Relationship Between Phenotyping and Individualized Absolute Risk Differences in Sepsis: A Secondary Analysis of Two Approaches in Two Multicenter Trials.

OBJECTIVES

Sepsis trials likely include patients who vary in response to therapeutic interventions. The optimal approach to identify such differences in treatment response remains unclear. Estimating individualized absolute risk differences (iARDs) to model treatment response at an individual patient level using supervised effect models applied to randomized trial data may be informative. We explored the relationship between two subgrouping approaches and a recently published iARD model for the effect of early goal-directed therapy (EGDT) resuscitation in sepsis.

DESIGN

Secondary analysis of the Protocolized Care for Early Septic Shock (ProCESS) and Australasian Resuscitation in Sepsis Evaluation (ARISE) trials. We applied clinical subtypes (α, β, γ, δ) to 829 ProCESS and 1588 ARISE patients and biologic "hyperinflammatory" and "nonhyperinflammatory" subphenotypes to 363 ProCESS patients with biomarker data using established methods. We predicted iARDs with supervised learning using clinical variables as predictors and 90-day mortality as the primary outcome. We evaluated iARD variability within subgroups.

SETTING

Eighty-one sites worldwide.

PATIENTS/SUBJECTS: Adults with septic shock.

INTERVENTIONS: EGDT or usual care.

MEASUREMENTS AND MAIN

RESULTS

The average treatment effect of EGDT appeared to vary within both clinical and biologic subphenotypes. EGDT appeared potentially beneficial in the β and nonhyperinflammatory subphenotypes but harmful in the γ and hyperinflammatory subphenotypes. However, the predicted iARDs within each subgroup ranged from considerable harm to considerable benefit. For example, for the β-subtype, the average mortality reduction from EGDT was 8.5% (95% CI, -0.4 to 17.5), but the iARDs ranged from a 29% increase to a 16% reduction in mortality, with 39% of patients predicted to be harmed.

CONCLUSIONS

Although both clinical and biologic phenotyping may identify subgroups whose average treatment effect is beneficial or harmful, individual risks and benefits within subgroups still vary dramatically, raising concern that phenotyping may not reliably or safely personalize sepsis care.