Abstract
Accurately predicting surgical and transfusion needs in traumatic brain injury (TBI) patients remains challenging in emergency settings. We developed multiomics data fusion (MDF) models integrating clinical biomarkers, neural radiological imaging, and clinical text mining for predicting surgical intervention and blood transfusion requirements across four multicenter cohorts (N = 2219). The MDF models provided predictions a median of 3 hours before interventions, with surgical model F1 scores of 0.63-0.85 across external testing, outperforming single-domain approaches. The transfusion model demonstrated strong cross-center performance (validation/external F1 scores: 0.78, 0.74) and correlated well with actual transfusion volumes (validation: R = 0.687, external: R = 0.580). SHapley additive exPlanations revealed radiological features drove surgical predictions, while clinical parameters (lactate, GCS scores, pupillary reflex, and hemoglobin) were crucial for transfusion predictions. We also developed a simplified emergency model maintaining robust performance (validation AUC: 0.81, external AUC: 0.75). These models demonstrate cross-center generalizability and practical utility for emergency settings, supporting clinical implementation for improved TBI patient management.