AUTHOR=Yang Penglei , Yuan Jun , He Jie , Yu Lina , Gu Xue , Ding Xizhen , Chen Qihong 

TITLE=Investigating the feasibility of machine learning to guide personalized red blood cell (RBC) transfusion: analyzing the heterogeneity of RBC transfusion in septic patients with hemoglobin levels of 7–9 g/dL based on the causal forest model

JOURNAL=Frontiers in Pharmacology

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/pharmacology/articles/10.3389/fphar.2025.1615618

DOI=10.3389/fphar.2025.1615618

ISSN=1663-9812

ABSTRACT=BackgroundThis study utilized the causal forest algorithm to explore the heterogeneity of treatment effects of low-dose red blood cell (RBC) transfusion on the 90-day survival rate of sepsis patients with hemoglobin (Hb) levels of 7–9 g/dL to develop personalized transfusion strategies.MethodsThe data of patients the met the Sepsis-3 criteria with a minimum Hb level of 7–9 g/dL were obtained from the MIMIC-IV and MIMIC-III databases and divided into RBC transfusion and non-transfusion groups. Patients in both groups were paired using a propensity score matching analysis (PSM) after which a causal forest model was constructed using MIMIC-IV data. The model’s accuracy was analyzed using out-of-bag data. Individual treatment effects (ITE) of MIMIC-III patients were predicted and categorized into four subgroups: Quantile1 to Quantile4, based on the effect size. Kaplan-Meier survival curves were established for each Quantile to determine the survival rates.ResultsThe MIMIC-IV and MIMIC-III database comprised 1,652 and 868 patients, with 826 (50%) and 434 (50%) in the RBC transfusion group, respectively, after PSM. The mean prediction coefficient estimated by the causal forest was 1.00 with a standard error of 0.57, while the differential forest prediction coefficient was 1.64 with a standard error of 0.48, demonstrating the model’s ability to effectively identify differences in the impact of transfusion on survival rates among individuals. There was significant heterogeneity in the ITE among patients in the MIMIC-III validation cohort. Moreover, the ITE values were divided into Quantile1: −5.4% (−8.0%, −3.9%), Quantile2: −2.1% (−2.6%, −1.7%), Quantile3: −0.5% (−0.1%, +0.1%), and Quantile 4: +3.6% (+2.0%, +6.6%). The Kaplan-Meier curves and the log-rank test demonstrated that the RBC transfusion decreased the survival of patients in Quantile1 (p < 0.001) and Quantile2 (p = 0.011) but increased the survival of patients in Quantile4 (p < 0.001).ConclusionRBC transfusions among sepsis patients with Hb levels of 7–9 g/dL exhibit heterogenous treatment effects, which reduces the mortality of patients with high ITE. Although the causal forest model can guide personalized transfusion in these cases, randomized controlled trials are needed to validate these findings.