Heterogeneity in response to early goal-directed therapy in sepsis

Introduction: Heterogeneity of treatment effect (HTE) may explain negative results of contemporary clinical trials of early goal directed therapy (EGDT) in sepsis that contrasted prior trials that demonstrated benefit. We employed used state-of-the-art computational methodologies to identify HTE to EGDT at an individual patient level. Methods: We analyzed data from the ARISE and ProCESS multicenter international EGDT trials. We used causal forest models that incorporate baseline covariate, treatment, and outcome data to estimate individual absolute risk reductions (iARRs) whereby positive iARRs suggest benefit with EGDT. In our primary analysis, we trained causal forests in ARISE (derivation) and predicted iARRs for all patients in ProCESS (validation). We performed a formal test for HTE by testing for interaction between iARR and treatment with 90-day mortality as a primary outcome. We identified the variables contributing to treatment response using SHAP values. We performed a secondary analysis with ProCESS as a derivation set and ARISE as a validation set. Results: We predicted iARRs in individual ProCESS patients using a causal forest model trained in ARISE. Absolute risk reduction from EGDT in ProCESS ranged from -21% (90% CI: -35%, -7%, suggesting harm from EGDT) in the quintile of patients with the lowest predicted iARRs to +8% (90% CI: -5%, +22%) in the quintile of patients with the highest iARRs. The interaction between iARR prediction and treatment was significant (p< 0.001) providing evidence for HTE to EGDT. The variables contributing most to individual response to EGDT by SHAP analysis included albumin, glucose, temperature, lactate, blood urea nitrogen, and hemoglobin. Results were consistent in our secondary analyses, including identification of a subgroup harmed by EGDT and similar variables contributing to individual response. Conclusions: In a secondary analysis of two multicenter trials of EGDT, we found evidence of HTE using approaches that model treatment response at an individual patient level and we identified variables contributing to treatment response. We identified a subgroup of patients harmed by EGDT with the remaining patients having modest or no benefit from EGDT compared to usual care.