Precision therapy for endotoxic septic shock: a step toward personalized treatment from EUPHAS to TIGRIS Authors: Su Hwan Lee Published: 2026 Since the introduction of the Sepsis-1 definition in 1991 [1], several studies have been conducted to improve our understanding of the pathophysiology and management of sepsis. Subsequently, the Sepsis-3 Task Force redefined sepsis as life-threatening organ dysfunction caused by a dysregulated host response to infection, thereby refining the clinical definition by abandoning the systemic inflammatory response syndrome criteria [2]. Protocolized care promoted by the Surviving Sepsis Campaign, jointly initiated by the European Society of Intensive Care Medicine and the Society of Critical Care Medicine, has contributed substantially to reducing sepsis-related mortality across healthcare systems globally [3]. However, improvements in sepsis survival have plateaued in recent years, suggesting that the benefits of a “one-size-fits-all” approach may have reached their limit [4]. The evolution of the definition of sepsis suggests that sepsis remains a syndrome rather than a disease. Even among patients who meet the same diagnostic criteria, the underlying pathogen, predominant immune dysregulation, and mechanisms of organ injury may vary substantially. Against this background, recent efforts have shifted toward more individualized treatment strategies based on patient-specific clinical and biological characteristics [4]. The pathophysiology of sepsis is highly complex, involving interactions between host susceptibility, pathogen burden and virulence, and a wide range of pathogen-associated molecular patterns (PAMPs) [5]. Among these, endotoxin is the most extensively characterized PAMP. Approximately 10%–15% of patients with sepsis—and roughly one-third to one-half of those with septic shock—demonstrate elevated circulating endotoxin activity [5]. However, randomized controlled trials of pharmacologic endotoxin neutralization have failed to demonstrate meaningful improvements in clinical outcomes [5]. As an alternative strategy, polymyxin B hemoperfusion has been proposed to selectively remove endotoxin from the bloodstream [6]. Polymyxin B hemoperfusion has remained one of the most debated therapies for septic shock over the past 15 years [3,6]. The recent TIGRIS trial, together with the preceding randomized controlled trials of polymyxin B hemoperfusion—EUPHAS, ABDOMIX, and EUPHRATES—warrants attention less for the magnitude of its treatment effect than for what it demonstrates about how sepsis trials might be designed and interpreted [7-10]. Accordingly, this editorial does not simply summarize these four trials in chronological order but instead asks three questions: what has changed across these trials, whether those changes are credible, and what should follow (Table 1). Table 1.Summary of key randomized controlled trials of polymyxin B hemoperfusion in septic shockTrial (year)Country (n)PopulationEAA screeningStatistical frameworkEarly stoppingBlindingPrimary outcomePrimary endpoint met ResultConclusionEUPHAS (2009)Italy (10 ICUs), n=64Abdominal sepsis/septic shockNoFrequentistYes—terminated at interim analysisOpen-label; no shamMAP and vasopressor change at 72 hrYes (hemodynamic endpoints)28-Day: 32% vs. 53% (unadjusted HR, 0.43; APACHE II–adjusted HR, 0.36)—a secondary outcomeHypothesis-generating; effect size probably overestimatedABDOMIX (2015)France (18 ICUs), n=243Peritonitis-induced septic shockNoFrequentistNoOpen-label; no sham28-Day mortalityNo28-Day: 27.7% vs. 19.5% (P=0.14)Neutral; lower-than-expected control mortality and incomplete delivery (81/119 completed both sessions)EUPHRATES (2018)USA/Canada (55 ICUs), n=450Septic shock + EAA ≥0.60Yes (EAA ≥0.60)Frequentist, with interim analyses and protocol amendment adding MODS ≤9 exclusionYes—enrollment closed at 450 of a planned 650Double-blind; sham (façade) hemoperfusion28-Day mortalityNo28-Day: 37.7% vs. 34.5% (P=0.49)Neutral overall; post hoc signal confined to EAA 0.60–0.89 with MODS >9TIGRIS (2026)USA (19 ICUs), n=157; randomized (106 vs. 51; 2:1) of 14,890 screenedSeptic shock + EAA 0.60–0.89 + MODS >9Yes (EAA 0.60–0.89)Bayesian, with an informative prior derived from the corresponding EUPHRATES subgroup (75% weighting)a)NoOpen-label; masked outcome assessment28-Day mortality (Bayesian)Yes, by the prespecified Bayesian decision rule28-Day: 39% vs. 45%; PP of benefit 95.3% (aOR, 0.67; 95% CrI, 0.39–1.08)Supports benefit in an enriched population; the 28-Day result is sensitive to prior borrowing and ~1% of screened patients were randomized90-day: PP 99.4% (aOR, 0.54; 95% CrI, 0.32–0.87)EAA: endotoxin activity assay; ICU: intensive care unit; MAP: mean arterial pressure; HR: hazard ratio; APACHE: Acute Physiology and Chronic Health Evaluation; MODS: Multiple Organ Dysfunction Score; PP: posterior probability; aOR: adjusted odds ratio; CrI: credible interval.Mortality results are presented as polymyxin B versus control. In EUPHAS, 28-day mortality was a secondary outcome; the prespecified primary endpoints were hemodynamic.a)In TIGRIS, 157 patients were randomized; the primary analysis additionally borrowed information from the EUPHRATES treatable cohort (patients randomized at U.S. sites with MODS >9 and EAA 0.60–0.89) by means of an informative prior weighted at 75%. The posterior probability of benefit is conditional on that prior and is not equivalent to a P-value. Summary of key randomized controlled trials of polymyxin B hemoperfusion in septic shock EAA: endotoxin activity assay; ICU: intensive care unit; MAP: mean arterial pressure; HR: hazard ratio; APACHE: Acute Physiology and Chronic Health Evaluation; MODS: Multiple Organ Dysfunction Score; PP: posterior probability; aOR: adjusted odds ratio; CrI: credible interval. Mortality results are presented as polymyxin B versus control. In EUPHAS, 28-day mortality was a secondary outcome; the prespecified primary endpoints were hemodynamic. In TIGRIS, 157 patients were randomized; the primary analysis additionally borrowed information from the EUPHRATES treatable cohort (patients randomized at U.S. sites with MODS >9 and EAA 0.60–0.89) by means of an informative prior weighted at 75%. The posterior probability of benefit is conditional on that prior and is not equivalent to a P-value. The EUPHAS trial (2009), conducted across 10 Italian intensive care units (ICUs) and enrolling 64 patients with septic shock following emergency abdominal surgery, suggested a substantial survival benefit associated with polymyxin B hemoperfusion, with a reduction in 28-Day mortality (32% vs. 53%; unadjusted hazard ratio [HR], 0.43; Acute Physiology and Chronic Health Evaluation (APACHE) II–adjusted HR, 0.36) [8]. However, the trial was stopped early after an interim analysis, which may have overestimated the treatment effect. In addition, the primary endpoint was hemodynamic improvement rather than mortality, and endotoxin levels were not measured, leaving the mechanism of the observed benefit unclear. Despite these limitations, EUPHAS provided early evidence supporting the biological rationale for polymyxin B hemoperfusion and laid the groundwork for subsequent larger trials. The ABDOMIX trial (2015), which enrolled 243 patients with peritonitis-induced septic shock across 18 French ICUs, was designed to replicate the mortality benefit observed in EUPHAS [9]. However, no significant clinical benefit of polymyxin B hemoperfusion was observed, with a 28-day mortality rate of 27.7% in the intervention group versus 19.5% in the standard care group (P=0.14). The control group mortality rate was substantially lower than anticipated based on prior studies (37% in the trial design and 53% in EUPHAS), so the trial was substantially underpowered to detect the anticipated treatment effect. This may have reflected improvements in general critical care over time, which reduced the ability to detect a treatment effect. As in EUPHAS, endotoxin levels were not measured, raising the possibility that a substantial proportion of enrolled patients may not have had elevated endotoxin activity. The EUPHRATES trial (2018) introduced a pivotal advancement by incorporating a theragnostic design (using the same biomarker to identify the target condition and guide treatment allocation), prospectively enrolling 450 patients with confirmed endotoxemia (endotoxin activity assay [EAA] ≥0.60) across 55 North American centers and using a sham hemoperfusion circuit to achieve blinding. Despite this patient enrichment strategy, no significant difference in 28-day mortality was observed between polymyxin B hemoperfusion and standard care (37.7% vs. 34.5%, P=0.49) [7]. Subsequent post hoc analyses suggested a potential mechanistic limitation: the EAA shows reduced reliability above 0.90, a range in which endotoxin burden may exceed the adsorption capacity of the polymyxin B cartridge [11]. When the analysis was restricted to patients with intermediate endotoxin activity (0.60–0.89) and higher disease severity (Multiple Organ Dysfunction Score [MODS] >9), a consistent mortality signal was observed. This suggested that biomarker-based patient selection alone may be insufficient when the biomarker’s dynamic range is not aligned with the device’s treatment capacity, which provided the rationale for the design of the subsequent TIGRIS trial [11]. The TIGRIS trial, published in The Lancet Respiratory Medicine in 2026, is the most recent large randomized controlled trial evaluating polymyxin B hemoperfusion in septic shock, following nearly 17 years of investigation in this field [10]. The trial was conducted across 19 U.S. hospitals using a Bayesian design and enrolled 157 patients with endotoxic septic shock (ESS), defined as septic shock with an EAA value of 0.60–0.89 and MODS >9 [10]. These patient selection criteria were directly informed by post hoc analyses of EUPHRATES, which targeted a biologically plausible endotoxin range in which polymyxin B hemoperfusion was expected to be most effective [11]. The results indicated a high probability of clinical benefit with polymyxin B hemoperfusion [10]. At 28 days, the posterior probability of mortality benefit was 95.3% (APACHE II-adjusted odds ratio [OR], 0.67; 95% credible interval [CrI], 0.39–1.08), which increased to 99.4% at 90 days (OR, 0.54; 95% CrI, 0.32–0.87). The 90-day probability of benefit remained above 98% even under non-informative assumptions, and Kaplan–Meier survival curves showed progressive divergence over the 90-day period. Serious adverse events occurred in 30% of the intervention group and 22% of controls, with two events considered treatment-related and none resulting in permanent sequelae [10]. Three considerations may account for the attention TIGRIS is receiving, none of which depends on the magnitude of the treatment effect itself. First, TIGRIS is an uncommon example in critical care of a hypothesis generated by post hoc subgroup analysis being tested prospectively rather than being incorporated into practice or set aside. Whatever one ultimately concludes about polymyxin B hemoperfusion, this sequence—a neutral trial, a mechanistic explanation, a prospectively defined enriched population, and a dedicated trial—has rarely been completed in critical care. Second, TIGRIS makes falsifiable the hypothesis that the discordant results of earlier polymyxin B hemoperfusion trials reflected differences in patient selection rather than a lack of treatment efficacy. The trial was designed to test this hypothesis rather than merely to revisit it. Third, and perhaps most relevant to practice, ESS is a phenotype that can be operationalized at the bedside today. Transcriptomic endotypes remain largely research tools [12] and have not yet been prospectively used to allocate treatment, whereas the EAA can be measured within hours and MODS calculated from routinely available clinical and laboratory data. This is a practical advantage. The trade-off is a considerable restriction of the eligible population: approximately 1% of the 14,890 patients screened were ultimately randomized. This approach is analogous to the evolution of precision oncology, where treatment success has depended on molecular or biological stratification rather than broad syndromic diagnoses [13], but the sepsis literature now offers closer comparators. The SENECA project derived four clinical phenotypes (α, β, γ, δ) from routinely available data in 20,189 patients, demonstrated their reproducibility and their association with distinct host-response patterns, and—most relevant here—showed in simulation that the distribution of phenotypes within a trial population could alter that trial's conclusions [14]. The PROVIDE trial stratified patients by serum ferritin and monocyte HLA-DR expression into macrophage activation-like syndrome and immunoparalysis, and randomized them to anakinra or recombinant interferon-gamma accordingly [15]. The subsequent and much larger ImmunoSep trial met its primary endpoint of organ dysfunction improvement at day 9 in both immune phenotypes, but showed no significant difference in 28-day mortality [16]. Read alongside TIGRIS, the emerging pattern is that prospective biological selection is feasible and yields real but so far modest benefits—and that trials in this field increasingly succeed on organ-dysfunction or probabilistic endpoints rather than on unambiguous mortality reduction. In this context, ESS may represent a targetable subgroup, which would explain why unselected trials have failed to show benefit and why patient enrichment strategies are necessary [17]. The Bayesian framework used in TIGRIS allowed evidence from prior trials to be integrated into this enrichment strategy, but it also requires an approach to interpretation that differs from frequentist convention [18]: a 95% posterior probability of benefit is not equivalent to a P-value of 0.05 and carries no type I error guarantee. This distinction matters for how the two endpoints are weighed. Although the 28-day mortality was the prespecified primary endpoint, the 90-day results may provide a more reliable signal of the treatment effect. The TIGRIS trial had several limitations that warrant consideration. The trial was open-label, although study personnel were masked to treatment allocation and the primary outcome was an objective measure. Its Bayesian framework incorporated an informative prior derived from a subgroup of EUPHRATES (75% weighting), introducing dependence on prior assumptions. Importantly, the same EUPHRATES subgroup that generated the enrichment criteria for TIGRIS also supplied the prior distribution, so that dataset contributes to the analysis in two ways. Borrowed evidence is less persuasive than independent evidence, and this circularity is the most important limitation of the trial. The implications are evident in the results. When the trial data are analyzed without borrowing from EUPHRATES, the 28-day posterior probability of benefit falls from 95.3% to approximately 80% (OR, 0.80; 95% CrI, 0.38–1.49), whereas the 90-day probability remains above 98% (OR, 0.50; 95% CrI, 0.23–0.94). The 28-day treatment effect was therefore modest and substantially dependent on the prior: the adjusted odds ratio was 0.67 (95% CrI, 0.39–1.08), an interval that includes no effect, and the posterior probability of 95.3% exceeds 95% only marginally. The primary endpoint should not be described as robust, and warrants cautious interpretation in the absence of independent confirmation. The 90-day result was stronger and more stable (adjusted OR, 0.54; 95% CrI, 0.32–0.87; posterior probability 99.4%), and progressive separation of the survival curves is biologically plausible if early reversal of organ dysfunction has durable consequences. However, 90-day mortality was a key secondary endpoint; the strength of this finding is supportive evidence warranting independent confirmation rather than a substitute for a confirmatory trial. Additionally, the narrow eligibility criteria resulted in a very low enrollment rate (approximately 1% of screened patients), limiting the generalizability of these findings and indicating that only a small proportion of patients with septic shock meet the current criteria for endotoxin-targeted therapy. Key uncertainties remain regarding the optimal timing, dosing, and duration of polymyxin B hemoperfusion, and whether these findings can be replicated across different healthcare systems and patient populations. Whether additional biomarkers beyond EAA can refine patient selection remains unclear. Collectively, the four trials—EUPHAS, ABDOMIX, EUPHRATES, and TIGRIS—provide a growing body of evidence suggesting that conflicting results across trials are better explained by differences in patient selection than by a lack of treatment efficacy. The early survival signal observed in EUPHAS was not reproduced in ABDOMIX, likely because of the absence of endotoxin measurements. EUPHRATES introduced biomarker-guided patient selection but included patients with endotoxin levels beyond the effective adsorption range of the polymyxin B cartridge. TIGRIS addressed these limitations by prospectively enrolling patients with moderate endotoxemia and severe organ dysfunction, and observed a probable mortality benefit in this selected population, supporting the role of patient phenotyping in the design of future sepsis trials. These observations set the agenda for future work. A confirmatory randomized trial is required: blinded with a sham circuit as in EUPHRATES, conducted outside the United States to test external validity, and powered for 90-day mortality. The assay needs parallel work—point-of-care EAA measurement standardized, its performance characterized within the 0.60–0.89 band that now defines eligibility, and its value as a serial rather than a single static measurement determined. Optimal timing, number and duration of sessions, and interactions with source control and renal replacement therapy all remain undefined. Whether EAA combined with transcriptomic or proteomic markers, or with the dynamic response to a first session, identifies responders more precisely than a single static threshold is an open and testable question. Finally, cost-effectiveness, cartridge availability and the need for an extracorporeal circuit and anticoagulation will determine whether any confirmed benefit can be realized outside well-resourced centers. The trajectory from EUPHAS to TIGRIS provides a useful case study of how a therapy with a sound biological rationale can appear promising, subsequently fail, and then show benefit under more selective conditions, without any fundamental change in the underlying biology. TIGRIS provides the most credible evidence to date that polymyxin B hemoperfusion may reduce mortality in ESS in a population prospectively enriched on the basis of biological characteristics rather than defined solely by a clinical syndrome. However, it does not, by itself, establish efficacy: the primary endpoint is partly informed by borrowed prior information, the trial was open-label, and the eligible population represents only a small fraction of patients screened for septic shock. The more durable contribution may therefore be methodological. If confirmed, these findings would establish ESS as the first sepsis subgroup with a prospectively validated, biology-guided indication for targeted extracorporeal therapy. If not, the trial will nevertheless have demonstrated how a treatable trait can be defined, measured, and tested prospectively. Either outcome would represent progress. What should not follow is a return to enrolling undifferentiated patients with septic shock and expecting a different answer. Infographics Footnotes CONFLICT OF INTEREST No potential conflict of interest relevant to this article was reported. The author has received no research funding, consultancy fees, honoraria, travel support, or equipment from any manufacturer or distributor of polymyxin B hemoperfusion cartridges or EAAs, holds no related patent or equity interest, and has no financial or personal relationship with any entity involved in the trials discussed. Readers are referred to the funding and competing-interest declarations in the original trial reports; the TIGRIS and EUPHRATES trials were sponsored by the manufacturer of the device under evaluation. FUNDING None. ACKNOWLEDGMENTS None. AUTHOR CONTRIBUTIONS All the work was done by Su Hwan Lee. REFERENCES