A unified multi-modal foundation model for end-to-end emergency care

Against the backdrop of population aging and the rising burden of chronic diseases in China, emergency departments (EDs) in tertiary hospitals remain persistently overcrowded. ED overcrowding is a multifactorial phenomenon arising from the interplay of input, throughput, and output factors along the emergency care workflow. Existing artificial intelligence approaches for emergency care are predominantly task-specific. Although effective for isolated clinical tasks, they lack patient-centric, end-to-end integration and generalize poorly under heterogeneous and incomplete inputs, limiting scalable deployment in resource-heterogeneous real-world settings. Here we propose ED-Foundation, a unified emergency foundation model built on the BEIT-3 architecture. By jointly leveraging aligned image-text pairs and unaligned pure-text data, ED-Foundation learns continuous patient-centric representations within a single model, and a two-stage self-supervised learning framework further improves robustness to pervasive missingness. We evaluate ED-Foundation on nine downstream validation datasets spanning three core emergency scenarios that address bottlenecks at different points along the emergency care workflow: early emergency triage, outcome prediction during the clinical course, and clinical decision-making support in the emergency treatment phase. ED-Foundation consistently achieves state-of-the-art retrospective performance and maintains robustness under limited information and modality missingness, outperforming prior task-specific approaches and existing foundation models across diverse institutional and cross-system evaluation settings. These results provide retrospective evidence supporting the benchmark performance and representational transferability of ED-Foundation as a unified representation learning framework for emergency care tasks.