELSO Red Book, Chapter 55 | Moderate — established general methodology, applied to ECLS by the source without ECLS-specific outcome evidence |
30.14 Key points
- An ECMO patient generates more measurements per hour than almost any other patient in the hospital, and almost none of it is kept. The console measures continuously, the flow sheet records every 1 to 2 hours, the registry records a summary per run, and a published analysis uses whatever was not missing. Each step loses about an order of magnitude.
- The losses are not random. At every level, what disappears is the brief, the unstable, the uncharted, the unsurvived and the unreported. The bias has a direction, and the direction is toward reassurance.
- A blank on a flow sheet is not a normal value. Gaps cluster in the worst hours of a run, because that is when nobody was free to type.
- Write down transient events with a time, even after they have resolved. The alarm history is part of the documentation standard, and the circuit catastrophe at 03:00 almost always announced itself earlier as something nobody recorded.
- Pair every set value with its measured partner. Speed with flow, sweep with post-membrane carbon dioxide, set with delivered temperature. A falling flow at unchanged speed is a circuit finding; the same flow at reduced speed is a decision, and a chart carrying only one of the pair cannot tell them apart.
- Keep the daily circuit entry separate from the critical care note. It is the only way the stable day gets a trend.
- The rule Part V reached from five directions: the reference interval describes a person who was never here; the comparison that carries information is this patient's own earlier value. Its operational form is a column, not a sentence.
- On a parallel circulation, two instruments disagreeing is the normal state, and the pattern of disagreement carries the diagnosis. Ask first whether they are measuring the same blood — on venoarterial support the answer is frequently no, and then there is no contradiction to resolve.
- A monitor that always agrees with another monitor is adding no diagnostic information, and two concordant readings are not two pieces of evidence. A redundant channel still guards against the other one failing — but agreement is not confirmation.
- Registry complication rates are detection rates, and "per 1000 ECMO hours" rewards long runs. The defensible use is within-centre over time with a stable detection policy; between-centre comparison of an ascertainment-dependent complication is not defensible.
- A benchmarked centre-level result is partly a statement about the peer group. Small centres' estimates are deliberately shrunk toward the average, the adjustment model can only use well-recorded variables, and the comparator group is a construct.
- Fewer than half of the ECPR runs reported to the world's largest ECMO registry over four years carried the addendum needed to answer an ECPR question. Optional addenda are optional by design, and the centres that complete them are not a random sample.
- When two analyses of the same data disagree, ask which one modelled time. In the ECPR unloading paper, the target trial emulation found nothing and the pooled observational meta-analysis found benefit; immortal time bias runs in the direction of the second.
- Mechanical left ventricular unloading during ECPR has not been shown to improve survival and is associated with more renal, cardiovascular, haemorrhagic and haemolytic complications. Chapters 15 and 22 are owed this.
- Judge a prediction model by its performance in the window in which the decision is taken. A model needing the last 24 hours before decannulation to predict decannulation is describing, not predicting — and in the same study the first 24 hours of support predicted weaning at an area under the curve of 0.515, while a four-variable score available before cannulation reached 0.815.
- If a model's outcome is a treatment decision, the model has learned local practice and will not transfer.
- Four arithmetic checks, none needing statistics, catch most errors in an ECMO paper: does the percentage match its count; is the confidence interval plausibly shaped; does the "significant" subgroup have a significant interaction test; could the exposure have required survival to receive it. All four found something in the best paper in this chapter.
- Before adding a field to your data collection, name the question it will answer in two years. If you cannot, do not add it — a data burden with no question behind it degrades the fields that have one.
- Record the indication, in words, for anything a future analyst would treat as an exposure. Confounding by indication is unfixable when the indication was never written down.
- Improving surveillance will make your complication numbers worse before it makes your patients better. Say so, in writing, before you change the surveillance.
Chapter answer: Part V's five instruments do not compete; they answer different questions, and on a parallel circulation their disagreement is the normal state and the diagnosis. What resolves a disagreement is almost never a new measurement — it is the patient's own earlier value, displayed in a column beside today's, and one question asked before the argument starts: are we measuring the same thing? What survives afterwards is much less than what was measured, and what is lost is systematically the brief, the unstable and the unreported. So a unit that wants to know anything about itself in two years must decide today which few numbers it will keep, and write down why it did what it did — because the chart will otherwise record only that it was done.