AI-enabled forecasting of prehospital transfusion needs in patients with trauma: a multinational, registry-based, retrospective, machine learning development and validation study

BACKGROUND: Trauma is a major global cause of morbidity and mortality, with haemorrhage representing a leading preventable cause of early death. Timely blood transfusion is a crucial intervention, but current prehospital decision-making tools are scarce. Conventional triggers, such as haemoglobin concentrations, are often unreliable in the acute setting. There is a clear need for more robust, data-driven methods to guide transfusion decisions before hospital arrival. METHODS: We conducted a retrospective, machine learning development and validation study to predict the need for prehospital transfusion in patients with trauma using readily available prehospital data, including vital signs, injury patterns, and anticoagulant medication taken before hospitalisation occurred. The models were trained on data obtained from 364 350 patients in the American National Trauma Data Bank from Jan 1 to Dec 31, 2020, and externally validated on data from 54 210 patients from three additional trauma registries (TraumaRegister DGU, National Office of Clinical Audit-Major Trauma Audit, and Alberta Trauma Registry of Alberta Health Services), covering cases from Germany, Austria, Switzerland, Ireland, and Canada between Jan 1, 2007, and Sept 30, 2024. Binary classifiers were trained for individual blood products, while a multiclass model predicted optimal transfusion combinations, and a regressor for the optimal amount of packed red blood cells (PRBCs). FINDINGS: The machine learning models demonstrated high predictive accuracy in identifying patients requiring transfusion. In the external validation cohort, the area under the receiver operating characteristic curve for predicting any transfusion need was 0·87 (95% CI 0·86-0·87), and was 0·88 (0·87-0·89) for PRBCs. The machine learning-based predictions outperformed laboratory-based risk stratification upon emergency department arrival. Stratification into transfusion probability groups showed that patients in the high transfusion probability group (predicted transfusion probability >0·5) had the highest incidence of overall mortality (p INTERPRETATION: Machine learning-based prediction of transfusion needs enables prehospital identification of patients at high risk for haemorrhagic shock, supporting early intervention and resource mobilisation. This strategy might improve outcomes by facilitating timely availability of blood products. Our findings support the potential use of artificial intelligence-driven decision support tools into emergency trauma care workflows, but further confirmation is needed with prospective usability and effectiveness studies before clinical implementation. FUNDING: None.