Itsuki Osawa, MD
Clinical Fellow
The University of Tokyo Hospital, Japan
Disclosure information not submitted.
Daisuke Kudo, MD, PhD
Associate Professor
Division of Emergency and Critical Care Medicine, Tohoku University Graduate School of Medicine, Japan
Disclosure information not submitted.
Toshikazu Abe, MD, MPH, PhD
Director
Department of Emergency and Critical Care Medicine, Tsukuba Memorial Hospital, Japan
Disclosure information not submitted.
Mineji Hayakawa, MD, PhD
Associate Professor
Department of Emergency Medicine, Hokkaido University Hospital, Japan
Disclosure information not submitted.
Atsushi Shiraishi, MD, PhD
Director
Emergency and Trauma Center, Kameda Medical Center, Japan
Disclosure information not submitted.
Ryo Uchimido, MD, MPH
Project Assistant Professor
Department of Intensive Care Medicine, Tokyo Medical and Dental University, Japan
Disclosure information not submitted.
Kazuma Yamakawa, MD, PhD
Associate Professor
Department of Emergency Medicine, Osaka Medical and Pharmaceutical University, Japan
Disclosure information not submitted.
Kent Doi, MD, PhD
Professor
Department of Emergency and Critical Care Medicine, The University of Tokyo Hospital, Japan
Disclosure information not submitted.
Shigeki Kushimoto, MD, PhD
Professor
Division of Emergency and Critical Care Medicine, Tohoku University Graduate School of Medicine, Japan
Disclosure information not submitted.
Title: A Machine Learning-Based Estimation of Potential Targets of Polymyxin-B Hemoperfusion Use for Sepsis
Introduction: Polymyxin B hemoperfusion (PMX) was not shown to reduce 28-day mortality in patients with septic shock in the EUPHRATES trial. However, given that the trial estimated the treatment effects of PMX only in certain prespecified patient groups, we aimed to find previously unidentified patient subgroups with septic shock who may benefit from targeted use of PMX using a machine-leaning based algorithm.
Methods: We retrospectively analyzed data from multicenter registries in Japan to identify patients (aged ≥16 years) with septic shock who were admitted to hospitals where PMX was available. We identified 2131 patients in 34 ICUs from one registry as a derivation cohort, and 772 patients in 38 ICUs from two registries as a validation cohort. We applied the causal forest to the derivation cohort to estimate individual treatment effects (ITEs) of PMX on 28-day mortality adjusted for potential confounders. The causal forest is a machine learning-based algorithm that estimates the treatment effect on each individual conditional on the covariates. After ranking patients into quintiles of estimated ITEs, we compared the patient characteristics in each group. Moreover, we used the policytree algorithm to choose the best criteria for determining the optimal target for PMX, using covariates primarily related to septic coagulopathy. To confirm the consistency of our findings, we evaluated the treatment effect of PMX in the targeted subpopulation of the validation cohort using inverse probability of treatment weighting.
Results: Patients in the top quintile of estimated ITEs of PMX were more likely to have severe coagulopathy (i.e., a higher PT-INR and a lower fibrinogen level) on admission. Additionally, based on the results of the policytree algorithm, we selected the potential treatment targets of PMX as patients with septic shock and coagulopathy of PT-INR≥1.4 (46% of the derivation cohort). In the validation cohort, PMX use on the targeted subpopulation was significantly associated with a lower 28-day mortality (adjusted risk difference −15.7%; 95%CI −25.6 to −5.8%).
Conclusion: We found that parameters indicating septic coagulopathy could effectively identify potential targets of PMX for sepsis. Our findings could be beneficial for tailoring the use of PMX and bridging it to future interventional trials.