The question this chapter answers
Bleeding is the complication every ECMO team expects, and the word covers two different problems. One is a hole — a cannula site, a surgical wound, a drain — where the treatment is local and the patient's outcome barely changes. The other is a failure of the whole haemostatic system — bleeding into the brain, the lungs or the gut — and it carries some of the highest mortality associated with any complication on ECMO.
So the bedside question is not is the patient bleeding? but where, how fast, which of the two kinds it is, and which of five mechanisms is driving it.
This chapter describes bleeding as a complication: how often it happens, which bleeds matter, why ECMO patients bleed, and what to do at each site before the systemic treatment begins. Chapter 45 owns the systemic treatment.
Evidence search date: 24 September 2026. Two sources are read as complete full text and are the Verified anchors: the ELSO Registry analysis of bleeding and thrombotic events in adult venoarterial support, and its companion in adult venovenous support, from the same research group. Primary sources are the ELSO Red Book 6th edition (adult respiratory management, adverse effects, the registry complication definitions, and procedures during ECLS), ECMO in the Adult Patient (the coagulation chapter and its bleeding table), ECMO: A Practical Guide to Management (cannulation-site bleeding and haematological complications), ECPR and Resuscitative ECMO (the bleeding patient), Comprehensive Healthcare Simulation: ECMO Simulation (registry complication tables and a haemorrhage scenario), and the ISCCM Manual of RRT and ECMO in ICU.
What this chapter covers — and what it does not
This chapter owns | This chapter does not own |
Bleeding as a complication — definitions, how often it happens, which sites, which bleeds are associated with death | Systemic bleeding management. Chapter 45 owns it: holding or reducing anticoagulation, laboratory targets, antifibrinolytics, factor concentrates, protamine and the decision to come off support early |
The five mechanisms of bleeding on ECMO, each as a question to ask at the bedside | The coagulation physiology behind them. Chapter 39 owns it, including acquired von Willebrand syndrome in depth |
Site-by-site local measures — what is done at the cannula, the wound, the airway and the tracheostomy before systemic treatment | Transfusion thresholds and product strategy. Chapter 46 owns them |
The diagnostic approach to a falling haemoglobin, including occult bleeding | Intracranial haemorrhage. Chapter 38a owns neurological complications; this chapter records it only as part of the bleeding hierarchy |
What the registries say about who bleeds, and whether bleeding can be predicted | Device-related coagulopathy and the exchange that reverses it. Chapter 31 §31.5 owns it; this chapter places it in the differential |
Anticoagulation monitoring (Chapter 42), heparin (Chapter 40), tamponade (Chapter 26), and cannula-site problems that are mechanical (Chapter 34) |
36.1 Two kinds of bleeding
The structural claim of this chapter · A bleeding site and a bleeding patient are different problems
The ELSO Registry analyses that anchor this chapter divide bleeding the same way in both configurations.
- Surgical bleeding — cannulation site, surgical site, and tamponade. A defined hole in a defined place.
- Medical bleeding — intracranial, pulmonary and gastrointestinal haemorrhage. Bleeding from a vascular bed that was intact, because the haemostatic system has failed.
The division is not a convenience. It is where the mortality signal lives. In the venovenous analysis, "medical bleeding events were associated with high mortality, whereas none of the surgical bleeding events was significantly associated with mortality." In the venoarterial analysis, cannulation-site bleeding had no significant association with in-hospital mortality (OR 1.05, 95% CI 0.91 to 1.22), while intracranial haemorrhage had an odds ratio of 7.71, pulmonary haemorrhage 3.08 and gastrointestinal haemorrhage 1.95 (§36.4).
Three consequences, and the chapter is built on them.
- The commonest bleed is the least informative. Cannula-site bleeding is the most frequent bleeding event in both registries and is not associated with death. It needs local treatment and a check that it is not something else (§36.6).
- The rarer bleeds carry the prognosis. A patient who starts bleeding into the lungs or the gut is telling the team that the whole system is failing, and the question becomes why (§36.5, §36.7).
- A falling haemoglobin is not located until it is located. Visible bleeding draws attention; the bleed that matters may be in the retroperitoneum, the pleura or the gut (§36.6).
Certainty: moderate for the association pattern, which two large registry analyses report concordantly; the association is not proof that medical bleeding causes death (§36.9).
36.2 How often ECMO patients bleed
Evidence · Two registry analyses, one method, two configurations
Venoarterial. Chung M, Cabezas FR, Nunez JI, Kennedy KF, Rick K, Rycus P, Mehra MR, Garan AR, Kociol RD, Grandin EW. "Hemocompatibility-Related Adverse Events and Survival on Venoarterial Extracorporeal Life Support: An ELSO Registry Analysis." JACC Heart Fail 2020;8(11):892–902. DOI 10.1016/j.jchf.2020.09.004, PMC8427553. Full text read.
Venovenous. Nunez JI, Gosling AF, O'Gara B, Kennedy KF, Rycus P, Abrams D, Brodie D, Shaefi S, Garan AR, Grandin EW. "Bleeding and thrombotic events in adults supported with venovenous extracorporeal membrane oxygenation: an ELSO registry analysis." Intensive Care Med 2022;48(2):213–224. DOI 10.1007/s00134-021-06593-x, PMC9178906. Full text read.
Both used the ELSO Registry for adults supported 2010 to 2017, excluded patients with pre-existing coagulopathy, bleeding or stroke and those with multiple runs, and used multivariable logistic regression for in-hospital mortality.
Venoarterial (11,984 patients) | Venovenous (7,579 patients) | |
Any bleeding or thrombotic event | 44.1 percent of patients | 40.2 percent of patients |
Bleeding as a share of all events | 62.1 percent — bleeding dominates | About 45 percent — thrombotic events were 54.9 percent |
Patients with any bleeding | 32.8 percent | About 23 percent — the book's arithmetic from the published proportions (37 percent bleeding only plus 21.2 percent both, of the 40.2 percent with any event) |
Commonest bleeding events, as share of all events | Cannulation site 23.5 percent; surgical site 20.8 percent | Cannulation site 15.5 percent; surgical site 9.6 percent |
Medical bleeding, as share of all events | 12.6 percent | 18.7 percent |
Haemorrhagic stroke | 2 percent of patients | 4.5 percent of events |
In-hospital mortality, whole cohort | 58.6 percent | 34.9 percent |
Trend 2010 to 2017 | Overall events fell; medical bleeding barely fell (p for trend 0.08) | Bleeding and thrombosis fell; medical bleeding fell more modestly |
The reading this chapter takes. Venoarterial support bleeds; venovenous support clots. The venovenous authors suggest why: longer runs, no arterial cannula, fewer surgical procedures, and a general tendency toward less aggressive anticoagulation. And in both, the bleeding that improved over the decade was surgical; the medical bleeding that carries the mortality did not.
Certainty: moderate — very large, multicentre, but self-reported, with a registry definition that misses minor bleeding (§36.3), and no anticoagulation or laboratory data.
Why published bleeding rates range from 10 to 60 percent
The sources give very different figures for the same complication.
Source | Figure |
ISCCM manual | Bleeding in 10 to 30 percent |
ECMO in the Adult Patient, respiratory failure | Gastrointestinal bleeding 5 percent, pulmonary haemorrhage 10 percent, intracranial haemorrhage under 5 percent |
ELSO Registry, venovenous (Nunez) | About 23 percent of patients |
ELSO Registry, venoarterial (Chung) | 32.8 percent of patients; the authors cite published rates of 25 to 60 percent |
Single centre, adult ECLS (Mazzeffi, Ann Thorac Surg 2016, abstract read) | Serious bleeding in 56.1 percent (74 of 132); 10 events per 100 days |
Much of the spread is definition, and the rest is case mix. A registry that counts a bleed only when it crosses a transfusion threshold (§36.3) reports fewer than a single centre that counts every serious event. A bleeding rate means nothing without the definition beside it.
36.3 What counts as a bleed
Pitfall · The registry only sees bleeding that crossed a threshold
The Red Book reproduces the ELSO Registry complication definitions. Bleeding is counted when it needs something.
- Gastrointestinal haemorrhage: upper or lower GI haemorrhage requiring red cell transfusion (more than 20 mL/kg in 24 hours, or more than 3 units in 24 hours in neonates and paediatrics), and/or endoscopic intervention, and/or a haemostatic agent.
- Peripheral, mediastinal cannulation-site or surgical-site bleeding: site bleeding requiring red cell transfusion at the same thresholds, and/or surgical intervention (including intravascular haemostatic agents).
The venoarterial authors state the consequence themselves. More than half their cohort had no reported bleeding or thrombotic complication, "which is somewhat lower than expected based on general clinical experience," and may reflect under-reporting or "the registry thresholds to define significant HRAEs (e.g., >3 U PRBCs/24 h for a bleeding event), resulting in minor events not being recorded."
What this means when reading the literature. The registry is a record of major bleeding. The steady ooze from a cannula site, the blood-stained secretions after suctioning, the haematoma at an old line site — none of these reaches the registry unless it crosses a threshold. A unit comparing its own bleeding rate with registry figures must use the same definition, or it will find it has more bleeding than the registry and conclude, wrongly, that it has a problem.
And the ELSO definitions do not include every bleed. The venovenous authors note the registry does not collect upper respiratory bleeding — so epistaxis and oropharyngeal bleeding are invisible to it.
Clinical pearl · "Minor" bleeding adds up
ECMO in the Adult Patient: minor bleeding — at the cannula, line insertions or chest drains — "often does not require any intervention other than local pressure. Persistent 'minor' bleeding (e.g. from a neck line suture) may result in significant volume loss if left unchecked for many hours and should not be left untreated."
Chart it as a volume, not an adjective. Dressings weighed or changed and counted, drain output measured, units transfused recorded against time. A bleed described only as "minor" on every shift for three days may have cost the patient several units — and the transfusion count, not the adjective, is what the registry and §36.9 suggest may matter.
36.4 Which bleeds matter
Evidence · The hierarchy, from the two registry analyses
Adjusted odds of in-hospital death associated with each event, and — for venovenous support — the crude mortality among patients who had it.
Event | Venoarterial: adjusted OR (95% CI) | Venovenous: adjusted OR (95% CI) | Venovenous: mortality with the event |
Intracranial haemorrhage | 7.71 (4.79 to 12.41) | 5.71 (4.02 to 8.09) | 73.2 percent |
Pulmonary haemorrhage | 3.08 (2.19 to 4.33) | 2.02 (1.54 to 2.67) | 53.6 percent |
Gastrointestinal haemorrhage | 1.95 (1.51 to 2.52) | 1.54 (1.20 to 1.98) | 48.6 percent |
Tamponade | 1.40 (1.09 to 1.79) | Not significant | 46.6 percent |
Surgical-site bleeding | 1.31 (1.13 to 1.52) | Not significant | 39.6 percent |
Cannulation-site bleeding | 1.05 (0.91 to 1.22) — not significant | Not significant | 39.3 percent |
For comparison: ischaemic stroke | 2.31 (1.77 to 3.00) | 4.50 (2.55 to 7.97) | 73.9 percent |
For comparison: any bleeding against any thrombosis | — | Bleeding 1.69 (1.49 to 1.93); thrombosis 1.23 (1.08 to 1.41) | Whole-cohort mortality 34.9 percent |
The events add up. In venoarterial support the adjusted odds of death were 1.43 for one event, 1.86 for two and 3.27 for three or more; in venovenous support 1.66, 1.68 and 2.16. And in the venoarterial analysis time on ECLS was not independently associated with mortality, which the authors take as evidence that these events are "morbid and not simply markers of longer runs."
The reading this chapter takes. The hierarchy is the same in both configurations and it follows the vascular bed, not the volume of blood. Bleeding into the brain, the lungs or the gut is associated with death; bleeding from a hole in the skin is not. The venovenous authors suggest why: surgical bleeding is "more amenable to correction and less likely to provoke organ damage than hemorrhage in the brain, lungs or gastrointestinal tract."
One inconsistency to record. The venoarterial abstract lists female sex as a risk factor for bleeding; the results describe it as a trend, and the discussion says women "trended toward a higher risk." The book treats female sex as unconfirmed.
Certainty: moderate for the ordering, concordant across two very large analyses; low that any single odds ratio is causal (§36.9).
Clinical pearl · The same bed-space, two conversations
- A patient bleeding from the cannula site needs a dressing, a stitch, a look at the anticoagulation and a check that the cannula is intact (§36.6, Chapter 34). It is a nursing and procedural problem, and in the registries it was not associated with death.
- A patient who starts bleeding from the airway or the gut needs the whole haemostatic system examined (§36.7), a senior decision about anticoagulation (Chapter 45), and a conversation that recognises this is one of the complications most strongly associated with death on ECMO.
Name which conversation is happening. A handover that says "bleeding" without saying where loses the most important piece of information.
36.5 Why ECMO patients bleed: five mechanisms
Physiology · Five mechanisms, often at once, each owned by another chapter
The sources give overlapping lists. The Red Book: bleeding complications are common "due to anticoagulation, underlying critical illness, and a high occurrence of conditions with bleeding diatheses (eg, renal failure, sepsis, thrombocytopenia, acquired von Willebrand factor deficiency)." The ISCCM manual names overanticoagulation, coagulopathy, platelet dysfunction or thrombocytopenia, and acquired von Willebrand syndrome, and adds hyperfibrinolysis.
Mechanism | What drives it | The bedside question | Owner |
1. The anticoagulant | Heparin or a direct thrombin inhibitor; the ISCCM manual calls heparin overdose "one of the common causes of bleeding in patients on ECMO" | Is the anticoagulant effect higher than intended, by an assay that measures it? | Chapters 40 to 42 |
2. The circuit consuming the patient | Contact activation, thrombin generation, fibrinogen consumption, device-induced hyperfibrinolysis; the ISCCM manual adds endothelial release of tissue plasminogen activator | Is D-dimer rising and the platelet count falling over three to four days? Chapter 31 §31.5: then the circuit belongs near the top of the list, and the treatment is an exchange rather than more product | Chapter 31 §31.5 |
3. Platelets — number and function | Shear from the pump; the ISCCM manual cites reduced glycoprotein Ibα and glycoprotein VI; dilution at initiation with a crystalloid prime; heparin-induced thrombocytopenia | Is the count low, and is the function low even if the count is not? | Chapters 39 and 44 |
4. Acquired von Willebrand syndrome | Shear destroys the largest von Willebrand multimers. The ISCCM manual: "seen in most patients on ECMO, but it is highly under recognized" | Have multimers been requested by name? Chapter 29 found the screening ratio misses about half, because antigen rises while function falls | Chapters 29 and 39 |
5. The patient and the procedures | Renal failure, sepsis, liver dysfunction, acidosis at cannulation; and every invasive procedure | What has been done to this patient in the last 24 hours, and what organ failure is present? | This chapter, §36.6 and §36.8 |
The practical point is that they coexist. A patient on heparin, with renal failure, a falling platelet count and loss of large von Willebrand multimers has four of the five at once. Correcting one and finding the bleeding continues is not a failure of that correction; it is the next mechanism showing.
Certainty: high that the mechanisms exist and coexist, as the primary texts state; the five-question framing is this book's.
Pitfall · Procedures are the most preventable cause
ECMO in the Adult Patient is direct: "Avoidance of any unnecessary surgical procedures is the most important way to prevent bleeding." Its list of culprits includes the obvious and the easily forgotten:
- Operating-theatre procedures such as thoracotomy.
- Bedside procedures — vascular access, chest drain insertion, tracheostomy.
- "Seemingly 'minor' procedures" — nasogastric tube insertion and transoesophageal echocardiography.
- Lumbar puncture is contraindicated. Intramuscular injections and venepuncture for blood sampling should be avoided.
The Red Book adds chest tube placement and excessive in-line suctioning, and asks that all procedures be performed with caution. Its procedures chapter asks for "absolute hemostasis" when operating on a patient on ECLS — aggressive packing, draining, and sometimes closing a cavity to tamponade it or leaving it open for re-exploration.
The rule this book draws: every procedure on a patient on ECMO has a bleeding cost, and the cost is paid by the patient, not the operator. Ask whether it is needed today, and whether the person doing it has done it on a patient on ECMO before.
36.6 Site by site: what is done before systemic treatment
Local and site-specific measures, from the primary sources
Site | What the sources say | What the book adds |
Cannulation site — the commonest | ECMO: A Practical Guide to Management: "Among all bleeding complications, cannulation sites are the most common source of bleeding. Although most cases are minimal, intervention is sometimes required." First verify adequate anticoagulation levels, platelet count, prothrombin time and fibrinogen; anticoagulation targets can be reduced and topical haemostatic agents applied; surgical cannulation may need re-exploration if conservative measures fail. Avoiding large incisions and over-dilatation at insertion makes the cannula fit tightly. ECMO in the Adult Patient: local pressure and packing first | Persistent oozing that does not respond to local measures is a reason to inspect the cannula itself — Chapter 34 §34.8's cracked cannula oozed for 13 days before the crack was found |
Surgical and procedural sites | Red Book: "absolute hemostasis" — aggressive packing, draining, closing a cavity to tamponade it, or leaving it open for exploration | The bleeding cost of a procedure is paid for days; §36.5 |
Tracheostomy | Red Book: bleeding from irritation around a tracheostomy "can be easily controlled by removing the tracheostomy, intubating the patient orally and packing the tracheostomy wound." The tracheostomy tube can usually be replaced in several days | A reversible manoeuvre worth knowing before the bleeding becomes a crisis |
Airway and lungs | The Red Book lists oropharyngeal bleeding and pulmonary haemorrhage among serious bleeding complications and excessive in-line suctioning as a cause. A narrative review in the project library suggests frequent bronchoscopy to clear the airway | Pulmonary haemorrhage is the second most lethal bleed in both registries (§36.4); it is a medical bleed and goes straight to §36.7 |
Gastrointestinal tract | Listed as a serious complication by the primary texts. Zhang and colleagues (BMC Gastroenterol 2026, abstract read) found GI bleeding in 17.8 percent (62 of 349) of venoarterial patients at one centre, associated with higher 28-day mortality (HR 2.25, 1.67 to 3.05) after weighting; independent predictors were continuous renal replacement therapy, baseline INR and baseline calcium | A medical bleed; §36.7. Endoscopic timing is Chapter 45's decision |
Chest, abdomen, retroperitoneum — the occult sites | ECMO in the Adult Patient: "External visible bleeding (e.g. from the site of a chest drain) should prompt assessment for occult bleeding (e.g. into the pleural space)." The Red Book lists haemothorax; a narrative review lists intrathoracic, abdominal or retroperitoneal haemorrhage | Femoral cannulation is the setting for retroperitoneal bleeding — every retroperitoneal injury in Chapter 34 §34.9's malposition data followed femoral insertion |
Pericardium | Tamponade is counted as a surgical bleed in both registries | Chapter 26 owns it: on a circuit it has no physical signs, and a rising central venous pressure is a tamponade sign until imaged |
Brain | Red Book: when significant bleeding or intracranial haemorrhage is identified, anticoagulation should be stopped immediately, coagulopathy reversed and platelets kept above 50,000/mm³; any change in neurological examination should prompt evaluation. ECMO in the Adult Patient: intracranial haemorrhage "should not be considered a reason to end support with ECMO," and venovenous ECMO "can be continued for weeks without any other anticoagulant than the heparin coating on the circuit" | Chapter 38a owns it. It is the most lethal event in both registries |
Certainty: moderate for the local measures, which are consistent expert practice; low for the single-centre GI data.
Danger · The circuit can be the first thing to notice a bleed
ECPR and Resuscitative ECMO: "drainage insufficiency (chatter) may be an early indicator of bleeding and intravascular volume loss, its absence does not rule out significant haemorrhage." The ECMO simulation text's haemorrhage scenario shows the same sequence: bleeding from a cannulation site, then tachycardia, hypotension, a falling central venous pressure, falling pre- and post-membrane pressures and low-flow alarms.
So a new low-flow alarm with a more negative inlet pressure is a haemorrhage question as well as a cannula question (Chapter 34 §34.2). And the reverse is also true: a stable circuit does not mean the patient is not bleeding, because the circuit only notices once the intravascular volume has fallen far enough to starve the drainage cannula.
ECPR and Resuscitative ECMO adds that during volume resuscitation for bleeding, ECMO blood flow may need to be adjusted to avoid ongoing drainage insufficiency, and that the circuit's heater can be used to keep the patient normothermic — hypothermia worsens coagulopathy.
36.7 A falling haemoglobin: the five questions
The diagnostic sequence — this book's synthesis of §36.5 and §36.6
Question 1 — Where is it? Look at every site: cannula, wounds, drains, airway, nasogastric aspirate, stool, urine. If the visible bleeding does not account for the fall in haemoglobin, image the occult sites — chest, abdomen, retroperitoneum — before concluding that it is dilution or haemolysis. A falling haemoglobin can also be haemolysis (Chapter 37) or phlebotomy (Chapter 29 records daily phlebotomy volume as a number to chart).
Question 2 — How fast? Units transfused per 24 hours, drain output per hour, haemoglobin trend. The registry threshold for a significant bleed is transfusion-based (§36.3), and so is Mazzeffi's signal (§36.9).
Question 3 — Is the anticoagulant higher than intended? By an assay that measures anticoagulant effect, not by the activated clotting time alone (Chapter 42).
Question 4 — Is the circuit consuming the patient? Rising D-dimer and falling platelets over three to four days, with or without a low fibrinogen — Chapter 31 §31.5's pattern. If yes, the answer may be an exchange, not more product.
Question 5 — Is platelet function or von Willebrand factor the problem? A normal platelet count does not exclude platelet dysfunction, and a normal von Willebrand activity-to-antigen ratio does not exclude acquired von Willebrand syndrome. Request multimers by name (Chapter 29, Chapter 39).
And alongside all five: what organ failures and procedures are present — renal failure, liver dysfunction, acidosis, a recent line or drain.
What this sequence does not do is choose the treatment. That is Chapter 45's, and the sources disagree about the targets (§36.12).
36.8 Who bleeds
Evidence · Risk factors from the two registry analyses
Venoarterial (Chung) — any bleeding | Venovenous (Nunez) — any bleeding | |
Higher risk | Central cannulation (OR 1.89); each day on ECLS (OR 1.06); renal failure (OR 1.3); polymethylpentene oxygenator (OR 1.15); older age (OR 1.05 per 10 years); pre-ECLS cardiopulmonary bypass (OR 1.79) | Pre-ECMO acute kidney injury; vasopressor use before ECMO; longer ECMO duration; younger age; higher pH |
Lower risk | Later year of support (OR 0.93 per year); higher pH (OR 0.93 per 0.1); Asian race | Later year of support; older age |
Medical bleeding specifically | Longer time on ECLS (OR 1.07 per day); concomitant respiratory failure (OR 1.3); lower pH | Pre-ECMO acute kidney injury, vasopressors, longer duration; lower P/F ratio (weak) |
Three things stand out.
- Kidney failure appears in both. The venovenous authors link it to the platelet dysfunction of renal failure, and Zhang's single-centre GI data also found continuous renal replacement therapy an independent predictor. The patient on renal replacement is a patient at higher bleeding risk.
- Duration appears in both, for bleeding and for thrombosis. Every day adds risk. Both groups of authors draw the same conclusion: minimise ECMO duration and reassess readiness for decannulation frequently.
- Age runs in opposite directions. Older age raised bleeding risk on venoarterial support and lowered it on venovenous support. The venovenous authors suggest residual confounding — younger patients are offered longer runs. Treat age as unresolved.
Certainty: low to moderate — associations in registry data with no anticoagulation or laboratory variables. They describe who bleeds, not why.
36.9 Controversy 1 · Does bleeding kill, or does it mark the patient who was going to die?
The clinical question. Patients who bleed on ECMO die more often. Is the bleeding causing the deaths — so that preventing it would save lives — or is it a marker of how sick they already were?
Position A: bleeding is causal, at least for medical bleeding. Both registry analyses found bleeding associated with death after adjustment (venovenous any-bleeding OR 1.69), a dose-response with the number of events, and a hierarchy that follows the organ damaged — brain, lungs, gut — rather than the volume lost. The venoarterial authors add that time on ECLS was not independently associated with mortality, so the events are "morbid and not simply markers of longer runs." A haemorrhage into the brain is not plausibly a mere marker.
Position B: much of the association is confounding, and transfusion may be the real signal. Mazzeffi M, Greenwood J, Tanaka K, et al. "Bleeding, Transfusion, and Mortality on Extracorporeal Life Support: ECLS Working Group on Thrombosis and Hemostasis." Ann Thorac Surg 2016;101(2):682–689. DOI 10.1016/j.athoracsur.2015.07.046. Abstract read. In 132 adults over about three years at one centre, serious bleeding occurred in 56.1 percent. The crude odds ratio for death with bleeding was 2.22 (1.00 to 4.94); after adjustment it was 0.90 (0.37 to 2.19). But each unit of red cells transfused carried an odds ratio of 1.03 (1.00 to 1.06) after adjustment. The authors' conclusion: "only the amount of red blood cell transfusion is associated with inhospital mortality after controlling for confounding variables."
Where the evidence actually sits.
- All of it is observational. The registry analyses have no anticoagulation or laboratory data and cannot separate a bleed caused by the circuit from a bleed caused by the disease. Mazzeffi's study is small — its adjusted interval runs from 0.37 to 2.19, compatible with a large harm or a large benefit.
- The surgical-bleeding result is the internal control. In both registries, cannulation-site bleeding was not associated with death. If bleeding were only a marker of illness severity, every kind of bleed should carry some of that marker. That the harmless-looking bleeds are harmless and the organ bleeds are not is an argument that the organ damage matters. This is the book's reading.
- Transfusion is on the causal path. More bleeding means more transfusion; adjusting for one while measuring the other splits the same effect two ways. Mazzeffi's two results are not necessarily in conflict with the registries — they may be the same signal carried by different variables. Also the book's reading.
What this book concludes.
- Medical bleeding should be treated as dangerous in its own right, not as a marker to be noted. The organ-specific hierarchy is consistent across configurations and plausible.
- Site bleeding should be treated as a problem of blood loss and transfusion, and the transfusion count should be tracked, because it was the variable associated with death in the one study that adjusted for both.
- Neither position supports a claim that any particular intervention reduces bleeding deaths. That requires trials, and Chapter 45 records what exists.
Certainty: low. The hierarchy is robust; causation is not established for any bleed.
36.10 Controversy 2 · Can we predict who will bleed?
The clinical question. If the patients who will bleed could be identified at cannulation, their anticoagulation could be set lower from the start. Can they?
Position A: yes, the risk factors are known and models exist. The registry analyses identify consistent predictors (§36.8). Willers and colleagues built the HEROES models from the ELSO Registry: HEROES V-V from 18,658 adults (2010 to 2020), 21.1 percent of whom bled (Artif Organs 2022;46(5):932–952, abstract read), and HEROES V-A from 28,767 adults, 29.0 percent of whom bled (Artif Organs 2022;46(11):2266–2283, abstract read). The authors describe each as the best available model for its configuration. A single-centre model for gastrointestinal bleeding on venoarterial support (Zhang 2026) reported an area under the curve of 0.797 in training and 0.753 in an external cohort.
Position B: not usefully, yet. The HEROES V-V model's area under the curve was 0.63, and HEROES V-A's was 0.66. An area of 0.5 is a coin toss; 0.63 to 0.66 means that for a random pair of patients, one who bled and one who did not, the model ranks them correctly roughly two times in three. That is a modest improvement on chance, from the largest datasets in the field. Both were internally validated only, and the HEROES V-A authors write that external validation is warranted. The better-looking GI model rests on 62 events in its derivation cohort and 10 in its external one; its external interval runs from 0.531 to 0.976.
Where the evidence actually sits. The largest models discriminate weakly; the better-performing model is small and wide. None has been tested as a tool to guide anticoagulation, and none includes the laboratory variables — anticoagulant effect, platelets, fibrinogen, von Willebrand function — that §36.5 says drive bleeding.
What this book concludes.
- Use the risk factors, not a score. Renal failure, pre-ECMO vasopressors, acidosis, central cannulation, recent surgery and a long run are each worth writing down at cannulation and each day, because they should shape how aggressively the team anticoagulates (Chapter 42, Chapter 45).
- Do not let a score set anticoagulation. A model that discriminates at 0.63 cannot safely decide who gets less heparin.
- The missing variables are the opportunity. A model that added laboratory haemostasis to the registry predictors is the obvious next study.
Certainty: moderate that current models discriminate weakly — their own reported AUCs show it; low for any clinical use.
36.11 A worked case: the dressing was not where the blood went
Worked example · Day 4 of femoral venoarterial ECMO, and three units overnight
The patient. Day 4 of femoro-femoral venoarterial ECMO for post-infarction cardiogenic shock. Flow 3.8 L/min, speed 3,500 rpm, inlet pressure minus 55 mmHg. Heparin by infusion. Continuous renal replacement therapy since day 2.
The night. The arterial cannula dressing has been oozing since cannulation and has been changed four times. Haemoglobin falls from 9.8 to 7.1 g/dL between 20:00 and 06:00; three units of red cells are given. At 04:00 a low-flow alarm sounds; the inlet pressure has become minus 95 mmHg and the drainage line is twitching. Flow recovers partly after fluid. Blood pressure is maintained on a rising noradrenaline dose.
The reading, question by question (§36.7).
Question | Finding | Meaning |
Where is it? | Dressing oozing, but the ooze does not account for a fall of 2.7 g/dL and three units | The visible site is not the main site. Image the occult ones — and femoral cannulation makes the retroperitoneum the first to look at |
How fast? | Three units in ten hours, still falling | A major bleed by any definition |
The circuit's view | Chatter and a more negative inlet pressure at 04:00 | ECPR and Resuscitative ECMO's early sign of intravascular volume loss — the circuit noticed before the team did |
Anticoagulant, circuit, platelets, von Willebrand factor | Sent; results pending | §36.7 questions 3 to 5 run in parallel, not after imaging |
Patient factors | Renal replacement therapy; rising vasopressor | Two of the registry risk factors for bleeding (§36.8) |
Computed tomography shows a large right retroperitoneal haematoma at the level of the arterial cannula.
What should happen, in order.
- Resuscitate and support, with the ECMO flow adjusted to stop the chatter while volume is restored, and the heater used to keep the patient warm.
- Stop treating the dressing as the problem. It is a surgical-type bleed at the skin, not associated with death in the registries. The retroperitoneal bleed is a vessel injury in an occult place, which is why it was missed.
- Call the surgical and interventional teams, and hand the systemic decisions — anticoagulation, products, antifibrinolytics — to Chapter 45's pathway.
- Record the transfusion count as a number with times.
The counterfactual. Change one finding — the haemoglobin is stable at 9.6 g/dL, no units given, no chatter. Then the oozing dressing is exactly what it looks like: the commonest bleed on ECMO, with no association with death, needing local pressure, a check of the anticoagulant and platelets, and a look at the cannula (§36.6).
The teaching point. The blood that is seen is not the measure of the blood that is lost. When the haemoglobin and the visible bleeding disagree, believe the haemoglobin and look for where the rest went.
This case is constructed to illustrate the mechanisms described in this chapter. The values are illustrative and are not drawn from a published report.
36.12 What this chapter could not close
Open questions, recorded rather than concealed
- Whether any bleed causes death, rather than accompanying it, is not established (§36.9). Every source is observational.
- The registry definitions miss minor and upper-airway bleeding (§36.3). No retrieved study reports the full burden of bleeding with a denominator of patient-days and a definition that includes it.
- Neither registry analysis has anticoagulation or laboratory data, so no retrieved study links a specific anticoagulant level, platelet count or von Willebrand result to a specific bleed in adults on ECMO.
- The sources disagree on the laboratory targets for a bleeding patient. ECMO in the Adult Patient: platelets over 150,000, INR under 1.5, fibrinogen over 2.0 g/L. ECPR and Resuscitative ECMO: fibrinogen over 1.5 g/L, platelets often at least 50,000, INR under 1.3. The Red Book, for significant bleeding or intracranial haemorrhage: platelets over 50,000. Chapter 45 must reconcile them; this chapter records the disagreement.
- No bleeding prediction model has been externally validated or tested as a guide to anticoagulation (§36.10).
- The venoarterial abstract and results disagree on female sex as a bleeding risk factor (§36.4).
- No source retrieved gives evidence for steroids or bronchoscopic strategies in pulmonary haemorrhage on ECMO; the suggestion appears in one narrative review only.
Corrections and additions owed to other chapters
- Chapter 31 (Circuit Thrombosis) is owed confirmation that §31.5 is now used as question 4 of the bleeding diagnostic sequence.
- Chapter 38a (Neurological Complications) is owed intracranial haemorrhage as the most lethal bleeding event in both registries (adjusted OR 7.71 venoarterial, 5.71 venovenous; 73.2 percent mortality venovenous), the Red Book's stop-reverse-platelets-over-50,000 instruction, and ECMO in the Adult Patient's position that it is not a reason to end support.
- Chapter 42 (Anticoagulation Monitoring) is owed the registry risk factors (§36.8) as inputs to the intensity of anticoagulation.
- Chapter 45 (Bleeding Management) is owed §36.7's five questions as its entry point, the three sets of disagreeing laboratory targets, the Red Book's statement that anticoagulation can be safely held for prolonged periods, particularly if blood flow is over 2.5 to 3 L/min, and ECMO in the Adult Patient's escalation table.
- Chapter 46 (Transfusion Strategy) is owed Mazzeffi's finding that red cell units, not bleeding, carried the adjusted association with death (OR 1.03 per unit), and the rule that transfusion counts are charted with times.
- Chapter 55 (AKI and CRRT) is owed renal failure and renal replacement therapy as bleeding risk factors in both registries and in Zhang's GI data.
- Chapter 88 (Morning Round Template) is owed bleeding recorded by site, volume and category (surgical or medical), never as "bleeding" alone.
Carried forward and still open: Chapter 34's missing cannula-days denominator; Chapter 35's missing adult incidence of circuit air; Chapter 33's retrograde-flow interval.
36.13 Evidence and certainty summary
Claim | Best evidence | Certainty |
About a third of adults on venoarterial and about a quarter on venovenous support have a registry-defined bleed | Chung 2020 (11,984) and Nunez 2022 (7,579), ELSO Registry. Full text read | Moderate — large, but threshold-defined and self-reported |
Venoarterial support is bleeding-dominant (62.1 percent of events); venovenous is thrombosis-dominant (54.9 percent) | Same | Moderate |
Cannulation-site bleeding is the commonest bleed and is not associated with death | Same; venoarterial OR 1.05 (0.91 to 1.22) | Moderate |
Intracranial, pulmonary and gastrointestinal haemorrhage are associated with death, in that order, in both configurations | Same | Moderate for the ordering; low for causation |
Complications have a cumulative association with mortality | Same; venoarterial OR 1.43, 1.86, 3.27 for 1, 2, 3 or more | Moderate |
Surgical bleeding fell over 2010 to 2017; medical bleeding barely did | Same | Moderate |
Registry definitions count bleeding only above a transfusion or intervention threshold, and exclude upper-airway bleeding | Red Book registry definitions; both registry analyses' limitations | High as a description of the definitions |
Five mechanisms — anticoagulant, circuit consumption, platelets, von Willebrand factor, patient and procedure — coexist | Red Book; ISCCM manual; ECMO: A Practical Guide to Management | High that they exist; the framing is this book's |
Avoiding unnecessary procedures is the most important way to prevent bleeding | ECMO in the Adult Patient | Moderate — expert statement |
Renal failure, pre-ECMO vasopressors, acidosis, central cannulation and duration of support are associated with bleeding | Chung 2020; Nunez 2022; Zhang 2026 for renal replacement and GI bleeding | Low to moderate |
Chatter can be an early sign of haemorrhage; its absence does not exclude it | ECPR and Resuscitative ECMO; ECMO simulation text | Moderate — mechanism clear |
Tracheostomy bleeding can be controlled by removing the tube, intubating orally and packing | ELSO Red Book | Moderate — expert technique |
After adjustment in one centre, bleeding was not associated with death but each red cell unit was (OR 1.03) | Mazzeffi 2016, 132 patients. Abstract read | Low — small, wide intervals |
GI bleeding on venoarterial support: 17.8 percent at one centre, HR 2.25 for 28-day mortality | Zhang 2026. Abstract read | Low — single centre |
Registry bleeding prediction models discriminate weakly (AUC 0.63 V-V, 0.66 V-A) | Willers 2022, HEROES V-V and V-A. Abstracts read | Moderate that they are weak; not externally validated |
Intracranial haemorrhage is not by itself a reason to end support; venovenous ECMO can run for weeks on the circuit coating alone | ECMO in the Adult Patient; Red Book on holding anticoagulation at flows over 2.5 to 3 L/min | Low to moderate — expert statements |
36.14 Key points
- "Bleeding" is two problems. Surgical bleeding — cannula, wound, tamponade — is a hole. Medical bleeding — brain, lung, gut — is a failure of the haemostatic system.
- The commonest bleed is the least informative. Cannulation-site bleeding is the most frequent in both registries and is not associated with death.
- The rarer bleeds carry the prognosis. Intracranial, pulmonary and gastrointestinal haemorrhage, in that order, are associated with death in both configurations.
- Venoarterial support bleeds; venovenous support clots. Bleeding was 62.1 percent of events on venoarterial support and about 45 percent on venovenous.
- Complications add up. Each additional bleeding or thrombotic event raises the association with death.
- Surgical bleeding has fallen over a decade; medical bleeding has not.
- Registry bleeding is major bleeding. Minor and upper-airway bleeding are not counted; compare like with like.
- "Minor" bleeding adds up. Chart it as a volume and a transfusion count, never as an adjective.
- Five mechanisms, often together: the anticoagulant, the circuit consuming the patient, platelets, von Willebrand factor, and the patient's organs and procedures.
- When one correction does not stop the bleeding, the next mechanism is showing.
- Avoiding unnecessary procedures is the most important way to prevent bleeding — including nasogastric tubes and transoesophageal echocardiography.
- Persistent cannula-site oozing that local measures do not stop is a reason to inspect the cannula itself.
- Tracheostomy bleeding can be stopped by removing the tube, intubating orally and packing the wound.
- When the haemoglobin and the visible bleeding disagree, believe the haemoglobin. Image the chest, abdomen and — after femoral cannulation — the retroperitoneum.
- The circuit may notice first. New chatter with a more negative inlet pressure can be haemorrhage.
- Rising D-dimer with falling platelets over days points at the circuit; the answer may be an exchange, not more product (Chapter 31).
- A normal platelet count and a normal von Willebrand ratio do not exclude dysfunction. Request multimers by name.
- Renal failure, vasopressors before ECMO, acidosis, central cannulation and a long run are associated with bleeding. Every day on the circuit adds risk.
- Bleeding scores discriminate weakly (AUC 0.63 to 0.66). Use the risk factors; do not let a score set anticoagulation.
- Whether bleeding causes death is not proven; the organ bleeds should be treated as though it does, and the transfusion count should be tracked because it may carry the signal.
Chapter answer: Bleeding on ECMO is common — about a third of adults on venoarterial support and about a quarter on venovenous support have a bleed large enough for the registry to record — but the word covers two different events. Bleeding from a cannula site, a wound or a drain is the commonest and is not associated with death; it needs local pressure, a check of the anticoagulant and platelets, and a look at the cannula. Bleeding into the brain, the lungs or the gut is rarer and is among the complications most strongly associated with death in both configurations, and it means the haemostatic system is failing for one or more of five reasons: too much anticoagulant, a circuit consuming the patient's clotting system, low or dysfunctional platelets, loss of large von Willebrand multimers, or the patient's own organ failure and recent procedures. So when the haemoglobin falls, find out where the blood went — which may not be where it is visible — how fast, and which mechanisms are active, before choosing the treatment that Chapter 45 describes. Whether preventing these bleeds would save lives has not been shown, and the models that try to predict them are only modestly better than chance; the best current tools are a careful look at every site, a transfusion count charted with times, and fewer procedures.