Abstract

Massive transfusion (MT) prediction in trauma remains limited by scoring systems with suboptimal discrimination and generalizability. We developed a multimodal joint fusion model integrating structured electronic health record (EHR) data with admission chest X-rays to support early MT risk stratification. From 33,824 trauma patients (2014-2024), 10,090 met inclusion criteria, including 226 MT cases (2.24%). Structured presentation variables were modeled using a multilayer perceptron, while imaging features were extracted using a pretrained DenseNet-121; intermediate feature fusion enabled end-to-end learning. Class imbalance was addressed with SMOTE for structured data and geometric augmentation for imaging. On the held-out validation set, the fusion model achieved an AUC of 0.669. At the selected threshold, sensitivity was 0.72 and specificity 0.84. Grad-CAM visualizations demonstrated attention over clinically relevant thoracoabdominal regions. Despite limited positive predictive value due to low event prevalence, these findings demonstrate feasibility of workflow-aligned multimodal MT prediction using routinely available data, warranting prospective multi-center validation.

Back to List