Skip to main content

Foundational Trials

SAFE

Image: Shutterstock / sfam_photo

SAFE Study Investigators. A comparison of albumin and saline for fluid resuscitation in the intensive care unit. N Engl J Med 2004;350:2247-2256.

Publication

Context & Rationale

  • Background
    • Albumin was widely used for intravascular volume expansion on the basis of its oncotic properties and the expectation that less infused volume would be required than with crystalloids.
    • The 1998 Cochrane review associated albumin administration with a 6-percentage-point absolute increase in mortality, provoking a major challenge to established practice. Its small trials included different indications, including hypovolaemia, burns and hypoalbuminaemia.1
    • A subsequent meta-analysis of 55 trials involving 3504 patients did not find a significant overall mortality increase. The conflicting syntheses left substantial uncertainty about the safety and clinical value of albumin.2
  • Research Question/Hypothesis
    • In a heterogeneous population of adults requiring ICU fluid resuscitation, does using 4% albumin rather than 0.9% saline alter all-cause mortality within 28 days?
    • The stated null hypothesis was no mortality difference. The trial was powered to detect a difference; it did not specify a formal equivalence or non-inferiority margin.
  • Why This Matters
    • A common, costly supportive treatment lacked reliable evidence for a patient-important benefit, and a plausible safety concern required a sufficiently large randomised comparison.
    • SAFE also tested whether an international critical care collaboration could deliver a large, blinded trial of routine bedside care. Its design was published while recruitment was continuing.3

Design & Methods

  • Research Question:
    • Whether the choice of albumin or saline for ICU intravascular fluid resuscitation affects 28-day mortality, organ failure, duration of organ support or length of stay.
  • Study Type:
    • Investigator-initiated, international, multicentre, parallel-group, double-blind randomised trial in 16 closed multidisciplinary ICUs in academic tertiary hospitals in Australia and New Zealand.
    • Recruitment ran from November 2001 to June 2003. Central allocation used a secure website and minimisation, with stratification by institution and trauma at ICU admission.
    • The study was conducted by ANZICS Clinical Trials Group, the Australian Red Cross Blood Service and the George Institute for International Health. Funding included public agencies and CSL, which manufactured and supplied the masked fluids; data management and analysis were independent of the funding agencies.
  • Population:
    • Adults aged ≥18 years for whom the clinician judged that intravascular volume depletion required fluid resuscitation and considered both study fluids appropriate. The requirement had to be additional to maintenance, nutrition and replacement of ongoing losses.4
    • At least one supporting sign was required: heart rate >90/min; systolic pressure <100 mm Hg, mean arterial pressure <75 mm Hg, a fall of 40 mm Hg from baseline, or vasoactive support to maintain these pressures; central venous pressure <10 mm Hg; pulmonary capillary wedge pressure <12 mm Hg; respiratory variation in systolic or mean pressure >5 mm Hg; capillary refill >1 second; or urine output <0.5 mL/kg/hour for one hour.4
    • Exclusions included admission after cardiac surgery or liver transplantation, burns, planned plasmapheresis, previous albumin reaction, objection to blood products, brain death or likely brain death within 24 hours, and expected death within 24 hours with treatment limitations beyond a resuscitation restriction.4
    • Patients were also excluded after prior resuscitation prescribed in the study ICU during the current hospital admission, after resuscitation in a transferring non-study ICU, or after previous enrolment with completed SAFE follow-up. Pre-ICU resuscitation outside these circumstances was not prohibited.4
    • Prespecified subgroup comparisons concerned the presence or absence of trauma, severe sepsis and ARDS. Severe sepsis used infection, systemic inflammatory response criteria and associated organ dysfunction; these were pre-Sepsis-3 definitions.4
  • Intervention:
    • 4% human albumin solution, Albumex (CSL), for all ICU resuscitation episodes until death, ICU discharge or 28 days after randomisation.
    • Study fluid was supplied in masked 500-mL bottles. The treating clinician determined the amount and rate according to clinical response; a fixed cumulative dose, universal infusion rate or maximum resuscitation volume was not specified.
    • Albumin was prescribed to restore or maintain intravascular volume. There was no target serum albumin concentration and no scheduled replacement regimen for hypoalbuminaemia.
    • Maintenance fluids, specific replacement fluids, enteral or parenteral nutrition and blood products remained discretionary. Study allocation did not determine every intravenous fluid administered.
  • Comparison:
    • 0.9% sodium chloride for the same ICU resuscitation indications and treatment period, delivered using matching masked bottles and administration sets.
    • Monitoring, haemodynamic targets, vasoactive drugs and other care were determined by the treating team. Non-study resuscitation fluid could be administered when clinically required; actual departures from allocation were recorded.
    • Fluids administered outside the ICU were not controlled by the trial.
  • Blinding:
    • Patients, clinical staff and study personnel were blinded through specially designed masking cartons and administration sets. The masking method was formally tested before the trial.
    • Serum albumin measurements remained available and subsequently differed between groups, creating a possible indirect clue to allocation despite concealment of the fluid itself.
  • Statistics:
    • Power calculation: 7000 patients were required for 90% power to detect a 3-percentage-point absolute mortality difference from an anticipated 15% baseline mortality; the α used for this calculation was Not reported in the index paper or supplied appendix.
    • Analyses followed allocated treatment using available observations, without imputation of missing data. Categorical comparisons used χ² or Fisher’s exact tests; continuous comparisons used unpaired t-tests.
    • Kaplan–Meier survival estimates and log-rank testing assessed time to death. Subgroup heterogeneity was evaluated with tests for a common relative risk.
    • An independent data and safety monitoring committee reviewed interim analyses after 2333 and 4666 enrolments. A numerical stopping boundary and multiplicity-adjusted subgroup significance threshold were Not reported in the index paper or supplied appendix.
  • Follow-Up Period:
    • Primary outcome: all-cause mortality within 28 days of randomisation.
    • Secondary outcomes included survival time, new organ failure, duration of mechanical ventilation and renal replacement therapy, and ICU and hospital stay during the 28-day period.
    • The cardiovascular, respiratory, renal, haematological and hepatic SOFA components were assessed at baseline, daily for seven days and then every third day while in the ICU through day 28. Long-term function and quality of life were not index-trial outcomes.

Key Results

This trial was not stopped early. Recruitment reached 7000 allocations. Three duplicate randomisations were removed, leaving 6997 unique patients: 3497 assigned albumin and 3500 saline. The published primary analysis comprised 3473 and 3460 patients, respectively. Continuous values below are means ±SD; effect estimates and confidence intervals are reproduced as published.

Outcome 4% albumin 0.9% saline Effect p value / 95% CI Notes
Death within 28 days 726/3473 (20.9%) 729/3460 (21.1%) RR 0.99; absolute difference −0.2 percentage points RR: 95% CI 0.91 to 1.09; P=0.87. Absolute difference: 95% CI −2.1 to +1.8 percentage points Primary outcome; albumin minus saline for the absolute difference.
ICU length of stay 6.5±6.6 days 6.2±6.2 days Mean difference +0.24 days 95% CI −0.06 to +0.54; P=0.44 Measured within the first 28 days.
Hospital length of stay 15.3±9.6 days 15.6±9.6 days Mean difference −0.24 days 95% CI −0.70 to +0.21; P=0.30 Includes time in the ICU; follow-up truncated at day 28.
Duration of mechanical ventilation 4.5±6.1 days 4.3±5.7 days Mean difference +0.19 days 95% CI −0.08 to +0.47; P=0.74 No demonstrated reduction in organ support.
Duration of renal replacement therapy 0.48±2.28 days 0.39±2.0 days Mean difference +0.09 days 95% CI −0.0 to +0.19; P=0.41 Duration within 28 days; not an incident-AKI endpoint.
New organ failure No new failure: 1397 (52.7%); one organ: 795 (30.0%) No new failure: 1424 (53.3%); one organ: 796 (29.8%) Not reported P=0.85 for the complete distribution; 95% CI Not reported Available in 2649 albumin and 2673 saline patients; test includes zero to five new organ failures.
Study fluid on day 1 1183.9±973.6 mL 1565.3±1536.1 mL Not reported P<0.001; 95% CI Not reported Published overall albumin:saline volume ratio over the first four days was approximately 1:1.4.
Net positive fluid balance on day 1 1543.6±1619.7 mL 1990.5±2061.7 mL Not reported P<0.001; 95% CI Not reported Fluid balance was also less positive with albumin on days 2 and 3.
Packed red cells on day 2 106.5±321.4 mL 61.1±235.2 mL Not reported P<0.001; 95% CI Not reported The authors reported 71 mL more packed red cells per albumin-assigned patient over the first four days; mechanism uncertain.
28-day mortality: trauma 81/596 (13.6%) 59/590 (10.0%) RR 1.36 95% CI 0.99 to 1.86; P=0.06 Prespecified subgroup; trauma versus non-trauma interaction P=0.04.
28-day mortality: traumatic brain injury 59/241 (24.5%) 38/251 (15.1%) RR 1.62 95% CI 1.12 to 2.34; P=0.009 Post hoc examination within the trauma subgroup; baseline brain injury had been recorded prospectively.
28-day mortality: severe sepsis 185/603 (30.7%) 217/615 (35.3%) RR 0.87 95% CI 0.74 to 1.02; P=0.09 Prespecified subgroup; sepsis versus no-sepsis interaction P=0.06.
28-day mortality: ARDS 24/61 (39.3%) 28/66 (42.4%) RR 0.93 95% CI 0.61 to 1.41; P=0.72 Prespecified subgroup; ARDS versus no-ARDS interaction P=0.74; very imprecise.
  • Overall mortality was nearly identical, with no demonstrated improvement in organ support or length of stay. The mortality interval nevertheless includes potentially important absolute benefit and harm.
  • Albumin produced measurable physiological and volume differences, but the approximately 1:1.4 volume ratio was much smaller than a commonly assumed several-fold crystalloid requirement.
  • The brain injury finding was the principal safety signal. The sepsis result remained uncertain: neither its within-subgroup comparison nor its interaction test established a benefit.

Internal Validity

  • Randomisation and allocation:
    • Central concealed allocation and stratification by site and trauma substantially reduced selection bias. Almost 7000 unique patients provided considerable protection against chance imbalance in the overall comparison.
    • Baseline age was 58.6 versus 58.5 years and APACHE II score 18.7 versus 19.0. Mechanical ventilation was being used in 63.8% versus 64.8%. Mean baseline serum albumin was 27.4 versus 27.7 g/L.
    • The groups were clinically similar overall. Baseline central venous pressure was slightly higher with albumin allocation, 9.0 versus 8.6 mm Hg; a small isolated imbalance does not establish failed randomisation.
  • Follow-up and analysis population:
    • The supplementary flow diagram starts with 7000 allocations and lists 26 albumin and 41 saline allocations outside analysis, including two and one duplicate randomisations. The published analysed denominators are 3473 and 3460; duplicate allocations should not be mistaken for additional patients lost to follow-up.4
    • Most missing records followed withheld or withdrawn consent. Follow-up was very complete, but analysis necessarily depended on available outcomes rather than including every unique randomised patient with an observed endpoint.
    • Delayed consent was used in 6628 patients (94.7%), enabling recruitment during urgent treatment. Consent-related loss was small and was documented.
  • Performance and detection bias:
    • Masking reduced treatment-dependent decisions about fluid administration and co-interventions. Mortality was an objective primary outcome.
    • Albumin concentrations could give clinicians an indirect indication of treatment, although formal unblinding frequency was Not reported. Decisions about discharge and organ support were less objective than mortality.
    • The double-blind design makes a large systematic imbalance in unreported co-interventions less plausible, although it cannot exclude every difference in care.
  • Adherence, timing and treatment separation:
    • Ninety albumin-assigned and 107 saline-assigned patients received no study fluid. Additional non-study resuscitation fluid was given to 309 (8.8%) and 375 (10.7%), respectively.
    • Errors accounted for 189 versus 190 additional-fluid exposures; clinician preference accounted for 68 versus 103. These departures could attenuate the effect of a policy of exclusive allocated-fluid use.
    • Eligibility generally captured the first resuscitation requirement managed within the study ICU. Time from illness onset or injury to randomisation, and complete preceding emergency-department or theatre fluid exposure, were Not reported in the index paper.
    • Albumin had been administered during the preceding 72 hours to 127 albumin-assigned patients (3.7%) and 135 saline-assigned patients (3.9%). Prior exposure was therefore present but balanced.
    • Exposure was sufficient to produce clear separation: day-2 serum albumin was 30.8±6.4 versus 24.5±5.9 g/L, and day-2 study-fluid volume was 602.7±892.7 versus 954.0±1484.4 mL.
    • Mean arterial pressure remained similar, including 84.4 versus 84.2 mm Hg on day 2. The comparison therefore achieved broadly similar conventional haemodynamic endpoints with different fluid composition and volume.
  • Outcome assessment and statistical rigour:
    • A single, clinically important primary endpoint, a large completed sample and independent interim oversight support the overall result. There was no early termination that might exaggerate an apparent treatment effect.
    • The trial estimated an average effect across diagnoses. Heterogeneity does not invalidate that average, but it limits application to an individual disease subgroup.
    • New-organ-failure data were available for fewer patients than mortality data. Shorter organ-support durations can reflect recovery or earlier death, so they require interpretation alongside mortality.
    • Subgroup estimates had less precision than the primary result and were not protected by a reported multiplicity adjustment. The post hoc brain injury analysis requires separate assessment.

Conclusion on Internal Validity: Strong for the average 28-day mortality effect of an albumin-versus-saline ICU resuscitation strategy. Confidence is lower for small treatment differences, organ-specific safety and subgroup effects because of incomplete secondary-outcome data, treatment contamination, multiplicity and limited subgroup precision.

External Validity

  • Population representativeness:
    • The broad medical and surgical population, clinician-directed dosing and approximately 21% mortality are relevant to adult multidisciplinary ICUs. Participants were not uniformly in profound shock.
    • Children, burns, postoperative cardiac surgery, liver transplantation, planned plasmapheresis and patients already resuscitated within the specified ICU settings were excluded.
    • Patients with a specific indication or contraindication to either fluid were ineligible. The study therefore addresses clinical equipoise between the two fluids, rather than every possible indication for albumin.
    • The number screened and the proportion of all potentially eligible admissions enrolled were Not reported in the supplied flow diagram.
  • Applicability to current practice:
    • The most direct application is adult ICU volume resuscitation with 4% Albumex versus 0.9% saline in a well-resourced health system.
    • The results do not directly compare albumin with balanced crystalloids, 4% with 20–25% albumin, or albumin replacement to a serum concentration target.
    • They do not establish the optimal amount or timing of fluid, the benefit of a fluid bolus versus no bolus, or the value of albumin for specific cirrhosis-related indications.
    • Resource availability, patient selection and contemporary resuscitation practice affect transportability. The general population result must be considered separately from the subsequent evidence in traumatic brain injury.

Conclusion on External Validity: Strong for comparable adult mixed-ICU resuscitation, but moderate across contemporary critical care as a whole. Excluded populations, different albumin preparations and indications, prehospital care and settings with limited organ support require their own evidence.

Strengths & Limitations

  • Strengths:
    • A large completed trial addressing a frequent treatment and an important safety controversy with all-cause mortality as its primary endpoint.
    • International multicentre recruitment, concealed allocation, formally tested masking and very complete mortality ascertainment.
    • Clinician-directed treatment increased practical relevance while clear differences in albumin exposure and resuscitation volume demonstrated delivery of the intervention.
    • Prespecified diagnostic subgroup comparisons created a basis for clinically important subsequent investigation.
  • Limitations:
    • No formal equivalence framework; modest mortality effects remained compatible with the confidence interval.
    • A 28-day horizon, incomplete organ-failure observations and limited assessment of long-term recovery or specific adverse reactions.
    • Discretionary co-interventions, some non-study resuscitation and unmeasured pre-ICU exposure limited interpretation of the biological effect of albumin in isolation.
    • Limited precision in subgroups and restricted applicability to other concentrations, comparators and excluded patient groups.

Interpretation & Why It Matters

  • What SAFE established
    • For the average eligible adult ICU patient, 4% albumin did not improve survival or reduce organ-support requirements compared with saline.
    • The primary mortality confidence interval substantially weakened the earlier claim of a large general mortality penalty from albumin, while leaving room for modest effects in either direction.
  • Clinical value
    • Reducing infused resuscitation volume and increasing serum albumin did not translate into improved patient-important outcomes. Physiological advantages alone were insufficient justification for routine albumin use.
    • Crystalloid resuscitation remained a defensible default. The editorial highlighted that absence of demonstrated benefit makes the additional acquisition cost of albumin difficult to justify routinely.5
  • A lasting methodological contribution
    • SAFE demonstrated that a familiar bedside treatment could be evaluated under rigorous blinded conditions at international scale.
    • Its legacy also illustrates why an average treatment effect cannot establish safety in every subgroup, and why a provocative subgroup observation needs more detailed follow-up before its clinical meaning is settled.

Controversies & Other Evidence

  • Similar outcomes did not prove equivalence:
    • The authors described the treatments as clinically equivalent, including in their reply to correspondence. The statistical design supported failure to detect a difference, without a prespecified equivalence margin and a formal equivalence analysis. The accompanying editorial explicitly distinguished those conclusions.56
    • The published absolute mortality interval, −2.1 to +1.8 percentage points, is clinically more informative than P=0.87 alone: it excludes the large average excess suggested in 1998 but does not exclude every worthwhile benefit or important harm.
    • Using the 3-percentage-point difference sought in the power calculation as a retrospective equivalence margin would change the inferential question after the results were known. Non-adherence and non-study fluids also tend to make a superiority comparison more similar; they do not strengthen an equivalence claim.
  • Were the patients sufficiently hypovolaemic, and was the treatment contrast adequate?
    • The relatively preserved baseline pressure and central venous pressure prompted the criticism that SAFE mainly studied mild haemodynamic instability. The investigators replied that clinical practice treats volume depletion before profound hypovolaemia develops. Both observations are relevant: profound shock was not required, but substantial illness severity and mortality make dismissal of the cohort as insufficiently ill unconvincing.6
    • A single supporting sign, such as tachycardia or a central venous pressure below 10 mm Hg, did not establish fluid responsiveness. SAFE evaluated fluid choice once a clinician decided to resuscitate; it did not validate those entry signs as a contemporary resuscitation algorithm.
    • The roughly 1:1.4 albumin:saline ratio reflects administered study fluid under clinician-directed treatment. It is not a direct measurement of intravascular retention or a universal dose-conversion factor.
    • The substantial volume of non-study maintenance and replacement fluid was permitted in both groups. It should be distinguished from actual non-study resuscitation or crossover; counting all non-study fluid as protocol failure misrepresents the tested strategy.
  • Subgroups: interaction, multiplicity and clinical plausibility:
    • Trauma was prespecified and used in allocation stratification; the trauma interaction was P=0.04. The increased mortality within traumatic brain injury emerged from a post hoc examination of that subgroup, despite prospective recording of baseline brain injury.
    • A small within-subgroup P value is not itself proof that the treatment effect differs from that in other patients. Multiple subgroup examinations and the smaller numbers increase the possibility of a chance finding.
    • The severe-sepsis point estimate favoured albumin, but the interaction P=0.06 and RR confidence interval crossing 1 did not establish a disease-specific survival advantage. The investigators explicitly rejected an inference of proven benefit in their reply.6
  • SAFE-TBI: stronger evidence of harm, with the original randomised comparison retained:
    • The 2007 follow-up reviewed baseline records and CT scans, added patients identified through diagnostic coding and excluded misclassified cases. Of 515 potentially eligible patients, 460 were retained: 231 albumin and 229 saline. Assessors remained blinded.7
    • At 24 months, mortality was 71/214 (33.2%) versus 42/206 (20.4%): RR 1.63; 95% CI 1.17 to 2.26; P=0.003. For severe injury, GCS 3–8, mortality was 41.8% versus 22.2%: RR 1.88; 95% CI 1.31 to 2.70; P<0.001.7
    • Adjustment for major baseline prognostic factors retained an association with mortality: OR 1.70; 95% CI 1.03 to 2.83; P=0.04. The moderate-injury estimate was imprecise, RR 0.74; 95% CI 0.31 to 1.79; P=0.50, and should not be interpreted as evidence of safety.7
    • Favourable neurological outcome at 24 months occurred in 47.3% versus 60.6%: RR 0.78; 95% CI 0.65 to 0.94; P=0.007. This difference largely reflected mortality; the survivor-only comparison did not demonstrate a difference in functional recovery.7
    • Post hoc analysis does not automatically convert randomised treatment assignment into an observational exposure. The comparison retained the original allocation in a subgroup defined by baseline injury, supporting a causal interpretation more strongly than a treatment-selected cohort. However, retrospective case reclassification, 17 versus 23 patients lost to 24-month follow-up and analysis prompted by an existing signal reduce certainty.
    • The follow-up established persistence of the signal and examined injury severity more carefully, but most deaths had already occurred by day 28. It therefore provides corroboration within the same trial, rather than independent replication or evidence of a newly emerging late toxic effect.7
  • Brain injury mechanism and formulation matter:
    • The 2007 correspondence proposed dilutional coagulopathy and questioned whether fixed fluid allocation prevented appropriate individualisation. Maintenance and replacement fluids were discretionary, and treatment volume was clinician-directed; randomisation concerned resuscitation-fluid type. Increased red-cell use was observed, but this did not establish coagulopathy as the cause of excess mortality.8
    • In SAFE-TBI, the binary measure of intracranial hypertension, pressure >30 mm Hg on two readings, did not differ significantly. A later analysis of 321 patients with intracranial pressure monitoring used longitudinal modelling to address informative loss of measurements and associated albumin with rising pressure during the first week in patients whose monitoring ended in that week.79
    • That analysis supports an intracranial mechanism but is post hoc and sensitive to monitoring, treatment and dropout processes; it is not definitive proof of mediation. Albumex was relatively hypotonic, approximately 260 mOsm/L, so the trial cannot separate an albumin-molecule effect from an effect of its formulation. It provides no positive evidence that another albumin preparation is safe for routine resuscitation in traumatic brain injury.9
  • Sepsis and hypoalbuminaemia: follow-up analyses did not settle efficacy:
    • The detailed SAFE severe-sepsis analysis found an adjusted mortality OR of 0.71; 95% CI 0.52 to 0.97; P=0.03, using the 919/1218 patients with complete baseline data. This more favourable complete-case adjusted result did not replace the uncertain unadjusted randomised estimate, and renal replacement therapy use was similar: 18.7% versus 18.2%.10
    • The baseline-albumin analysis included 6045 patients. Mortality ORs were 0.87 for albumin ≤25 g/L and 1.09 above 25 g/L, with unadjusted interaction P=0.08. An adjusted dichotomised interaction reached P=0.04, but the continuous-albumin interaction was absent and multiple comparisons were unadjusted. The evidence did not identify a robust serum-albumin threshold for treatment benefit.11
    • Low serum albumin can be prognostic without being a treatment target. SAFE administered albumin for volume resuscitation, so these analyses must also be distinguished from trials deliberately correcting albumin concentration.
  • Subsequent sepsis trials: ALBIOS and ARISS:
    • ALBIOS randomised 1818 patients with severe sepsis to crystalloids with or without 20% albumin targeting serum albumin ≥30 g/L. Mortality was 31.8% versus 32.0% at 28 days (RR 1.00; 95% CI 0.87 to 1.14; P=0.94) and 41.1% versus 43.6% at 90 days (RR 0.94; 95% CI 0.85 to 1.05; P=0.29). Its exploratory septic-shock finding did not establish an overall benefit.12
    • ARISS, published in 2026, tested 20% albumin replacement targeting ≥30 g/L in septic shock. It stopped for poor recruitment after 440 of a planned 1662 patients. Ninety-day mortality was 43.3% versus 45.9% (RR 0.94; 95% CI 0.76 to 1.17; P=0.71). Albumin was permitted for early resuscitation in controls and was received by 48.6%, further limiting the treatment contrast. The trial did not establish benefit and remained imprecise.13
    • These trials address related but distinct strategies: concentrated albumin supplementation alongside crystalloids, rather than substitution of 4% albumin for saline at each resuscitation episode. Taken together, they do not provide a reproducible mortality benefit sufficient to support routine albumin in sepsis.
  • Systematic and observational evidence:
    • The 2018 Cochrane review found little or no mortality difference for albumin or fresh frozen plasma versus crystalloids: RR 0.98; 95% CI 0.92 to 1.06 at the end of follow-up, across 20 studies and 13,047 participants. Its grouping of albumin with plasma, varied populations and reliance on large existing trials mean it is supportive synthesis rather than an independent replication of SAFE or evidence that all colloids share the same safety profile.14
    • In a 2024 analysis of 1541 CLASSIC participants, albumin use was associated with gastrointestinal infection, higher noradrenaline doses and treatment site. Albumin itself was not randomised in that analysis. This documents substantial clinical selection and helps explain why observational comparisons of albumin recipients and non-recipients are vulnerable to confounding by indication.15
  • Population and setting cannot be ignored: FEAST:
    • In African children with severe infection and impaired perfusion, FEAST found similar 48-hour mortality with albumin and saline boluses, 10.6% versus 10.5%, but lower mortality with no bolus, 7.3%. Any bolus versus control gave RR 1.45; 95% CI 1.13 to 1.86; P=0.003. This does not overturn SAFE’s adult comparison; it demonstrates that similarity between two fluids cannot establish the benefit or safety of giving a bolus in a different population and care setting.16
  • Cost, safety and sponsorship:
    • Acquisition-cost comparisons strongly favoured saline, but SAFE did not include a funded formal cost-effectiveness evaluation. Historical wholesale-price calculations in the correspondence were not measurements of total trial costs and should not be presented as current economic estimates.6
    • Greater red-cell use may reflect haemodilution, altered coagulation or other treatment effects; the trial did not establish the mechanism. Absence of an overall mortality difference does not fully characterise rare reactions or disease-specific harm.
    • CSL’s financial and material contribution and share ownership disclosed by two investigators warrant transparency. Independent data analysis, concealed treatment and objective mortality ascertainment are substantive safeguards; sponsorship alone is not evidence that the observed result was biased.
  • Current guidelines and the modern bedside decision:
    • The 2024 ESICM fluid-choice guideline conditionally favours crystalloids over albumin in general critical illness and sepsis, with moderate-certainty evidence. For traumatic brain injury, it conditionally favours isotonic saline over albumin and over balanced crystalloids, with very-low-certainty evidence.17
    • The 2024 International Collaboration for Transfusion Medicine Guidelines advises against routine albumin as first-line volume replacement or simply to increase serum albumin in critically ill adults. Its recommendations for specific cirrhosis complications concern different indications.18
    • The 2026 Surviving Sepsis Campaign suggests crystalloids alone over crystalloids with supplemental albumin, with moderate-certainty evidence. Its remarks allow selective supplementation after large crystalloid volumes or in cirrhosis and advise avoiding albumin in traumatic brain injury. It generally favours balanced crystalloids over saline for initial sepsis resuscitation, with saline preferred when traumatic brain injury coexists.19
    • The practical conclusion is to reassess whether fluid is indicated, use an appropriate crystalloid as the usual resuscitation fluid, and reserve albumin for a defined indication or selected circumstance. SAFE does not establish saline as the optimal crystalloid for every patient.

Summary

  • SAFE randomised 6997 unique adult ICU patients to clinician-directed resuscitation with 4% albumin or 0.9% saline under double-blind conditions.
  • Twenty-eight-day mortality was 20.9% versus 21.1%: RR 0.99; 95% CI 0.91 to 1.09; P=0.87. This was a precise overall comparison, but not formal proof of equivalence.
  • Albumin reduced study-fluid requirements, with an approximate first-four-day ratio of 1:1.4, without demonstrated improvements in organ support or length of stay.
  • The post hoc traumatic brain injury signal persisted at 24 months in SAFE-TBI and supports avoidance of albumin for resuscitation in this population.7
  • The sepsis signal did not develop into a reproducible mortality benefit in subsequent trials. Current guidance supports crystalloids as the default, with selective albumin use for specific circumstances.121319

Overall Takeaway

SAFE was a landmark trial because it replaced a major uncertainty about a ubiquitous treatment with a large, rigorously blinded comparison of patient-important outcomes. It showed no overall survival advantage from 4% albumin over saline, while its brain injury follow-up demonstrated why general reassurance cannot be extended to every diagnosis. Its enduring clinical lesson is to select resuscitation fluids by demonstrated benefit, formulation and patient context rather than physiological appeal alone.

Overall Summary

  • In adult mixed-ICU resuscitation, 4% albumin and saline produced similar 28-day mortality; albumin’s modest volume advantage did not improve clinical outcomes. Use crystalloids routinely, reserve albumin for selected indications, and avoid albumin resuscitation in traumatic brain injury.

Bibliography

Last updated September 10th, 2026