Abstract
BACKGROUND: To assess heterogeneous treatment effects of high fresh frozen plasma (FFP) to red blood cell (RBC) transfusion ratios in patients with severe blunt trauma and to identify subgroups that derive the greatest survival benefit. METHODS: This multicenter retrospective cohort study used data from the Japan Trauma Data Bank (2019–2023). Adults with severe blunt trauma (Injury Severity Score ≥ 16) who received transfusions were included. Patients were categorized into high-FFP (FFP:RBC > 1) and low-FFP (FFP:RBC ≤ 1) groups. A causal forest machine learning model was applied to a derivation cohort (2019–2021) to estimate conditional average treatment effects (CATEs) and identify subgroups with the highest predicted benefit. Findings were validated in a separate cohort (2022–2023). RESULTS: Among 6,679 patients, in-hospital mortality was 23.3% in the derivation and 23.2% in the validation cohort. Causal forest analysis revealed lactate level and Glasgow Coma Scale (GCS) score as key effect modifiers. A therapeutic target subgroup—defined as lactate ≥ 4.5 mmol/L and GCS ≤ 12—comprised 20.7% of the validation cohort. This subgroup showed a substantially greater mortality reduction with high-FFP transfusion (risk difference –13.3%, 95% CI –22.4 to –4.2%; number needed to treat [NNT] 7.5), compared with the overall cohort (risk difference –3.3%, 95% CI –6.7 to 0.5%; NNT 32.1). Results were consistent across sensitivity analyses. CONCLUSIONS: High FFP-to-RBC transfusion ratios may confer the greatest benefit in patients with impaired consciousness and metabolic acidosis. Identifying high-benefit subgroups using machine learning could support more individualized transfusion strategies in trauma care. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s13054-025-05678-z.