Treatment effects of Xuebijing injection in patients with sepsis by clinical phenotype: a post hoc analysis of the EXIT-SEP trial

Abstract

Background: Xuebijing injection (XBJ) could improve the outcomes of sepsis patients. However, sepsis is a heterogeneous syndrome, and it remains unclear which patients benefit the most. We aimed to identify the sepsis phenotypes most likely to benefit from XBJ treatment.

Methods: This post hoc analysis of the EXIT-SEP trial included sepsis patients from 45 intensive care units (ICUs) in China between October 2017 and June 2019. The XBJ group received 100 mL of XBJ every 12 h for 5 days, while the placebo group was given a volume-matched saline. Consensus k-means clustering was performed to reproduce previously identified sepsis phenotypes in relation to the SENECA classification (α, β, γ, and δ) based on 19 variables such as age, sex, temperature, and biochemistry/lab results. Clinical characteristics and outcomes (28-day mortality, ventilator-free days, ICU-free days), as well as the heterogeneity of treatment effects (HTE), were compared between the four different phenotypes and across treatment groups. We also developed a probabilistic model for phenotype assignment and evaluated its performance in derivation and internal validation cohorts. The EXIT-SEP trial is registered with ClinicalTrials.gov (NCT03238742).

Findings: Among 1760 patients (878 in the XBJ group and 882 in the placebo group), four sepsis phenotypes (α: 28.2%, β: 24.0%, γ: 28.9%, and δ: 18.9%) were replicated based on the SENECA classification. Phenotype α had the lowest 28-day mortality. Phenotype β was associated with older age, chronic illness, and renal dysfunction. Phenotype γ was characterized by respiratory dysfunction. Phenotype δ was associated with acidosis, elevated alanine transaminase, coagulation dysfunction, shock, and the highest 28-day mortality (32.5%). Compared with placebo, XBJ treatment was associated with lower 28-day mortality in patients with phenotype γ (p = 0.003) and δ (p = 0.033), while the treatment-by-phenotype interaction was not statistically significant. Additionally, patients with phenotype δ who received XBJ had more ventilator-free days and ICU-free days than those with phenotype α, with p for interaction < 0.001 for both outcomes. Finally, a parsimonious classifier model demonstrated good accuracy in phenotype prediction, with AUROCs of 0.937 (95% CI: 0.916-0.957) for α, 0.893 (0.861-0.924) for β, 0.945 (0.927-0.964) for γ, and 0.900 (0.866-0.935) for δ in the internal validation cohort.

Interpretation: We replicated four sepsis phenotypes in the EXIT-SEP cohort, with patterns similar to previously established phenotypes. XBJ treatment was associated with lower 28-day mortality in patients with phenotypes γ and δ, but these findings require further validation.

Funding: This study was funded by the National Natural Science Foundation of China; Noncommunicable Chronic Diseases-National Science and Technology Major Project; Zhongda Hospital Affiliated to Southeast University, Jiangsu Province High-Level Hospital Construction Funds; Nanjing Technology Development Program; The Pilot Project of the Flagship Hospital of Integrated Traditional Chinese and Western Medicines in Zhongda Hospital affiliated to Southeast University.

Keywords: Heterogeneity of treatment effect; Phenotypes; Sepsis; Xuebijing injection.

Fig. 1

Study flowchart. XBJ = Xuebijing injection.

Fig. 2

Heatmap showing the distribution of clinical variables across the four phenotypes. Each row represents a clinical variable, and each column corresponds to a phenotype. The color gradient indicates the number of standard deviations by which the average value in each phenotype deviates from the overall cohort mean, with red indicating higher and blue indicating lower values. The dendrogram on the left represents hierarchical clustering of variables based on their distribution similarity, allowing visualization of clinically related groupings. BUN = Blood urea nitrogen; SBP = Systolic blood pressure; PaO₂ = Partial pressure of oxygen; P/F ratio = PaO₂/FiO₂ ratio; CRP = C-reactive protein; WBC = White blood cell; ALT = Alanine transaminase.

Fig. 3

Kaplan–Meier curves for 28-day mortality by phenotype in sepsis patients.

Fig. 4

Comparison of 28-day cumulative survival between XBJ and placebo groups across α(A), β(B), γ(C), and δ(D)phenotypes using Kaplan–Meier curves. XBJ = Xuebijing injection.

Fig. 5

Receiver operating characteristic (ROC) curves for the multinomial regression model predicting sepsis phenotypes. The ROC curves display the ability to discriminate between each phenotype and all other phenotypes. Panel A shows the ROC for Phenotype α, Panel B for Phenotype β, Panel C for Phenotype γ, and Panel D for Phenotype δ. For each phenotype, the area under the curve (AUC) is reported for both the derivation and validation datasets, along with 95% confidence intervals (CIs). The AUC values reflect performance in distinguishing each phenotype (α vs. non-α, β vs. non-β, γ vs. non-γ, and δ vs. non-δ). The dashed diagonal line represents the performance of a random classifier.